Predicting Bitcoin Prices
Practicing Deep Learning Time Series Forecasting on BitCoin Data
From ZeroToMastery Tensorflow Course
!nvidia-smi -LGPU 0: Tesla K80 (UUID: GPU-3d431c8e-89fd-3326-713b-0dfee7b92355)Table of Contents
Get Data
Download all historical data from here: https://www.coindesk.com/price/bitcoin/
## Import time series data with pandas
import pandas as pd
df = pd.read_csv('/content/BTC_USD_2014-11-04_2021-10-25-CoinDesk.csv',
parse_dates=['Date'],
index_col=["Date"]) # put date as index
df.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
df.info()<class 'pandas.core.frame.DataFrame'>
DatetimeIndex: 2548 entries, 2014-11-04 to 2021-10-25
Data columns (total 5 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 Currency 2548 non-null object
1 Closing Price (USD) 2548 non-null float64
2 24h Open (USD) 2548 non-null float64
3 24h High (USD) 2548 non-null float64
4 24h Low (USD) 2548 non-null float64
dtypes: float64(4), object(1)
memory usage: 119.4+ KB# How many samples
len(df)2548df.tail().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
Note: we have 2548 rows, but generally deep learning models need a lot more data to be effective
Smaller sample size is a problem that is common to time series problems
Note: Seasonality = number of samples per year. This data has a seasonality of 365
Important Time series patterns
Trend: time series has a clear long term increase/decrease
Seasonal: time of year / day of week / season / etc.
Cyclic: has no fixed time frame
# Only want closing price
bitcoin_prices = pd.DataFrame(df["Closing Price (USD)"]).rename(columns={"Closing Price (USD)": "Price"})
bitcoin_prices.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
import matplotlib.pyplot as plt
bitcoin_prices.plot(figsize=(10,7))
plt.ylabel("BTC Price")
plt.title("Price of Bitcoin from Nov 2014 to Oct 2021")
plt.legend(fontsize=14);Importing time series data with Python's CLV module
# Importing and formatting historical bitcoin data with Python
import csv
from datetime import datetime
timesteps = []
prices = []
with open("/content/BTC_USD_2014-11-04_2021-10-25-CoinDesk.csv", 'r') as f:
csv_reader = csv.reader(f, delimiter=",")
next(csv_reader) # will skip first row
for line in csv_reader:
timesteps.append(datetime.strptime(line[1], "%Y-%m-%d")) # get the dates as dates not strings
prices.append(float(line[2])) # get closing price as float
# View first 10 of each
timesteps[:10], prices[:10]([datetime.datetime(2014, 11, 4, 0, 0),
datetime.datetime(2014, 11, 5, 0, 0),
datetime.datetime(2014, 11, 6, 0, 0),
datetime.datetime(2014, 11, 7, 0, 0),
datetime.datetime(2014, 11, 8, 0, 0),
datetime.datetime(2014, 11, 9, 0, 0),
datetime.datetime(2014, 11, 10, 0, 0),
datetime.datetime(2014, 11, 11, 0, 0),
datetime.datetime(2014, 11, 12, 0, 0),
datetime.datetime(2014, 11, 13, 0, 0)],
[324.71833,
332.45666,
336.585,
346.775,
344.81166,
343.065,
358.50166,
368.07666,
376.99666,
442.10666])# Plot from CSV
import matplotlib.pyplot as plt
import numpy as np
plt.figure(figsize=(10, 7))
plt.plot(timesteps, prices)
plt.title("Price of Bitcoin from Nov 2014 to Oct 2021", fontsize=16)
plt.xlabel("Date")
plt.ylabel("BTC Price");Creating train test splits
# Create train and test splits the right way for time series data
split_size = int(0.8 * len(prices)) # 80% train, 20% test
# Create train data splits (everything before the split)
X_train, y_train = timesteps[:split_size], prices[:split_size]
# Create test data splits (everything after the split)
X_test, y_test = timesteps[split_size:], prices[split_size:]
len(X_train), len(X_test), len(y_train), len(y_test)(2038, 510, 2038, 510)# Plot train test splits
plt.figure(figsize=(10, 7))
plt.scatter(X_train, y_train, s=5, label="Train data")
plt.scatter(X_test, y_test, s=5, label="Test data")
plt.xlabel("Date")
plt.ylabel("BTC Price")
plt.legend(fontsize=14)
plt.show();Side note: This article is fun read about differences in time series compared to regular machine learning
https://towardsdatascience.com/3-facts-about-time-series-forecasting-that-surprise-experienced-machine-learning-practitioners-69c18ee89387
Create a plotting function
def plot_time_series(timesteps, values, format=".", start=0, end=None, label=None):
"""
Plots timesteps against values
Parameters
---------
timesteps: array of timestep values
values: array of values across time
format: style of plot, default = scatter
start: where to start the plot
end: where to end the plot
label: label to show on plot about values
"""
plt.plot(timesteps[start:end], values[start:end], format, label=label)
plt.xlabel("Time")
plt.ylabel("BTC Price")
if label:
plt.legend(fontsize=14) # make label bigger
plt.grid(True)# Try out our plotting function
plt.figure(figsize=(10, 7))
plot_time_series(timesteps=X_train, values=y_train, label="Train data")
plot_time_series(timesteps=X_test, values=y_test, label="Test data")Modelling Experiments
0: Naive
1: Dense, horizon = 1, window = 7
2: Dense, horizon = 1, window = 30
3: Dense, horizon = 7, window = 30
4: Conv1D, 1 7
5: LSTM 1 7
6: Dense but multivariate
7: N-BEATs 1 7
8: Ensemble
9: Future prediction
10: Silly model
Model 0: Naive Forecast
The formula looks like this:
y^t=yt−1
The prediction at time t will be equal to previous timestep at time t-1
naive_forecast = y_test[:-1]
naive_forecast[:10], naive_forecast[-10:]([9667.0607542595,
9811.367094602,
9659.6612981892,
9677.0509945125,
9737.458558786,
9791.9278997681,
9786.1420204782,
9874.967047524,
9260.8326409877,
9464.2280968215],
[61273.805275885,
60744.9849380446,
60965.4654336171,
61707.4813118213,
63901.0730802098,
65972.8631297649,
62603.5750699116,
60689.2382648626,
61124.3471256221,
60936.1508514192])# Plot naive forecast
plt.figure(figsize=(10,7))
#plot_time_series(timesteps=X_train, values=y_train, label="Train data")
plot_time_series(timesteps=X_test, values=y_test,start=350,format='-', label="Test data")
plot_time_series(timesteps=X_test[1:], values=naive_forecast, start=350, format='-', label='Naive Forecast')Review: Different Evaluation Metrics
Regression problem so can use regression metrics:
MAE - interpretable - forecasts lead to median
RMSE - interpretable - forecasts lead to mean
MASE - comparing model's performance to naive forecast
MAPE - percentage errors can be dangerous if y=0
Essentially looking for this: how do our model's forecast compare against actual values?
See this link: https://otexts.com/fpp3/accuracy.html
import tensorflow as tf
# MASE implemented courtesy of sktime - https://github.com/alan-turing-institute/sktime/blob/ee7a06843a44f4aaec7582d847e36073a9ab0566/sktime/performance_metrics/forecasting/_functions.py#L16
def mean_absolute_scaled_error(y_true, y_pred):
"""
Implement MASE (assuming no seasonality of data).
"""
mae = tf.reduce_mean(tf.abs(y_true - y_pred))
# Find MAE of naive forecast (no seasonality)
mae_naive_no_season = tf.reduce_mean(tf.abs(y_true[1:] - y_true[:-1])) # our seasonality is 1 day (hence the shifting of 1 day)
return mae / mae_naive_no_season# Put all eval metrics into a function
def evaluate_preds(y_true, y_pred):
# Make sure float32 (for metric calculations)
y_true = tf.cast(y_true, dtype=tf.float32)
y_pred = tf.cast(y_pred, dtype=tf.float32)
# Calculate various metrics
mae = tf.keras.metrics.mean_absolute_error(y_true, y_pred)
mse = tf.keras.metrics.mean_squared_error(y_true, y_pred) # puts and emphasis on outliers (all errors get squared)
rmse = tf.sqrt(mse)
mape = tf.keras.metrics.mean_absolute_percentage_error(y_true, y_pred)
mase = mean_absolute_scaled_error(y_true, y_pred)
return {"mae": mae.numpy(),
"mse": mse.numpy(),
"rmse": rmse.numpy(),
"mape": mape.numpy(),
"mase": mase.numpy()}naive_results = evaluate_preds(y_true=y_test[1:],
y_pred=naive_forecast)
naive_results{'mae': 924.30414,
'mape': 2.6047938,
'mase': 0.99834156,
'mse': 2055079.2,
'rmse': 1433.5548}tf.reduce_min(y_test), tf.reduce_mean(y_test), tf.reduce_max(y_test)(<tf.Tensor: shape=(), dtype=float32, numpy=9022.153>,
<tf.Tensor: shape=(), dtype=float32, numpy=32127.76>,
<tf.Tensor: shape=(), dtype=float32, numpy=65972.86>)We can guess how the model is doing by looking at where the majority of values are
For example, if most values are closer to 60k, then an error of 1000 is not that bad
Side note: Other models we could use
This notebook is focused on tensorflow, but maybe for future work can use those and compare to the final best tensorflow model
Moving average https://machinelearningmastery.com/moving-average-smoothing-for-time-series-forecasting-python/
ARIMA (Autoregression Integrated Moving Average) https://machinelearningmastery.com/arima-for-time-series-forecasting-with-python/
sktime (Scikit-Learn for time series) https://github.com/alan-turing-institute/sktime
TensorFlow Decision Forests (random forest, gradient boosting trees) https://www.tensorflow.org/decision_forests
Facebook Kats (purpose-built forecasting and time series analysis library by Facebook) https://github.com/facebookresearch/Kats
LinkedIn Greykite (flexible, intuitive and fast forecasts) https://github.com/linkedin/greykite
Windowing our Time Series
We window to turn our data into a supervised learning problem
HORIZON = 1 # predict 1 step at a time
WINDOW_SIZE = 7 # use a week worth of timesteps to predict the horizon# Create function to label windowed data
def get_labelled_window(x, horizon=HORIZON):
"""
Creates labels for windowed dataset
If horizon=1
input: [0,1,2,3,4,5,6,7] -> output: ([0,1,2,3,4,5,6], [7])
"""
return x[:, :-horizon], x[:, -horizon:]# Test out the window labelling function
test_window, test_label = get_labelled_window(tf.expand_dims(tf.range(8)+1, axis=0), horizon=HORIZON)
print(f"Window: {tf.squeeze(test_window).numpy()} -> Label: {tf.squeeze(test_label).numpy()}")Window: [1 2 3 4 5 6 7] -> Label: 8Great let's now use this basic function to iterate through all the data
For loops would take a while so we will use numpy array indexing: https://towardsdatascience.com/fast-and-robust-sliding-window-vectorization-with-numpy-3ad950ed62f5
np.arange(4)array([0, 1, 2, 3])np.expand_dims(np.arange(4), axis=0)array([[0, 1, 2, 3]])np.expand_dims(np.arange(4), axis=0).Tarray([[0],
[1],
[2],
[3]])# Simple example of the above matrix addition
np.expand_dims(np.arange(4), axis=0).T + np.expand_dims(np.arange(8), axis=0)array([[ 0, 1, 2, 3, 4, 5, 6, 7],
[ 1, 2, 3, 4, 5, 6, 7, 8],
[ 2, 3, 4, 5, 6, 7, 8, 9],
[ 3, 4, 5, 6, 7, 8, 9, 10]])import numpy as np
# Create function to view NumPy arrays as windows copying the blogpost
def make_windows(x, window_size=7, horizon=1):
"""
Turns a 1D array into a 2D array of sequential windows of window_size.
"""
# 1. This is array 1
window_step = np.expand_dims(np.arange(window_size+horizon), axis=0)
# print(f"Window step:\n {window_step}")
# 2. Add horizontal array and vertical array together
window_indexes = window_step + np.expand_dims(np.arange(len(x)-(window_size+horizon-1)), axis=0).T # this is the vertical array [[0],[1],[2],...,[T]]
# print(f"Window indexes:\n {window_indexes[:3], window_indexes[-3:], window_indexes.shape}")
# 3. Index on the target array (time series) with 2D array of multiple window steps
windowed_array = x[window_indexes]
# 4. Get the labelled windows
windows, labels = get_labelled_window(windowed_array, horizon=horizon)
return windows, labelsfull_windows, full_labels = make_windows(np.array(prices), window_size=WINDOW_SIZE, horizon=HORIZON)
len(full_windows), len(full_labels)(2541, 2541)# View the first 3 windows/labels
for i in range(3):
print(f"Window: {full_windows[i]} -> Label: {full_labels[i]}")Window: [324.71833 332.45666 336.585 346.775 344.81166 343.065 358.50166] -> Label: [368.07666]
Window: [332.45666 336.585 346.775 344.81166 343.065 358.50166 368.07666] -> Label: [376.99666]
Window: [336.585 346.775 344.81166 343.065 358.50166 368.07666 376.99666] -> Label: [442.10666]Turning windows into training and test sets
full_windows[:5], full_labels[:5](array([[324.71833, 332.45666, 336.585 , 346.775 , 344.81166, 343.065 ,
358.50166],
[332.45666, 336.585 , 346.775 , 344.81166, 343.065 , 358.50166,
368.07666],
[336.585 , 346.775 , 344.81166, 343.065 , 358.50166, 368.07666,
376.99666],
[346.775 , 344.81166, 343.065 , 358.50166, 368.07666, 376.99666,
442.10666],
[344.81166, 343.065 , 358.50166, 368.07666, 376.99666, 442.10666,
389.00333]]), array([[368.07666],
[376.99666],
[442.10666],
[389.00333],
[394.92666]]))def make_train_test_splits(windows, labels, test_split=0.2):
"""
Splits matching pairs of windows and labels into train and test splits
"""
split_size = int(len(windows) * (1-test_split)) # will default to 0.8 train
train_windows = windows[:split_size]
train_labels = labels[:split_size]
test_windows = windows[split_size:]
test_labels = labels[split_size:]
return train_windows, test_windows, train_labels, test_labelstrain_windows, test_windows, train_labels, test_labels = make_train_test_splits(full_windows, full_labels)
len(train_windows), len(test_windows), len(train_labels), len(test_labels)(2032, 509, 2032, 509)train_windows[:5], train_labels[:5](array([[324.71833, 332.45666, 336.585 , 346.775 , 344.81166, 343.065 ,
358.50166],
[332.45666, 336.585 , 346.775 , 344.81166, 343.065 , 358.50166,
368.07666],
[336.585 , 346.775 , 344.81166, 343.065 , 358.50166, 368.07666,
376.99666],
[346.775 , 344.81166, 343.065 , 358.50166, 368.07666, 376.99666,
442.10666],
[344.81166, 343.065 , 358.50166, 368.07666, 376.99666, 442.10666,
389.00333]]), array([[368.07666],
[376.99666],
[442.10666],
[389.00333],
[394.92666]]))# Check to see if train labels are the same (before and after window split)
np.array_equal(np.squeeze(train_labels[:-HORIZON-1]), np.array(y_train[WINDOW_SIZE:-1]))TrueModeling Checkpoint
We want to compare the best performance/epoch of each model
import os
# Create a function to implement a ModelCheckpoint callback with a specific filename
def create_model_checkpoint(model_name, save_path="model_experiments"):
return tf.keras.callbacks.ModelCheckpoint(filepath=os.path.join(save_path, model_name), # create filepath to save model
verbose=0, # only output a limited amount of text
save_best_only=True) # save only the best model to fileModel 1: Dense 1, 7
Single dense layer with 128 and relu
Output layer with linear activation (no activation)
Adam and MAE
Batch size 128 (data is pretty small so can do bigger batches)
100 epochs
Could tune the hyperparameters but that could be an extension for later
import tensorflow as tf
from tensorflow.keras import layers
# Set random seed
tf.random.set_seed(42)
# Create model
model_1 = tf.keras.Sequential([
layers.Dense(128, activation='relu'),
layers.Dense(HORIZON)
], name='model_1_dense')
# Compile
model_1.compile(loss='mae',
optimizer=tf.keras.optimizers.Adam(),
metrics=['mae', "mse"])
# Fit the model
history_1 = model_1.fit(x=train_windows,
y=train_labels,
epochs=100,
verbose=1,
batch_size=128,
validation_data=(test_windows, test_labels),
callbacks=[create_model_checkpoint(model_name=model_1.name)])Epoch 1/100
14/16 [=========================>....] - ETA: 0s - loss: 1125.1427 - mae: 1125.1427 - mse: 3623911.5000 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 4s 78ms/step - loss: 1052.2838 - mae: 1052.2838 - mse: 3281413.0000 - val_loss: 4038.3960 - val_mae: 4038.3960 - val_mse: 27835144.0000
Epoch 2/100
14/16 [=========================>....] - ETA: 0s - loss: 376.2725 - mae: 376.2725 - mse: 457702.9062INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 47ms/step - loss: 370.0820 - mae: 370.0820 - mse: 441171.4688 - val_loss: 2101.2781 - val_mae: 2101.2781 - val_mse: 8439465.0000
Epoch 3/100
15/16 [===========================>..] - ETA: 0s - loss: 247.7293 - mae: 247.7293 - mse: 244647.3125INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 44ms/step - loss: 248.8695 - mae: 248.8695 - mse: 246683.7969 - val_loss: 1932.4431 - val_mae: 1932.4431 - val_mse: 7341109.5000
Epoch 4/100
15/16 [===========================>..] - ETA: 0s - loss: 236.4443 - mae: 236.4443 - mse: 217470.1406INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 236.8721 - mae: 236.8721 - mse: 219932.3750 - val_loss: 1686.6390 - val_mae: 1686.6390 - val_mse: 6009384.0000
Epoch 5/100
13/16 [=======================>......] - ETA: 0s - loss: 225.7914 - mae: 225.7914 - mse: 214336.6094INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 51ms/step - loss: 228.8478 - mae: 228.8478 - mse: 211895.9531 - val_loss: 1581.4629 - val_mae: 1581.4629 - val_mse: 5512004.0000
Epoch 6/100
13/16 [=======================>......] - ETA: 0s - loss: 219.0259 - mae: 219.0259 - mse: 193449.3438INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 47ms/step - loss: 220.2377 - mae: 220.2377 - mse: 200366.5781 - val_loss: 1567.6823 - val_mae: 1567.6823 - val_mse: 5342175.5000
Epoch 7/100
14/16 [=========================>....] - ETA: 0s - loss: 217.9900 - mae: 217.9900 - mse: 200676.9844INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 214.8801 - mae: 214.8801 - mse: 192315.7188 - val_loss: 1524.3950 - val_mae: 1524.3950 - val_mse: 5073459.5000
Epoch 8/100
15/16 [===========================>..] - ETA: 0s - loss: 202.3831 - mae: 202.3831 - mse: 168473.8125INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 210.5827 - mae: 210.5827 - mse: 185379.6562 - val_loss: 1470.5625 - val_mae: 1470.5625 - val_mse: 4827353.5000
Epoch 9/100
14/16 [=========================>....] - ETA: 0s - loss: 206.6307 - mae: 206.6307 - mse: 180371.8594INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 204.5752 - mae: 204.5752 - mse: 175921.7188 - val_loss: 1414.4553 - val_mae: 1414.4553 - val_mse: 4464321.0000
Epoch 10/100
16/16 [==============================] - 0s 7ms/step - loss: 201.6268 - mae: 201.6268 - mse: 170200.0469 - val_loss: 1425.2523 - val_mae: 1425.2523 - val_mse: 4339175.5000
Epoch 11/100
15/16 [===========================>..] - ETA: 0s - loss: 195.5575 - mae: 195.5575 - mse: 161337.8906INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 192.8863 - mae: 192.8863 - mse: 158605.2344 - val_loss: 1356.2078 - val_mae: 1356.2078 - val_mse: 4055710.5000
Epoch 12/100
16/16 [==============================] - 0s 6ms/step - loss: 192.2759 - mae: 192.2759 - mse: 155920.8594 - val_loss: 1360.0175 - val_mae: 1360.0175 - val_mse: 3945523.2500
Epoch 13/100
15/16 [===========================>..] - ETA: 0s - loss: 185.4322 - mae: 185.4322 - mse: 144037.6719INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 48ms/step - loss: 185.7385 - mae: 185.7385 - mse: 145905.3750 - val_loss: 1321.9495 - val_mae: 1321.9495 - val_mse: 3745769.5000
Epoch 14/100
16/16 [==============================] - 0s 6ms/step - loss: 180.8296 - mae: 180.8296 - mse: 141572.0000 - val_loss: 1408.8519 - val_mae: 1408.8519 - val_mse: 4009162.5000
Epoch 15/100
15/16 [===========================>..] - ETA: 0s - loss: 180.9534 - mae: 180.9534 - mse: 137292.4219INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 181.0478 - mae: 181.0478 - mse: 138626.0312 - val_loss: 1252.2047 - val_mae: 1252.2047 - val_mse: 3543531.0000
Epoch 16/100
16/16 [==============================] - 0s 7ms/step - loss: 176.3963 - mae: 176.3963 - mse: 132472.8281 - val_loss: 1266.6610 - val_mae: 1266.6610 - val_mse: 3355603.7500
Epoch 17/100
14/16 [=========================>....] - ETA: 0s - loss: 173.3926 - mae: 173.3926 - mse: 131475.6250INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 43ms/step - loss: 171.4510 - mae: 171.4510 - mse: 128554.3125 - val_loss: 1240.1663 - val_mae: 1240.1663 - val_mse: 3230422.2500
Epoch 18/100
14/16 [=========================>....] - ETA: 0s - loss: 168.5066 - mae: 168.5066 - mse: 126398.6406INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 57ms/step - loss: 169.1706 - mae: 169.1706 - mse: 124567.4609 - val_loss: 1175.9944 - val_mae: 1175.9944 - val_mse: 3057932.5000
Epoch 19/100
16/16 [==============================] - ETA: 0s - loss: 164.9955 - mae: 164.9955 - mse: 119707.5156INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 164.9955 - mae: 164.9955 - mse: 119707.5156 - val_loss: 1159.4547 - val_mae: 1159.4547 - val_mse: 2976652.2500
Epoch 20/100
15/16 [===========================>..] - ETA: 0s - loss: 162.1346 - mae: 162.1346 - mse: 114131.6719INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 164.2351 - mae: 164.2351 - mse: 118974.2969 - val_loss: 1147.7201 - val_mae: 1147.7201 - val_mse: 2894545.7500
Epoch 21/100
16/16 [==============================] - 0s 7ms/step - loss: 160.2294 - mae: 160.2294 - mse: 114881.9844 - val_loss: 1214.4899 - val_mae: 1214.4899 - val_mse: 3045749.2500
Epoch 22/100
16/16 [==============================] - 0s 6ms/step - loss: 161.4803 - mae: 161.4803 - mse: 115629.6250 - val_loss: 1180.6323 - val_mae: 1180.6323 - val_mse: 2915320.5000
Epoch 23/100
15/16 [===========================>..] - ETA: 0s - loss: 160.6385 - mae: 160.6385 - mse: 114708.5078INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 159.1492 - mae: 159.1492 - mse: 111753.2578 - val_loss: 1110.0690 - val_mae: 1110.0690 - val_mse: 2714468.7500
Epoch 24/100
16/16 [==============================] - 0s 6ms/step - loss: 155.2771 - mae: 155.2771 - mse: 108248.1172 - val_loss: 1113.2184 - val_mae: 1113.2184 - val_mse: 2688186.0000
Epoch 25/100
15/16 [===========================>..] - ETA: 0s - loss: 156.8250 - mae: 156.8250 - mse: 108347.1250INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 157.5860 - mae: 157.5860 - mse: 108855.6406 - val_loss: 1078.8622 - val_mae: 1078.8622 - val_mse: 2624271.7500
Epoch 26/100
16/16 [==============================] - 0s 7ms/step - loss: 162.6721 - mae: 162.6721 - mse: 110743.4766 - val_loss: 1268.4597 - val_mae: 1268.4597 - val_mse: 3222643.7500
Epoch 27/100
13/16 [=======================>......] - ETA: 0s - loss: 158.8755 - mae: 158.8755 - mse: 108227.4062INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 161.1631 - mae: 161.1631 - mse: 113015.6562 - val_loss: 1066.5618 - val_mae: 1066.5618 - val_mse: 2557681.5000
Epoch 28/100
16/16 [==============================] - 0s 6ms/step - loss: 152.1051 - mae: 152.1051 - mse: 102798.0000 - val_loss: 1075.0272 - val_mae: 1075.0272 - val_mse: 2547176.0000
Epoch 29/100
15/16 [===========================>..] - ETA: 0s - loss: 146.8097 - mae: 146.8097 - mse: 99783.5078INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 149.1455 - mae: 149.1455 - mse: 102317.8125 - val_loss: 1049.2566 - val_mae: 1049.2566 - val_mse: 2505042.0000
Epoch 30/100
14/16 [=========================>....] - ETA: 0s - loss: 149.3239 - mae: 149.3239 - mse: 103105.3203INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 149.9327 - mae: 149.9327 - mse: 102230.4922 - val_loss: 1040.1512 - val_mae: 1040.1512 - val_mse: 2494505.7500
Epoch 31/100
14/16 [=========================>....] - ETA: 0s - loss: 145.9488 - mae: 145.9488 - mse: 96257.1875INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 147.7430 - mae: 147.7430 - mse: 101104.7500 - val_loss: 1035.3419 - val_mae: 1035.3419 - val_mse: 2456552.2500
Epoch 32/100
15/16 [===========================>..] - ETA: 0s - loss: 147.5689 - mae: 147.5689 - mse: 100420.4688INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 148.4546 - mae: 148.4546 - mse: 101211.2344 - val_loss: 1028.4969 - val_mae: 1028.4969 - val_mse: 2450137.0000
Epoch 33/100
14/16 [=========================>....] - ETA: 0s - loss: 146.9856 - mae: 146.9856 - mse: 97509.6172 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 147.1712 - mae: 147.1712 - mse: 100658.8672 - val_loss: 1022.3632 - val_mae: 1022.3632 - val_mse: 2428050.7500
Epoch 34/100
12/16 [=====================>........] - ETA: 0s - loss: 143.4796 - mae: 143.4796 - mse: 101960.2500INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 145.7424 - mae: 145.7424 - mse: 98862.5312 - val_loss: 1022.1561 - val_mae: 1022.1561 - val_mse: 2390078.7500
Epoch 35/100
16/16 [==============================] - 0s 7ms/step - loss: 145.4932 - mae: 145.4932 - mse: 98751.2500 - val_loss: 1077.5400 - val_mae: 1077.5400 - val_mse: 2517487.7500
Epoch 36/100
13/16 [=======================>......] - ETA: 0s - loss: 148.6425 - mae: 148.6425 - mse: 104479.6172INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 44ms/step - loss: 148.6831 - mae: 148.6831 - mse: 101312.3125 - val_loss: 1010.8162 - val_mae: 1010.8162 - val_mse: 2384694.0000
Epoch 37/100
16/16 [==============================] - 0s 7ms/step - loss: 148.5485 - mae: 148.5485 - mse: 97635.0391 - val_loss: 1011.3127 - val_mae: 1011.3127 - val_mse: 2388210.0000
Epoch 38/100
16/16 [==============================] - 0s 7ms/step - loss: 152.7205 - mae: 152.7205 - mse: 99604.3672 - val_loss: 1040.7112 - val_mae: 1040.7112 - val_mse: 2407394.7500
Epoch 39/100
16/16 [==============================] - 0s 6ms/step - loss: 144.3329 - mae: 144.3329 - mse: 95290.5312 - val_loss: 1017.5256 - val_mae: 1017.5256 - val_mse: 2348195.0000
Epoch 40/100
13/16 [=======================>......] - ETA: 0s - loss: 145.3250 - mae: 145.3250 - mse: 97471.1094INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 142.2780 - mae: 142.2780 - mse: 94249.9062 - val_loss: 992.9414 - val_mae: 992.9414 - val_mse: 2314894.0000
Epoch 41/100
16/16 [==============================] - 0s 6ms/step - loss: 144.4900 - mae: 144.4900 - mse: 96034.8906 - val_loss: 1078.0189 - val_mae: 1078.0189 - val_mse: 2511313.2500
Epoch 42/100
15/16 [===========================>..] - ETA: 0s - loss: 144.1443 - mae: 144.1443 - mse: 95577.6484INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 44ms/step - loss: 146.1954 - mae: 146.1954 - mse: 97416.1797 - val_loss: 987.5748 - val_mae: 987.5748 - val_mse: 2284384.0000
Epoch 43/100
16/16 [==============================] - 0s 7ms/step - loss: 145.4785 - mae: 145.4785 - mse: 96275.2422 - val_loss: 1009.1605 - val_mae: 1009.1605 - val_mse: 2364442.7500
Epoch 44/100
16/16 [==============================] - 0s 6ms/step - loss: 147.1797 - mae: 147.1797 - mse: 98086.0547 - val_loss: 1105.3702 - val_mae: 1105.3702 - val_mse: 2594204.0000
Epoch 45/100
13/16 [=======================>......] - ETA: 0s - loss: 143.4602 - mae: 143.4602 - mse: 88536.6797INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 147.0417 - mae: 147.0417 - mse: 97479.0078 - val_loss: 981.8096 - val_mae: 981.8096 - val_mse: 2274694.0000
Epoch 46/100
15/16 [===========================>..] - ETA: 0s - loss: 140.0057 - mae: 140.0057 - mse: 91701.1094 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 57ms/step - loss: 140.9648 - mae: 140.9648 - mse: 93108.9297 - val_loss: 980.0170 - val_mae: 980.0170 - val_mse: 2268344.0000
Epoch 47/100
16/16 [==============================] - 0s 8ms/step - loss: 140.0634 - mae: 140.0634 - mse: 92529.8906 - val_loss: 981.1994 - val_mae: 981.1994 - val_mse: 2248048.7500
Epoch 48/100
16/16 [==============================] - 0s 7ms/step - loss: 141.1929 - mae: 141.1929 - mse: 93225.9531 - val_loss: 1020.2706 - val_mae: 1020.2706 - val_mse: 2390321.2500
Epoch 49/100
14/16 [=========================>....] - ETA: 0s - loss: 143.5843 - mae: 143.5843 - mse: 91065.1094 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 143.9884 - mae: 143.9884 - mse: 94434.5781 - val_loss: 971.7318 - val_mae: 971.7318 - val_mse: 2237890.2500
Epoch 50/100
16/16 [==============================] - 0s 6ms/step - loss: 150.8942 - mae: 150.8942 - mse: 99836.5391 - val_loss: 1157.8853 - val_mae: 1157.8853 - val_mse: 2759950.5000
Epoch 51/100
14/16 [=========================>....] - ETA: 0s - loss: 151.7400 - mae: 151.7400 - mse: 100450.4375INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 151.5239 - mae: 151.5239 - mse: 101250.7422 - val_loss: 968.7036 - val_mae: 968.7036 - val_mse: 2221557.7500
Epoch 52/100
16/16 [==============================] - 0s 6ms/step - loss: 138.7729 - mae: 138.7729 - mse: 90782.7422 - val_loss: 969.0363 - val_mae: 969.0363 - val_mse: 2219408.7500
Epoch 53/100
15/16 [===========================>..] - ETA: 0s - loss: 138.5810 - mae: 138.5810 - mse: 89031.8438 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 139.7062 - mae: 139.7062 - mse: 91717.7266 - val_loss: 967.3604 - val_mae: 967.3604 - val_mse: 2213458.0000
Epoch 54/100
14/16 [=========================>....] - ETA: 0s - loss: 141.3379 - mae: 141.3379 - mse: 91776.3203 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 142.2512 - mae: 142.2512 - mse: 92233.3594 - val_loss: 962.0714 - val_mae: 962.0714 - val_mse: 2203328.2500
Epoch 55/100
16/16 [==============================] - 0s 6ms/step - loss: 139.5733 - mae: 139.5733 - mse: 91622.0938 - val_loss: 1099.4904 - val_mae: 1099.4904 - val_mse: 2560293.7500
Epoch 56/100
16/16 [==============================] - 0s 7ms/step - loss: 142.3359 - mae: 142.3359 - mse: 92018.7422 - val_loss: 1096.3219 - val_mae: 1096.3219 - val_mse: 2549114.2500
Epoch 57/100
16/16 [==============================] - 0s 6ms/step - loss: 148.8801 - mae: 148.8801 - mse: 99092.1562 - val_loss: 1141.7053 - val_mae: 1141.7053 - val_mse: 2697140.5000
Epoch 58/100
16/16 [==============================] - 0s 6ms/step - loss: 147.6814 - mae: 147.6814 - mse: 95996.1094 - val_loss: 1008.5109 - val_mae: 1008.5109 - val_mse: 2301740.7500
Epoch 59/100
16/16 [==============================] - 0s 7ms/step - loss: 137.2294 - mae: 137.2294 - mse: 89823.7969 - val_loss: 985.8022 - val_mae: 985.8022 - val_mse: 2243133.5000
Epoch 60/100
16/16 [==============================] - 0s 7ms/step - loss: 138.4461 - mae: 138.4461 - mse: 89879.0625 - val_loss: 998.8140 - val_mae: 998.8140 - val_mse: 2275104.7500
Epoch 61/100
16/16 [==============================] - 0s 7ms/step - loss: 138.9334 - mae: 138.9334 - mse: 92423.8281 - val_loss: 1046.4457 - val_mae: 1046.4457 - val_mse: 2398038.5000
Epoch 62/100
16/16 [==============================] - 0s 7ms/step - loss: 138.1750 - mae: 138.1750 - mse: 90686.6328 - val_loss: 963.2917 - val_mae: 963.2917 - val_mse: 2189547.2500
Epoch 63/100
15/16 [===========================>..] - ETA: 0s - loss: 136.1799 - mae: 136.1799 - mse: 89750.9219INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 47ms/step - loss: 136.7343 - mae: 136.7343 - mse: 89407.0781 - val_loss: 953.9246 - val_mae: 953.9246 - val_mse: 2181251.2500
Epoch 64/100
16/16 [==============================] - 0s 6ms/step - loss: 136.4803 - mae: 136.4803 - mse: 88897.3047 - val_loss: 976.1436 - val_mae: 976.1436 - val_mse: 2216084.2500
Epoch 65/100
16/16 [==============================] - 0s 9ms/step - loss: 137.0847 - mae: 137.0847 - mse: 89570.9531 - val_loss: 980.8021 - val_mae: 980.8021 - val_mse: 2226572.0000
Epoch 66/100
16/16 [==============================] - 0s 7ms/step - loss: 136.3562 - mae: 136.3562 - mse: 88769.1016 - val_loss: 961.8546 - val_mae: 961.8546 - val_mse: 2182614.7500
Epoch 67/100
16/16 [==============================] - 0s 7ms/step - loss: 137.7614 - mae: 137.7614 - mse: 89457.9297 - val_loss: 964.5009 - val_mae: 964.5009 - val_mse: 2186731.0000
Epoch 68/100
16/16 [==============================] - 0s 6ms/step - loss: 143.2756 - mae: 143.2756 - mse: 92493.5703 - val_loss: 1015.8807 - val_mae: 1015.8807 - val_mse: 2312432.7500
Epoch 69/100
12/16 [=====================>........] - ETA: 0s - loss: 142.3913 - mae: 142.3913 - mse: 93224.7500 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 47ms/step - loss: 139.2209 - mae: 139.2209 - mse: 89967.8750 - val_loss: 948.3441 - val_mae: 948.3441 - val_mse: 2154132.0000
Epoch 70/100
13/16 [=======================>......] - ETA: 0s - loss: 141.0554 - mae: 141.0554 - mse: 96066.3672INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 137.0358 - mae: 137.0358 - mse: 88676.4453 - val_loss: 944.1388 - val_mae: 944.1388 - val_mse: 2147949.0000
Epoch 71/100
16/16 [==============================] - 0s 7ms/step - loss: 137.4951 - mae: 137.4951 - mse: 89559.7891 - val_loss: 949.1281 - val_mae: 949.1281 - val_mse: 2152945.0000
Epoch 72/100
16/16 [==============================] - 0s 8ms/step - loss: 141.2230 - mae: 141.2230 - mse: 88946.1016 - val_loss: 988.2425 - val_mae: 988.2425 - val_mse: 2263519.2500
Epoch 73/100
16/16 [==============================] - 0s 7ms/step - loss: 141.0431 - mae: 141.0431 - mse: 88668.4922 - val_loss: 954.3595 - val_mae: 954.3595 - val_mse: 2159943.7500
Epoch 74/100
16/16 [==============================] - 0s 8ms/step - loss: 136.4250 - mae: 136.4250 - mse: 88413.8516 - val_loss: 952.0051 - val_mae: 952.0051 - val_mse: 2153421.5000
Epoch 75/100
14/16 [=========================>....] - ETA: 0s - loss: 145.3431 - mae: 145.3431 - mse: 96820.6875INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 141.5638 - mae: 141.5638 - mse: 92153.6172 - val_loss: 939.4951 - val_mae: 939.4951 - val_mse: 2129653.5000
Epoch 76/100
16/16 [==============================] - 0s 6ms/step - loss: 138.9685 - mae: 138.9685 - mse: 90930.5469 - val_loss: 947.6743 - val_mae: 947.6743 - val_mse: 2142410.2500
Epoch 77/100
16/16 [==============================] - 0s 7ms/step - loss: 135.7710 - mae: 135.7710 - mse: 88085.2969 - val_loss: 950.3411 - val_mae: 950.3411 - val_mse: 2164202.2500
Epoch 78/100
16/16 [==============================] - 0s 8ms/step - loss: 140.3474 - mae: 140.3474 - mse: 89952.8047 - val_loss: 957.8239 - val_mae: 957.8239 - val_mse: 2181345.5000
Epoch 79/100
14/16 [=========================>....] - ETA: 0s - loss: 138.0457 - mae: 138.0457 - mse: 86333.9297 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 136.9556 - mae: 136.9556 - mse: 86075.8984 - val_loss: 936.6367 - val_mae: 936.6367 - val_mse: 2124193.5000
Epoch 80/100
16/16 [==============================] - 0s 7ms/step - loss: 136.4273 - mae: 136.4273 - mse: 88075.7188 - val_loss: 965.5436 - val_mae: 965.5436 - val_mse: 2178103.7500
Epoch 81/100
16/16 [==============================] - 0s 7ms/step - loss: 134.7644 - mae: 134.7644 - mse: 87283.6094 - val_loss: 967.8107 - val_mae: 967.8107 - val_mse: 2184070.5000
Epoch 82/100
16/16 [==============================] - 0s 6ms/step - loss: 134.1701 - mae: 134.1701 - mse: 87228.1016 - val_loss: 957.0618 - val_mae: 957.0618 - val_mse: 2158057.7500
Epoch 83/100
16/16 [==============================] - 0s 7ms/step - loss: 135.6599 - mae: 135.6599 - mse: 88001.3828 - val_loss: 964.8870 - val_mae: 964.8870 - val_mse: 2197043.7500
Epoch 84/100
16/16 [==============================] - 0s 7ms/step - loss: 140.0131 - mae: 140.0131 - mse: 89544.5625 - val_loss: 1019.5710 - val_mae: 1019.5710 - val_mse: 2350208.0000
Epoch 85/100
13/16 [=======================>......] - ETA: 0s - loss: 138.3011 - mae: 138.3011 - mse: 92530.0938 INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 134.8940 - mae: 134.8940 - mse: 88015.3672 - val_loss: 936.4964 - val_mae: 936.4964 - val_mse: 2121844.5000
Epoch 86/100
16/16 [==============================] - 0s 6ms/step - loss: 135.6369 - mae: 135.6369 - mse: 88435.6406 - val_loss: 942.7554 - val_mae: 942.7554 - val_mse: 2138073.5000
Epoch 87/100
16/16 [==============================] - 0s 7ms/step - loss: 135.0664 - mae: 135.0664 - mse: 86687.9688 - val_loss: 964.7237 - val_mae: 964.7237 - val_mse: 2174232.2500
Epoch 88/100
16/16 [==============================] - 0s 8ms/step - loss: 136.0079 - mae: 136.0079 - mse: 86509.4453 - val_loss: 998.5038 - val_mae: 998.5038 - val_mse: 2259588.2500
Epoch 89/100
16/16 [==============================] - 0s 7ms/step - loss: 142.3703 - mae: 142.3703 - mse: 91063.0625 - val_loss: 959.3842 - val_mae: 959.3842 - val_mse: 2160678.5000
Epoch 90/100
16/16 [==============================] - 0s 6ms/step - loss: 144.7295 - mae: 144.7295 - mse: 92511.4453 - val_loss: 947.9583 - val_mae: 947.9583 - val_mse: 2150230.7500
Epoch 91/100
16/16 [==============================] - 0s 7ms/step - loss: 144.4417 - mae: 144.4417 - mse: 92074.1562 - val_loss: 979.2183 - val_mae: 979.2183 - val_mse: 2233493.5000
Epoch 92/100
16/16 [==============================] - 0s 7ms/step - loss: 139.8674 - mae: 139.8674 - mse: 91445.4375 - val_loss: 944.9947 - val_mae: 944.9947 - val_mse: 2129126.5000
Epoch 93/100
16/16 [==============================] - 0s 6ms/step - loss: 139.7284 - mae: 139.7284 - mse: 90204.4844 - val_loss: 1003.3015 - val_mae: 1003.3015 - val_mse: 2271156.5000
Epoch 94/100
16/16 [==============================] - 0s 7ms/step - loss: 142.8270 - mae: 142.8270 - mse: 89543.1953 - val_loss: 994.1068 - val_mae: 994.1068 - val_mse: 2247293.2500
Epoch 95/100
14/16 [=========================>....] - ETA: 0s - loss: 139.1342 - mae: 139.1342 - mse: 88337.8203INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 45ms/step - loss: 138.3101 - mae: 138.3101 - mse: 89367.2109 - val_loss: 934.3267 - val_mae: 934.3267 - val_mse: 2109674.0000
Epoch 96/100
16/16 [==============================] - 0s 6ms/step - loss: 134.0352 - mae: 134.0352 - mse: 85598.4219 - val_loss: 1023.6821 - val_mae: 1023.6821 - val_mse: 2360825.7500
Epoch 97/100
16/16 [==============================] - 0s 7ms/step - loss: 135.8769 - mae: 135.8769 - mse: 86870.6641 - val_loss: 938.7961 - val_mae: 938.7961 - val_mse: 2113813.5000
Epoch 98/100
14/16 [=========================>....] - ETA: 0s - loss: 135.1390 - mae: 135.1390 - mse: 88315.8906INFO:tensorflow:Assets written to: model_experiments/model_1_dense/assets
16/16 [==============================] - 1s 46ms/step - loss: 133.4204 - mae: 133.4204 - mse: 85778.6562 - val_loss: 933.0836 - val_mae: 933.0836 - val_mse: 2106266.2500
Epoch 99/100
16/16 [==============================] - 0s 8ms/step - loss: 132.9773 - mae: 132.9773 - mse: 85948.2109 - val_loss: 936.1680 - val_mae: 936.1680 - val_mse: 2108080.7500
Epoch 100/100
16/16 [==============================] - 0s 7ms/step - loss: 139.6587 - mae: 139.6587 - mse: 88771.3516 - val_loss: 1119.3230 - val_mae: 1119.3230 - val_mse: 2611163.2500naive_results{'mae': 924.30414,
'mape': 2.6047938,
'mase': 0.99834156,
'mse': 2055079.2,
'rmse': 1433.5548}Evaluate
# Evaluate
model_1.evaluate(test_windows, test_labels)16/16 [==============================] - 0s 3ms/step - loss: 1119.3231 - mae: 1119.3231 - mse: 2611163.5000
[1119.3231201171875, 1119.3231201171875, 2611163.5]# Load in best performing model
model_1 = tf.keras.models.load_model("model_experiments/model_1_dense/")
model_1.evaluate(test_windows, test_labels)16/16 [==============================] - 0s 3ms/step - loss: 933.0836 - mae: 933.0836 - mse: 2106266.2500
[933.0836181640625, 933.0836181640625, 2106266.25]Forecast on the test dataset
pred_1 = model_1.predict(x = test_windows)len(test_labels), len(pred_1)(509, 509)tf.keras.metrics.mean_absolute_error(tf.squeeze(test_labels), tf.squeeze(pred_1))<tf.Tensor: shape=(), dtype=float32, numpy=933.08374>model_1_results = evaluate_preds(y_true=tf.squeeze(test_labels), y_pred=tf.squeeze(pred_1))
model_1_results{'mae': 933.08374,
'mape': 2.6302116,
'mase': 1.0078244,
'mse': 2106266.2,
'rmse': 1451.2981}test_windows.shape(509, 7)len(test_labels)509offset = 300
plt.figure(figsize=(10, 7))
# Account for the test_window offset and index into test_labels to ensure correct plotting
plot_time_series(timesteps=X_test[1:], values=test_labels[:, 0], start=offset, label="Test_data")
plot_time_series(timesteps=X_test[1:], values=pred_1, start=offset, format="-", label="model_1_preds")def make_preds(model, input_data):
"""
make predictions and squeeze results
"""
forecast = model.predict(input_data)
return tf.squeeze(forecast)Model 2: Dense 30, 1
HORIZON = 1
WINDOW_SIZE = 30# Make windowed data
full_windows, full_labels = make_windows(np.array(prices), window_size=WINDOW_SIZE, horizon=HORIZON)
len(full_windows), len(full_labels)(2518, 2518)train_windows, test_windows, train_labels, test_labels = make_train_test_splits(full_windows, full_labels)
len(train_windows), len(test_windows), len(train_labels), len(test_labels)(2014, 504, 2014, 504)train_windows[:1], train_labels[:1](array([[324.71833, 332.45666, 336.585 , 346.775 , 344.81166, 343.065 ,
358.50166, 368.07666, 376.99666, 442.10666, 389.00333, 394.92666,
377.08 , 400.47166, 384.93 , 375.83333, 376.16 , 348.565 ,
352.42 , 356.8 , 367.83 , 381.20666, 375.498 , 372.819 ,
368.983 , 375.44 , 376.851 , 376.13218, 377.862 , 382.452 ]]),
array([[376.382]]))tf.random.set_seed(42)
# Create model
model_2 = tf.keras.Sequential([
layers.Dense(128, activation='relu'),
layers.Dense(HORIZON)
], name='model_2_dense')
model_2.compile(loss='mae',
optimizer=tf.keras.optimizers.Adam(),
metrics=['mae'])
model_2.fit(train_windows, train_labels,
epochs=100,
batch_size=128,
verbose=0,
validation_data=(test_windows,test_labels),
callbacks=[create_model_checkpoint(model_name=model_2.name)])INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_2_dense/assets
<keras.callbacks.History at 0x7f9e7517b5d0>Evaluate
model_2 = tf.keras.models.load_model("model_experiments/model_2_dense/")
model_2.evaluate(test_windows, test_labels)16/16 [==============================] - 0s 3ms/step - loss: 990.3993 - mae: 990.3993
[990.3992919921875, 990.3992919921875]# Get forecast predictions
model_2_preds = make_preds(model_2,
input_data=test_windows)model_2_results = evaluate_preds(y_true=tf.squeeze(test_labels), y_pred=model_2_preds)
model_2_results{'mae': 990.3992,
'mape': 2.788899,
'mase': 1.059855,
'mse': 2314995.8,
'rmse': 1521.511}Didn't perform as well as model 1 - could totally run a for loop with different window sizes
offset = 300
plt.figure(figsize=(10, 7))
# Account for the test_window offset and index into test_labels to ensure correct plotting
plot_time_series(timesteps=X_test[-len(test_windows):], values=test_labels[:, 0], start=offset, label="Test_data")
plot_time_series(timesteps=X_test[-len(test_windows):], values=model_2_preds, start=offset, format="-", label="model_2_preds")Model 3: Dense 30, 7
HORIZON = 7
WINDOW_SIZE = 30# Make windowed data
full_windows, full_labels = make_windows(np.array(prices), window_size=WINDOW_SIZE, horizon=HORIZON)
len(full_windows), len(full_labels)(2512, 2512)train_windows, test_windows, train_labels, test_labels = make_train_test_splits(full_windows, full_labels)
len(train_windows), len(test_windows), len(train_labels), len(test_labels)(2009, 503, 2009, 503)train_windows[:1], train_labels[:1](array([[324.71833, 332.45666, 336.585 , 346.775 , 344.81166, 343.065 ,
358.50166, 368.07666, 376.99666, 442.10666, 389.00333, 394.92666,
377.08 , 400.47166, 384.93 , 375.83333, 376.16 , 348.565 ,
352.42 , 356.8 , 367.83 , 381.20666, 375.498 , 372.819 ,
368.983 , 375.44 , 376.851 , 376.13218, 377.862 , 382.452 ]]),
array([[376.382 , 368.3618, 373.414 , 373.744 , 373.155 , 357.884 ,
348.898 ]]))tf.random.set_seed(42)
# Create model
model_3 = tf.keras.Sequential([
layers.Dense(128, activation='relu'),
layers.Dense(HORIZON)
], name='model_3_dense')
model_3.compile(loss='mae',
optimizer=tf.keras.optimizers.Adam(),
metrics=['mae'])
model_3.fit(train_windows, train_labels,
epochs=100,
batch_size=128,
verbose=0,
validation_data=(test_windows,test_labels),
callbacks=[create_model_checkpoint(model_name=model_3.name)])INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
INFO:tensorflow:Assets written to: model_experiments/model_3_dense/assets
<keras.callbacks.History at 0x7f9e7807a550>Evaluate
model_3 = tf.keras.models.load_model("model_experiments/model_3_dense/")
model_3.evaluate(test_windows, test_labels)16/16 [==============================] - 0s 3ms/step - loss: 2047.8856 - mae: 2047.8856
[2047.8856201171875, 2047.8856201171875]# Get forecast predictions
model_3_preds = make_preds(model_3,
input_data=test_windows)test_labels.shape, model_3_preds.shape((503, 7), TensorShape([503, 7]))model_3_results = evaluate_preds(y_true=test_labels, y_pred=model_3_preds)
model_3_results{'mae': array([ 198.34473 , 344.33746 , 227.27246 , 262.179 ,
258.46304 , 376.8178 , 545.6133 , 667.0472 ,
259.5064 , 273.26535 , 182.20229 , 238.57227 ,
204.28 , 205.19225 , 208.5664 , 179.56445 ,
174.07004 , 172.27469 , 249.48772 , 324.3662 ,
658.38574 , 401.90836 , 361.08316 , 192.99959 ,
115.62849 , 141.73856 , 71.96303 , 117.711914,
163.44824 , 157.82701 , 142.01186 , 56.97977 ,
125.28209 , 53.725445, 156.06584 , 256.89648 ,
180.32156 , 225.35645 , 164.63184 , 170.1988 ,
198.65291 , 231.8591 , 148.79256 , 173.79729 ,
188.39244 , 231.84334 , 245.54074 , 539.3613 ,
627.8876 , 803.3146 , 872.73175 , 1053.4825 ,
1283.8585 , 1223.984 , 488.4527 , 205.73186 ,
224.60506 , 190.05692 , 327.95117 , 450.90527 ,
178.5466 , 104.73912 , 220.82799 , 205.536 ,
348.13364 , 366.0572 , 304.0113 , 318.61996 ,
297.5872 , 206.25642 , 231.82547 , 187.90053 ,
230.45508 , 157.01967 , 397.63492 , 475.86063 ,
734.29535 , 459.06027 , 455.41336 , 338.95035 ,
322.3609 , 535.4078 , 407.27106 , 171.99791 ,
191.34445 , 292.96457 , 421.27762 , 586.0368 ,
832.2006 , 1162.0234 , 1574.3854 , 1485.5608 ,
945.8036 , 680.7335 , 138.21289 , 200.53474 ,
109.816124, 283.66574 , 323.6511 , 337.11383 ,
402.0537 , 310.8725 , 320.70215 , 240.15736 ,
197.00461 , 297.49875 , 482.01813 , 426.15765 ,
548.6476 , 417.54785 , 185.7867 , 143.07674 ,
208.96596 , 129.89453 , 214.2302 , 245.90096 ,
348.01813 , 340.75992 , 341.37766 , 242.69824 ,
151.43416 , 154.83594 , 262.8623 , 344.3146 ,
395.17188 , 461.6091 , 537.1222 , 499.5226 ,
372.60797 , 81.39886 , 174.20619 , 387.56754 ,
257.13422 , 441.0887 , 614.24384 , 706.9044 ,
922.39856 , 1084.5853 , 1090.9434 , 1091.9379 ,
360.91254 , 261.1013 , 248.30984 , 255.09724 ,
235.00168 , 305.48145 , 233.3086 , 142.6183 ,
332.70453 , 619.629 , 711.79034 , 774.89636 ,
994.37823 , 1076.2625 , 1027.3513 , 425.54617 ,
482.83762 , 456.38254 , 345.99707 , 469.71652 ,
756.25714 , 708.012 , 852.3648 , 989.6528 ,
1089.5575 , 1644.5303 , 1342.9972 , 960.204 ,
410.33176 , 498.6638 , 713.6239 , 947.9819 ,
1114.9084 , 1148.0881 , 1178.8998 , 974.8368 ,
741.6417 , 1038.5848 , 1141.1635 , 1209.2333 ,
200.67969 , 385.05804 , 674.4183 , 1059.7338 ,
663.9403 , 821.9721 , 786.01227 , 603.0148 ,
616.0787 , 721.7525 , 1301.8822 , 2269.0444 ,
2594.9653 , 2808.096 , 2998.5503 , 3234.464 ,
2263.4448 , 902.9032 , 520.41296 , 982.93164 ,
1308.087 , 1457.373 , 1629.0212 , 2366.7136 ,
2965.1985 , 2699.2488 , 2279.3882 , 2441.1575 ,
2570.7456 , 3510.4832 , 3328.179 , 4058.5444 ,
5065.7656 , 4623.383 , 4086.5723 , 4456.236 ,
3156.0378 , 2677.4397 , 2276.2935 , 4508.8906 ,
5825.0244 , 5611.3755 , 1661.7706 , 845.726 ,
867.4258 , 2666.938 , 3371.8198 , 3614.1462 ,
3292.4792 , 3546.8093 , 3638.3484 , 4005.7156 ,
1541.0896 , 1302.4001 , 1217.4744 , 1369.2963 ,
1229.257 , 2257.268 , 3417.4084 , 2871.5723 ,
2480.7087 , 2513.466 , 2763.73 , 4625.2593 ,
4782.5625 , 4349.855 , 4666.804 , 6162.767 ,
5578.535 , 7692.2266 , 4323.515 , 1063.9012 ,
1545.2483 , 1439.0078 , 2370.869 , 2822.2036 ,
3160.3398 , 2985.384 , 3794.0972 , 2580.2014 ,
3335.1987 , 5645.238 , 7315.0405 , 9767.505 ,
9322.964 , 5195.8384 , 3231.231 , 2128.3137 ,
1312.2494 , 1766.7611 , 3287.6228 , 2306.319 ,
2333.2136 , 2844.9722 , 3699.7417 , 4800.5845 ,
5849.413 , 5997.942 , 5407.9844 , 4060.5083 ,
2332.4036 , 1375.0045 , 1264.4559 , 2245.563 ,
4036.7266 , 2309.6167 , 2903.303 , 3510.4856 ,
4537.8877 , 6121.304 , 5190.4272 , 5551.523 ,
1387.3075 , 2136.9058 , 2498.737 , 3573.7214 ,
4086.5503 , 2073.905 , 2189.5073 , 1139.2963 ,
766.92914 , 1152.5441 , 1436.0898 , 1400.1356 ,
1102.2277 , 1371.721 , 1362.317 , 1386.7812 ,
1805.6608 , 2493.904 , 2941.188 , 1845.8928 ,
2140.1858 , 2488.168 , 3852.1536 , 5994.5693 ,
7607.3755 , 9059.399 , 8574.132 , 6773.6094 ,
5189.7554 , 4926.047 , 3062.0664 , 2121.145 ,
2596.2786 , 3622.7144 , 5636.5815 , 3238.9714 ,
1651.4006 , 1597.1512 , 1834.8834 , 1492.7494 ,
1455.7266 , 1486.4135 , 2031.476 , 1812.0419 ,
1687.6584 , 2623.7695 , 4156.2114 , 6564.0327 ,
8155.9546 , 8803.975 , 11318.192 , 9851.561 ,
8017.621 , 9746.167 , 8563.351 , 8970.545 ,
6508.9243 , 6760.962 , 3445.9973 , 3126.2214 ,
2015.6803 , 2641.9675 , 2195.7456 , 2624.8918 ,
3300.6958 , 3381.625 , 4271.9404 , 1961.774 ,
1496.9408 , 1526.1083 , 1528.4347 , 2068.452 ,
2956.1418 , 2970.3416 , 3116.106 , 1698.2333 ,
1426.6139 , 1770.2355 , 3493.432 , 1837.5988 ,
2148.6606 , 1192.2589 , 2073.352 , 1410.8516 ,
3233.982 , 5003.247 , 5502.7495 , 5344.1323 ,
3799.816 , 3266.5054 , 3503.0896 , 1119.5798 ,
1375.6322 , 1966.7006 , 1187.0098 , 1500.8019 ,
1921.2517 , 999.07367 , 920.08875 , 1811.0681 ,
1668.466 , 1165.034 , 690.83875 , 680.68304 ,
782.0067 , 1382.4303 , 1013.7397 , 1861.8976 ,
1206.8392 , 1241.0117 , 1980.798 , 2791.3293 ,
2935.9858 , 1723.6033 , 2224.018 , 1247.7949 ,
1025.0469 , 1316.5421 , 1534.9941 , 2456.0562 ,
3958.0015 , 4228.286 , 4914.943 , 5676.0522 ,
4785.6562 , 4855.253 , 3566.6394 , 1365.601 ,
757.0882 , 1197.4933 , 1638.3522 , 2444.0552 ,
1892.9219 , 2323.957 , 3070.6484 , 3792.4792 ,
3147.5178 , 2337.3125 , 1435.8783 , 1191.3075 ,
1024.8622 , 1149.8427 , 1211.0117 , 925.2601 ,
1683.8353 , 1875.5502 , 1942.6495 , 1976.1786 ,
2096.7969 , 2742.3337 , 2529.3784 , 528.9196 ,
1278.7181 , 2170.6467 , 2400.2388 , 2107.6958 ,
1438.7428 , 956.12054 , 1082.9531 , 997.81024 ,
1333.159 , 1813.6088 , 2040.3822 , 2176.8416 ,
2053.4414 , 2818.4353 , 3990.8303 , 5517.2505 ,
7213.7144 , 4508.144 , 2198.616 , 1504.3588 ,
1150.6914 , 1681.9755 , 1525.1842 , 2467.0864 ,
1753.2338 , 2303.058 , 3458.7266 , 3915.9727 ,
4562.4736 , 5186.7764 , 3279.924 , 1160.4564 ,
977.5156 , 1440.6864 , 1489.6523 , 2251.639 ,
2537.5564 , 3375.6067 , 4399.5317 , 6303.0083 ,
6276.4653 , 4917.6226 , 4017.4453 , 3827.3572 ,
5202.6626 , 5032.5195 , 1562.3047 , 1523.0033 ,
1910.0206 , 1785.5391 , 2249.7036 , 2012.0223 ,
2554.5286 , 2947.3333 , 2809.2305 , 1600.462 ,
1735.24 , 1548.6914 , 1886.9342 ], dtype=float32),
'mape': array([ 2.0409105 , 3.5966687 , 2.3846035 , 2.7481534 , 2.747307 ,
4.0041947 , 5.8015065 , 7.0797234 , 2.7551498 , 2.9169676 ,
1.941608 , 2.546255 , 2.1833947 , 2.1878464 , 2.225683 ,
1.9068556 , 1.8761367 , 1.845595 , 2.6970382 , 3.5297573 ,
7.1855345 , 4.391867 , 3.9610407 , 2.1175442 , 1.2631248 ,
1.5533594 , 0.789953 , 1.293295 , 1.7913997 , 1.7116746 ,
1.5376999 , 0.6098182 , 1.346609 , 0.5801435 , 1.6797042 ,
2.7794037 , 1.9568933 , 2.4455276 , 1.7867798 , 1.8507617 ,
2.1655111 , 2.5257678 , 1.6192667 , 1.8591682 , 2.0007024 ,
2.4336922 , 2.5278356 , 5.2029967 , 5.86003 , 7.350724 ,
7.8994884 , 9.449211 , 11.39213 , 10.823228 , 4.2928815 ,
1.7929399 , 1.9738612 , 1.6804116 , 2.8660011 , 3.9693837 ,
1.5590972 , 0.90401906, 1.9032898 , 1.7755817 , 3.0038102 ,
3.152892 , 2.6241322 , 2.7546659 , 2.5460298 , 1.713315 ,
1.9174682 , 1.5480202 , 1.9443216 , 1.3291662 , 3.3576999 ,
4.0618033 , 6.3079076 , 3.975036 , 3.962695 , 2.9549134 ,
2.8215885 , 4.6779375 , 3.5526416 , 1.4986718 , 1.6417801 ,
2.5823085 , 3.8424094 , 5.522741 , 7.9753056 , 11.12138 ,
15.235226 , 14.461771 , 9.256595 , 6.6554894 , 1.3533186 ,
1.9536266 , 1.0562544 , 2.6833112 , 3.0139103 , 3.1104934 ,
3.7169588 , 2.841987 , 2.934643 , 2.2072937 , 1.833329 ,
2.8234065 , 4.5616555 , 4.038537 , 5.213386 , 3.9686978 ,
1.7763933 , 1.3485821 , 1.9433869 , 1.2094533 , 2.006067 ,
2.3066087 , 3.2691407 , 3.2020674 , 3.2180798 , 2.2902763 ,
1.4234982 , 1.4335357 , 2.3897367 , 3.1134093 , 3.4988203 ,
4.040433 , 4.7129726 , 4.3660583 , 3.2488909 , 0.70869786,
1.5187101 , 3.3927531 , 2.2474601 , 3.6401849 , 4.8955703 ,
5.5116935 , 7.110484 , 8.36204 , 8.390585 , 8.274902 ,
2.706628 , 1.9726504 , 1.87443 , 1.9121913 , 1.7235445 ,
2.2373135 , 1.7240627 , 1.0321987 , 2.2308378 , 4.138693 ,
4.740137 , 5.116083 , 6.509969 , 7.047697 , 6.640906 ,
2.713436 , 3.1119843 , 2.8711429 , 2.1789086 , 2.8805835 ,
4.513125 , 4.1019797 , 4.924124 , 5.625858 , 6.0197773 ,
9.039097 , 7.318673 , 5.1672754 , 2.1930883 , 2.75807 ,
4.0843906 , 5.4146433 , 6.330515 , 6.518898 , 6.7290173 ,
5.584917 , 4.01358 , 5.426522 , 5.976067 , 6.2946773 ,
1.0542319 , 2.0336823 , 3.5821457 , 5.654707 , 3.5833917 ,
4.420351 , 4.2376714 , 3.2628787 , 3.3360868 , 3.6832948 ,
6.0888963 , 10.487547 , 11.508254 , 12.206615 , 12.961777 ,
13.961157 , 9.694232 , 3.8480554 , 2.170022 , 4.0295506 ,
5.3007073 , 5.6936 , 6.2191463 , 8.750816 , 10.779581 ,
9.5962305 , 7.6114736 , 7.8555236 , 8.174191 , 10.909031 ,
9.837456 , 11.323862 , 13.5916815 , 11.981682 , 10.50922 ,
11.431423 , 8.033039 , 6.9882803 , 6.3726015 , 12.515625 ,
16.24484 , 15.664029 , 4.616887 , 2.2658567 , 2.3547854 ,
7.8124776 , 10.034515 , 10.980551 , 10.185558 , 10.919814 ,
11.356761 , 12.605284 , 4.8053894 , 4.06745 , 3.692954 ,
4.093094 , 3.729272 , 6.5538034 , 9.735404 , 7.9521775 ,
6.7950597 , 6.7391343 , 7.2153134 , 11.548952 , 11.169111 ,
9.77751 , 10.211084 , 13.406171 , 11.977131 , 16.352287 ,
9.058758 , 2.2204459 , 3.0962837 , 2.888864 , 4.5459604 ,
5.3408623 , 5.8279524 , 5.3900776 , 7.02347 , 4.9445014 ,
6.46274 , 11.718033 , 15.263276 , 20.734089 , 19.76489 ,
11.09307 , 6.9334364 , 4.5595746 , 2.7395282 , 3.5962372 ,
6.633214 , 4.583344 , 4.4937057 , 5.2629957 , 6.7527466 ,
8.580757 , 10.195822 , 10.2611 , 9.275844 , 6.9551296 ,
3.9324408 , 2.3676229 , 2.1546001 , 3.910556 , 7.001995 ,
4.1098228 , 5.189389 , 6.4119377 , 8.408363 , 11.330194 ,
9.620076 , 10.289664 , 2.611399 , 3.9229493 , 4.3717723 ,
6.1837955 , 7.018744 , 3.5511944 , 3.748101 , 1.9426886 ,
1.3111055 , 1.9951535 , 2.4953012 , 2.4208422 , 1.8986394 ,
2.3767948 , 2.371073 , 2.3314166 , 2.9234915 , 3.9891224 ,
4.7085977 , 2.9526165 , 3.5060458 , 4.1939425 , 6.6803446 ,
10.563861 , 13.716027 , 16.765244 , 16.172094 , 13.179675 ,
10.133285 , 9.630142 , 6.029118 , 4.172738 , 4.8675203 ,
6.547346 , 10.114635 , 5.718496 , 2.918504 , 2.8566 ,
3.2145638 , 2.6727214 , 2.6089797 , 2.6572514 , 3.6168218 ,
3.134128 , 3.0801384 , 4.978535 , 8.034607 , 13.003127 ,
16.612455 , 18.73364 , 24.517635 , 22.514673 , 19.05419 ,
23.992016 , 21.67174 , 23.663872 , 17.450253 , 18.179699 ,
9.31843 , 8.521563 , 5.6955543 , 7.45401 , 6.1200705 ,
7.333215 , 9.205043 , 9.48326 , 11.911655 , 5.3824954 ,
4.079406 , 4.143833 , 4.2748485 , 5.9612446 , 8.4120655 ,
8.549851 , 8.900564 , 4.8705006 , 3.9907281 , 4.6707864 ,
9.085865 , 4.724092 , 5.5293584 , 3.1803977 , 5.3039885 ,
3.7732177 , 9.304856 , 14.697075 , 16.296993 , 15.8320055 ,
11.6358 , 10.115356 , 10.814154 , 3.4606152 , 4.0400305 ,
5.7444496 , 3.590705 , 4.322302 , 5.539883 , 2.8603876 ,
2.6533535 , 5.3358693 , 4.9170675 , 3.4665222 , 1.996088 ,
1.963587 , 2.33209 , 4.142221 , 3.0865688 , 5.6164494 ,
3.6786728 , 3.8498755 , 6.1602454 , 8.731968 , 9.264205 ,
5.5800796 , 7.1735873 , 4.067086 , 3.2938998 , 4.1956086 ,
4.6731973 , 7.141753 , 11.036282 , 11.313997 , 12.911396 ,
14.532192 , 11.928548 , 12.05362 , 8.829873 , 3.3527904 ,
1.8891755 , 3.0242498 , 4.1290917 , 6.1564865 , 4.7068853 ,
5.413719 , 6.8975077 , 8.471508 , 7.0060015 , 5.0862875 ,
3.0937424 , 2.575307 , 2.2411168 , 2.562805 , 2.6790764 ,
1.9999385 , 3.6560953 , 4.0793138 , 4.226768 , 4.1888237 ,
4.299208 , 5.620482 , 5.168273 , 1.0873759 , 2.6516943 ,
4.4800615 , 4.9738936 , 4.4172287 , 2.9971297 , 1.9806389 ,
2.2549436 , 2.0957136 , 2.7528627 , 3.5897672 , 4.0447927 ,
4.3221755 , 4.2354136 , 6.0061493 , 8.639593 , 12.0095215 ,
15.788022 , 9.869997 , 4.8399973 , 3.3010283 , 2.4443154 ,
3.551695 , 3.203864 , 5.224407 , 3.9855976 , 5.3052225 ,
7.933607 , 9.066291 , 10.634617 , 12.099351 , 7.6456466 ,
2.7139003 , 2.2784884 , 3.4095955 , 3.376904 , 4.923179 ,
5.473342 , 7.1610146 , 9.009227 , 12.562357 , 12.209492 ,
9.295086 , 7.428875 , 7.028468 , 9.477975 , 9.069722 ,
2.750283 , 2.6848319 , 3.2936757 , 2.9969134 , 3.7589424 ,
3.3451118 , 4.171559 , 4.656481 , 4.4283633 , 2.5137007 ,
2.753452 , 2.4734402 , 3.0146499 ], dtype=float32),
'mase': 2.2066739,
'mse': array([6.02635742e+04, 1.61393453e+05, 7.66824219e+04, 8.52572656e+04,
1.13504102e+05, 1.91161656e+05, 3.52221188e+05, 4.52515719e+05,
9.44122656e+04, 1.11110570e+05, 3.93382891e+04, 7.04189688e+04,
6.02438984e+04, 5.82169727e+04, 5.25739766e+04, 3.70784336e+04,
5.71524961e+04, 3.93154023e+04, 8.33688047e+04, 1.16277930e+05,
4.46706344e+05, 1.82590453e+05, 1.56936281e+05, 5.14491250e+04,
1.52098301e+04, 2.77714004e+04, 7.34899609e+03, 2.67268809e+04,
4.24213828e+04, 3.64512148e+04, 3.32481367e+04, 6.56644922e+03,
2.12645996e+04, 6.31993896e+03, 3.04794336e+04, 6.98643984e+04,
4.47920625e+04, 6.16715625e+04, 3.04689883e+04, 3.53427383e+04,
4.81881797e+04, 6.59483125e+04, 3.56257539e+04, 3.42030664e+04,
4.40084180e+04, 7.06022969e+04, 9.28485391e+04, 5.90253188e+05,
8.37575062e+05, 1.14411600e+06, 1.18824150e+06, 1.48737388e+06,
1.88306175e+06, 1.58527488e+06, 3.05822250e+05, 7.83743203e+04,
5.43062812e+04, 5.92518047e+04, 1.43609219e+05, 2.72906625e+05,
5.93235273e+04, 1.57208438e+04, 6.76435000e+04, 5.95815117e+04,
1.49439344e+05, 1.67541000e+05, 1.39870672e+05, 1.71601500e+05,
1.29803352e+05, 6.79795625e+04, 9.28080859e+04, 7.65459844e+04,
6.59152109e+04, 3.62826016e+04, 1.62309609e+05, 2.53965250e+05,
5.59725250e+05, 2.90691312e+05, 2.73631469e+05, 1.52240047e+05,
1.44111344e+05, 3.52751719e+05, 1.93066484e+05, 6.11461680e+04,
6.82383750e+04, 1.44429750e+05, 3.28503062e+05, 5.91617562e+05,
1.04241488e+06, 1.63229375e+06, 2.73104000e+06, 2.29322350e+06,
9.71668438e+05, 4.84436781e+05, 3.29003008e+04, 5.46350508e+04,
2.30860488e+04, 1.12870008e+05, 1.55262766e+05, 1.75493266e+05,
1.89354516e+05, 1.28518211e+05, 1.14358461e+05, 6.27497773e+04,
5.28453047e+04, 1.34363516e+05, 2.93759625e+05, 2.22686891e+05,
3.83397938e+05, 2.32191719e+05, 8.58148828e+04, 2.80607012e+04,
5.55600430e+04, 2.56328223e+04, 5.79876523e+04, 7.08189922e+04,
1.42848031e+05, 1.38269891e+05, 1.51377297e+05, 8.24241719e+04,
3.83368086e+04, 2.93569199e+04, 1.00090859e+05, 1.57954031e+05,
2.31765500e+05, 3.22995938e+05, 3.44943562e+05, 2.80493000e+05,
1.62750281e+05, 8.85123730e+03, 3.78679570e+04, 1.82778328e+05,
9.81686016e+04, 3.35493219e+05, 6.61173688e+05, 8.70286062e+05,
1.31481338e+06, 1.47086025e+06, 1.29646488e+06, 1.26271975e+06,
1.98772984e+05, 8.17516797e+04, 1.02111164e+05, 8.55700000e+04,
8.01586953e+04, 1.20107883e+05, 8.44372266e+04, 3.29432695e+04,
3.80255406e+05, 7.44121875e+05, 8.93110375e+05, 8.91499688e+05,
1.30313000e+06, 1.25025062e+06, 1.16127575e+06, 3.21425000e+05,
2.79804469e+05, 2.48583141e+05, 1.72024266e+05, 3.31551219e+05,
8.84960625e+05, 8.88029500e+05, 9.86135062e+05, 1.32494288e+06,
1.62982062e+06, 3.07474275e+06, 1.97816200e+06, 1.07651500e+06,
2.24665625e+05, 4.33981531e+05, 1.06579475e+06, 1.53990962e+06,
1.71817088e+06, 1.91520788e+06, 2.27477275e+06, 1.72612575e+06,
6.17495062e+05, 1.49849262e+06, 1.40931712e+06, 1.65482712e+06,
5.61167148e+04, 2.07506922e+05, 5.91383625e+05, 1.22947000e+06,
5.77613875e+05, 7.91609688e+05, 7.65791688e+05, 5.61649875e+05,
6.15332875e+05, 9.85750562e+05, 3.65632075e+06, 7.78465450e+06,
1.03309940e+07, 1.11687500e+07, 1.10862650e+07, 1.08677550e+07,
5.25551450e+06, 9.20579375e+05, 5.17527312e+05, 1.38823088e+06,
2.16241025e+06, 2.88264975e+06, 3.68373450e+06, 7.54384650e+06,
1.03672720e+07, 8.58287300e+06, 8.87176100e+06, 9.79217100e+06,
1.00405920e+07, 1.61764100e+07, 1.67255470e+07, 2.55444100e+07,
3.66071440e+07, 3.09434220e+07, 2.25679640e+07, 2.57305700e+07,
1.42721990e+07, 7.96055450e+06, 8.93561400e+06, 2.51467660e+07,
3.84612560e+07, 3.36552240e+07, 3.44210850e+06, 1.41157600e+06,
1.07750412e+06, 9.29582900e+06, 1.50594870e+07, 1.79308420e+07,
1.55374210e+07, 1.41847140e+07, 1.58231540e+07, 1.68342580e+07,
3.02586375e+06, 2.32551175e+06, 2.24363950e+06, 2.35121725e+06,
2.40093825e+06, 6.62534100e+06, 1.36296910e+07, 1.16804400e+07,
8.29417200e+06, 7.87098150e+06, 1.03223500e+07, 2.84853420e+07,
3.69317240e+07, 3.31269120e+07, 3.46628520e+07, 4.75282880e+07,
3.70332680e+07, 6.05394640e+07, 2.08697240e+07, 1.68151112e+06,
4.64034750e+06, 3.61903200e+06, 8.91104600e+06, 1.08811700e+07,
1.37434800e+07, 1.44052970e+07, 1.72384320e+07, 9.30735000e+06,
1.26587110e+07, 4.72894440e+07, 6.52601080e+07, 1.07295224e+08,
8.88539040e+07, 3.06110720e+07, 1.43460740e+07, 6.81347250e+06,
2.50146275e+06, 4.12097750e+06, 1.21763830e+07, 7.79444900e+06,
8.97457400e+06, 1.60861310e+07, 2.02304680e+07, 3.30884060e+07,
4.56147480e+07, 4.64308280e+07, 3.31400140e+07, 1.81215660e+07,
8.68264600e+06, 2.71753625e+06, 2.43745000e+06, 6.93750150e+06,
1.72550740e+07, 1.02480550e+07, 1.38872460e+07, 1.91899540e+07,
2.90499580e+07, 4.43009680e+07, 3.16381940e+07, 3.32760640e+07,
4.68822650e+06, 6.54313900e+06, 8.00927100e+06, 1.52268590e+07,
1.89474360e+07, 5.28554700e+06, 5.08103350e+06, 1.78393712e+06,
7.09489062e+05, 1.69709375e+06, 3.18260925e+06, 2.26034575e+06,
1.41440575e+06, 2.81559250e+06, 4.05267000e+06, 3.01963125e+06,
6.00106750e+06, 1.05654980e+07, 1.21671280e+07, 4.89837700e+06,
6.10186750e+06, 8.90079400e+06, 2.21908380e+07, 4.64206600e+07,
6.80947280e+07, 9.32291760e+07, 8.05212240e+07, 5.48445880e+07,
3.24067320e+07, 2.86741480e+07, 1.30346490e+07, 6.82166550e+06,
8.55098100e+06, 1.65672780e+07, 3.33959980e+07, 1.29440640e+07,
3.93819575e+06, 2.81149275e+06, 4.46547100e+06, 3.18340975e+06,
3.94254750e+06, 3.88937425e+06, 4.92533700e+06, 5.61514050e+06,
5.83295950e+06, 1.26648300e+07, 2.75646240e+07, 6.01588560e+07,
8.31906560e+07, 9.90140560e+07, 1.41365904e+08, 1.15172344e+08,
8.21034720e+07, 1.10768136e+08, 8.38470560e+07, 9.06310560e+07,
4.93695440e+07, 4.92976200e+07, 1.39268340e+07, 1.29656780e+07,
8.83610000e+06, 1.02600180e+07, 6.90018050e+06, 9.72071300e+06,
1.30933750e+07, 1.45546770e+07, 2.08824920e+07, 5.10763150e+06,
3.28746925e+06, 2.72773475e+06, 3.11880925e+06, 6.40201350e+06,
1.00027050e+07, 1.18081870e+07, 1.13493510e+07, 4.44851200e+06,
2.96390225e+06, 4.52715450e+06, 1.37967440e+07, 4.94238350e+06,
5.87400000e+06, 2.01933662e+06, 6.01318400e+06, 2.52787150e+06,
1.42805990e+07, 3.09676020e+07, 3.47320960e+07, 3.02570240e+07,
1.94623740e+07, 1.32472960e+07, 1.39777670e+07, 2.25988025e+06,
2.31521125e+06, 4.35876650e+06, 2.18231850e+06, 3.46242350e+06,
4.47223350e+06, 1.30399862e+06, 1.26344325e+06, 4.22225650e+06,
3.46305450e+06, 1.85592062e+06, 7.90947000e+05, 9.38050375e+05,
8.25064688e+05, 2.43869325e+06, 2.01115288e+06, 3.84190175e+06,
2.38604625e+06, 2.28371450e+06, 4.69497850e+06, 8.66315700e+06,
9.48999400e+06, 4.20857750e+06, 5.66952850e+06, 2.60072575e+06,
1.59580425e+06, 2.66830275e+06, 3.30068575e+06, 9.12381300e+06,
2.15875740e+07, 2.59771740e+07, 3.10645800e+07, 3.96665040e+07,
2.83818580e+07, 2.59971240e+07, 1.35758290e+07, 2.63723250e+06,
8.90591438e+05, 1.97722288e+06, 3.71191000e+06, 8.47786500e+06,
5.33029700e+06, 6.26918050e+06, 1.31502110e+07, 1.64368500e+07,
1.08721780e+07, 7.07728050e+06, 2.79165250e+06, 1.81114888e+06,
1.73915475e+06, 3.11838325e+06, 2.10635725e+06, 1.47970025e+06,
3.64875575e+06, 5.01025250e+06, 5.61969850e+06, 4.19229775e+06,
5.52940950e+06, 7.95149050e+06, 7.09606000e+06, 5.39287625e+05,
2.46630825e+06, 5.64409700e+06, 6.75420950e+06, 6.48041500e+06,
2.25266225e+06, 1.09417875e+06, 1.85317012e+06, 2.48694725e+06,
2.95841725e+06, 4.05653325e+06, 5.76684000e+06, 6.19136900e+06,
5.19941050e+06, 1.08823590e+07, 2.03956640e+07, 3.48016080e+07,
5.31685400e+07, 2.18782780e+07, 7.05036700e+06, 3.43575350e+06,
1.82266438e+06, 3.48586150e+06, 3.17687875e+06, 6.42254950e+06,
5.62411300e+06, 8.74643400e+06, 1.67537990e+07, 1.99799860e+07,
2.50424640e+07, 2.94314040e+07, 1.14567220e+07, 2.18096375e+06,
1.28479362e+06, 3.01544525e+06, 4.06863175e+06, 7.66859600e+06,
9.96111700e+06, 1.50847790e+07, 2.65057800e+07, 4.72068680e+07,
4.61459560e+07, 3.31260300e+07, 2.37195220e+07, 1.99815940e+07,
2.95850780e+07, 2.79226580e+07, 4.80961950e+06, 3.00543825e+06,
4.99730550e+06, 5.79615950e+06, 6.89437600e+06, 4.72613350e+06,
7.57346550e+06, 1.30055040e+07, 1.11136730e+07, 4.45417750e+06,
4.37515100e+06, 3.48337075e+06, 4.15199825e+06], dtype=float32),
'rmse': array([ 245.4864 , 401.73804, 276.91592, 291.98846, 336.90372,
437.22037, 593.48224, 672.69293, 307.26578, 333.33252,
198.33882, 265.36572, 245.44633, 241.28194, 229.29015,
192.55762, 239.06589, 198.28113, 288.73657, 340.9955 ,
668.36096, 427.30606, 396.15182, 226.824 , 123.32815,
166.64754, 85.72629, 163.48358, 205.96454, 190.92203,
182.34071, 81.03363, 145.82387, 79.49805, 174.5836 ,
264.31876, 211.64134, 248.33762, 174.5537 , 187.99666,
219.51805, 256.80402, 188.74785, 184.94073, 209.78185,
265.71094, 304.7106 , 768.27936, 915.1912 , 1069.6337 ,
1090.065 , 1219.5795 , 1372.247 , 1259.077 , 553.012 ,
279.95416, 233.0371 , 243.41695, 378.95804, 522.40466,
243.56422, 125.38278, 260.08365, 244.09326, 386.57385,
409.31775, 373.9929 , 414.24814, 360.28232, 260.7289 ,
304.64423, 276.66946, 256.73956, 190.47993, 402.87665,
503.94965, 748.1479 , 539.15796, 523.09796, 390.17947,
379.61996, 593.9291 , 439.3933 , 247.27753, 261.22476,
380.03915, 573.15186, 769.1668 , 1020.9872 , 1277.6125 ,
1652.5858 , 1514.3392 , 985.7324 , 696.01495, 181.3844 ,
233.74142, 151.94093, 335.9613 , 394.034 , 418.9192 ,
435.14886, 358.4944 , 338.16928, 250.49905, 229.88107,
366.5563 , 541.996 , 471.8971 , 619.19135, 481.86276,
292.94177, 167.51329, 235.71179, 160.10254, 240.80624,
266.11838, 377.95245, 371.84662, 389.07236, 287.09613,
195.7979 , 171.33862, 316.3714 , 397.4343 , 481.42032,
568.32733, 587.319 , 529.6159 , 403.4232 , 94.08102,
194.59691, 427.52585, 313.3187 , 579.2178 , 813.1259 ,
932.89124, 1146.6532 , 1212.7903 , 1138.6241 , 1123.708 ,
445.83963, 285.92252, 319.54837, 292.5235 , 283.1231 ,
346.56586, 290.58084, 181.5028 , 616.64856, 862.625 ,
945.04517, 944.19257, 1141.5472 , 1118.1461 , 1077.6251 ,
566.94354, 528.96545, 498.58112, 414.7581 , 575.8048 ,
940.7235 , 942.35315, 993.0434 , 1151.0616 , 1276.6443 ,
1753.4945 , 1406.4716 , 1037.5525 , 473.98904, 658.7727 ,
1032.3734 , 1240.9309 , 1310.7902 , 1383.9104 , 1508.235 ,
1313.821 , 785.80853, 1224.1293 , 1187.1467 , 1286.4009 ,
236.88966, 455.52927, 769.01465, 1108.8147 , 760.0091 ,
889.72455, 875.0952 , 749.4331 , 784.4316 , 992.84973,
1912.151 , 2790.0994 , 3214.1863 , 3341.968 , 3329.6042 ,
3296.628 , 2292.491 , 959.46826, 719.39374, 1178.2322 ,
1470.5134 , 1697.8368 , 1919.3058 , 2746.6064 , 3219.8247 ,
2929.6543 , 2978.5505 , 3129.2446 , 3168.6892 , 4021.9907 ,
4089.688 , 5054.148 , 6050.384 , 5562.681 , 4750.575 ,
5072.5312 , 3777.8564 , 2821.4456 , 2989.25 , 5014.655 ,
6201.714 , 5801.3125 , 1855.2921 , 1188.0977 , 1038.029 ,
3048.906 , 3880.6555 , 4234.4824 , 3941.7537 , 3766.2598 ,
3977.833 , 4102.957 , 1739.501 , 1524.9628 , 1497.8783 ,
1533.368 , 1549.4961 , 2573.9736 , 3691.8413 , 3417.666 ,
2879.9602 , 2805.527 , 3212.8413 , 5337.166 , 6077.148 ,
5755.598 , 5887.5166 , 6894.076 , 6085.4966 , 7780.7104 ,
4568.3394 , 1296.731 , 2154.1467 , 1902.3751 , 2985.1375 ,
3298.6619 , 3707.22 , 3795.4312 , 4151.919 , 3050.7952 ,
3557.9084 , 6876.7324 , 8078.3726 , 10358.34 , 9426.234 ,
5532.7275 , 3787.621 , 2610.263 , 1581.6013 , 2030.0192 ,
3489.4673 , 2791.854 , 2995.7595 , 4010.7517 , 4497.829 ,
5752.2524 , 6753.869 , 6814.017 , 5756.7363 , 4256.9434 ,
2946.633 , 1648.4951 , 1561.2334 , 2633.9138 , 4153.923 ,
3201.2585 , 3726.5598 , 4380.634 , 5389.8013 , 6655.8975 ,
5624.784 , 5768.541 , 2165.2312 , 2557.9558 , 2830.0654 ,
3902.1606 , 4352.865 , 2299.0317 , 2254.1147 , 1335.6411 ,
842.3117 , 1302.7256 , 1783.9869 , 1503.4447 , 1189.288 ,
1677.9728 , 2013.1245 , 1737.7086 , 2449.7075 , 3250.4612 ,
3488.141 , 2213.2278 , 2470.196 , 2983.4202 , 4710.7153 ,
6813.271 , 8251.952 , 9655.526 , 8973.362 , 7405.713 ,
5692.6914 , 5354.8247 , 3610.3533 , 2611.8318 , 2924.2063 ,
4070.2922 , 5778.9272 , 3597.7861 , 1984.4889 , 1676.7507 ,
2113.166 , 1784.2113 , 1985.585 , 1972.1497 , 2219.31 ,
2369.629 , 2415.152 , 3558.768 , 5250.202 , 7756.2144 ,
9120.891 , 9950.581 , 11889.739 , 10731.838 , 9061.097 ,
10524.645 , 9156.805 , 9520.035 , 7026.346 , 7021.2266 ,
3731.8672 , 3600.7883 , 2972.5579 , 3203.1265 , 2626.8193 ,
3117.806 , 3618.4768 , 3815.0596 , 4569.737 , 2260.007 ,
1813.1381 , 1651.5854 , 1766.0149 , 2530.2202 , 3162.7053 ,
3436.3044 , 3368.88 , 2109.1497 , 1721.5989 , 2127.7112 ,
3714.397 , 2223.1472 , 2423.6338 , 1421.0337 , 2452.1794 ,
1589.9281 , 3778.968 , 5564.854 , 5893.3936 , 5500.6387 ,
4411.618 , 3639.6836 , 3738.685 , 1503.2898 , 1521.5819 ,
2087.766 , 1477.2672 , 1860.7589 , 2114.7656 , 1141.9276 ,
1124.0299 , 2054.813 , 1860.9283 , 1362.3218 , 889.35205,
968.52997, 908.33075, 1561.6315 , 1418.1512 , 1960.077 ,
1544.6833 , 1511.1964 , 2166.79 , 2943.324 , 3080.5833 ,
2051.4817 , 2381.0774 , 1612.6765 , 1263.2515 , 1633.4941 ,
1816.779 , 3020.5652 , 4646.243 , 5096.781 , 5573.561 ,
6298.1353 , 5327.4624 , 5098.7373 , 3684.539 , 1623.9558 ,
943.71155, 1406.1376 , 1926.6318 , 2911.6775 , 2308.7437 ,
2503.8333 , 3626.322 , 4054.2388 , 3297.2986 , 2660.316 ,
1670.8239 , 1345.7893 , 1318.7701 , 1765.8944 , 1451.3295 ,
1216.4293 , 1910.1718 , 2238.3594 , 2370.5903 , 2047.5101 ,
2351.4697 , 2819.8386 , 2663.843 , 734.36206, 1570.4484 ,
2375.731 , 2598.8862 , 2545.666 , 1500.8872 , 1046.03 ,
1361.3119 , 1577.006 , 1720.005 , 2014.0839 , 2401.4246 ,
2488.246 , 2280.2214 , 3298.842 , 4516.156 , 5899.289 ,
7291.6763 , 4677.422 , 2655.2527 , 1853.5786 , 1350.0608 ,
1867.0461 , 1782.38 , 2534.275 , 2371.5212 , 2957.4373 ,
4093.1406 , 4469.898 , 5004.2446 , 5425.072 , 3384.778 ,
1476.8087 , 1133.4874 , 1736.5038 , 2017.0848 , 2769.2231 ,
3156.1238 , 3883.9128 , 5148.3765 , 6870.726 , 6793.0815 ,
5755.5215 , 4870.2695 , 4470.0776 , 5439.217 , 5284.1895 ,
2193.0845 , 1733.62 , 2235.4656 , 2407.5215 , 2625.7144 ,
2173.9673 , 2751.993 , 3606.3147 , 3333.7178 , 2110.4922 ,
2091.6863 , 1866.379 , 2037.6451 ], dtype=float32)}It is getting results for each week
# Need to adjust our evaluate_preds() function to work with multiple dimension outputs
def evaluate_preds(y_true, y_pred):
# Make sure float32 (for metric calculations)
y_true = tf.cast(y_true, dtype=tf.float32)
y_pred = tf.cast(y_pred, dtype=tf.float32)
# Calculate various metrics
mae = tf.keras.metrics.mean_absolute_error(y_true, y_pred)
mse = tf.keras.metrics.mean_squared_error(y_true, y_pred)
rmse = tf.sqrt(mse)
mape = tf.keras.metrics.mean_absolute_percentage_error(y_true, y_pred)
mase = mean_absolute_scaled_error(y_true, y_pred)
# Account for different sized metrics (for longer horizons, reduce to single number)
if mae.ndim > 0: # if mae isn't already a scalar, reduce it to one by aggregating tensors to mean
mae = tf.reduce_mean(mae)
mse = tf.reduce_mean(mse)
rmse = tf.reduce_mean(rmse)
mape = tf.reduce_mean(mape)
mase = tf.reduce_mean(mase)
return {"mae": mae.numpy(),
"mse": mse.numpy(),
"rmse": rmse.numpy(),
"mape": mape.numpy(),
"mase": mase.numpy()}model_3_results = evaluate_preds(y_true=test_labels, y_pred=model_3_preds)
model_3_results{'mae': 2047.8856,
'mape': 5.7667065,
'mase': 2.2066739,
'mse': 10158263.0,
'rmse': 2338.8376}# There's not actually a way to visualize this because of the 7 day window, but here's one way to look at it
offset = 300
plt.figure(figsize=(10, 7))
# Account for the test_window offset and index into test_labels to ensure correct plotting
plot_time_series(timesteps=X_test[-len(test_windows):], values=tf.reduce_mean(test_labels, axis=1), start=offset, label="Test_data")
plot_time_series(timesteps=X_test[-len(test_windows):], values=tf.reduce_mean(model_3_preds, axis=1), start=offset, format="-", label="model_3_preds")Comparing our models so far
pd.DataFrame({"naive": naive_results["mae"],
"horizon_1_window_7": model_1_results["mae"],
"horizon_1_window_30": model_2_results["mae"],
"horizon_7_window_30": model_3_results["mae"]}, index=["mae"]).plot(figsize=(10, 7), kind="bar");So why does naive model do so well?
Because of autocorrelation in the data. The value at t+1 is typically pretty close to value at t
https://towardsdatascience.com/how-not-to-use-machine-learning-for-time-series-forecasting-avoiding-the-pitfalls-19f9d7adf424
Model 4: Conv 1D model
HORIZON = 1 # predict next day
WINDOW_SIZE = 7 # use previous week worth of data# Create windowed dataset
full_windows, full_labels = make_windows(np.array(prices), window_size=WINDOW_SIZE, horizon=HORIZON)
len(full_windows), len(full_labels)(2541, 2541)# Create train/test splits
train_windows, test_windows, train_labels, test_labels = make_train_test_splits(full_windows, full_labels)
len(train_windows), len(test_windows), len(train_labels), len(test_labels)(2032, 509, 2032, 509)We need an input shape of (batch_size, timesteps, input_dim)
train_windows[0].shape(7,)# Before we pass our data to the Conv1D layer we need to reshape it in order to make sure it works
x = tf.constant(train_windows[0])
x<tf.Tensor: shape=(7,), dtype=float64, numpy=
array([324.71833, 332.45666, 336.585 , 346.775 , 344.81166, 343.065 ,
358.50166])># Before we pass our data to the Conv1D layer, we have to reshape it in order to make sure it works
expand_dims_layer = layers.Lambda(lambda x: tf.expand_dims(x, axis=1)) # add an extra dimension for timesteps
print(f"Original shape: {x.shape}") # (WINDOW_SIZE)
print(f"Expanded shape: {expand_dims_layer(x).shape}") # (WINDOW_SIZE, input_dim)
print(f"Original values with expanded shape:\n {expand_dims_layer(x)}")Original shape: (7,)
Expanded shape: (7, 1)
Original values with expanded shape:
[[324.71832]
[332.45667]
[336.585 ]
[346.775 ]
[344.81165]
[343.065 ]
[358.50165]]# Create Conv1D
tf.random.set_seed(42)
model_4 = tf.keras.Sequential([
expand_dims_layer,
layers.Conv1D(filters=128, kernel_size=5, padding='causal', activation='relu'), # causal makes sure it learns in the correct temporal order
layers.Dense(HORIZON)
], name='model_4_conv')
model_4.compile(loss='mae',
metrics=['mae'],
optimizer=tf.keras.optimizers.Adam())
model_4.fit(train_windows, train_labels,
epochs=100,
batch_size=32,
verbose=0,
validation_data=(test_windows,test_labels),
callbacks=[create_model_checkpoint(model_name=model_4.name)])INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
INFO:tensorflow:Assets written to: model_experiments/model_4_conv/assets
<keras.callbacks.History at 0x7f9e7a91b250>Evaluate
model_4 = tf.keras.models.load_model("model_experiments/model_4_conv/")
model_4.evaluate(test_windows, test_labels)16/16 [==============================] - 0s 4ms/step - loss: 923.6509 - mae: 923.6509
[923.6509399414062, 923.6509399414062]# Get forecast predictions
model_4_preds = make_preds(model_4,
input_data=test_windows)test_labels.shape, model_4_preds.shape((509, 1), TensorShape([509]))model_4_results = evaluate_preds(y_true=tf.squeeze(test_labels), y_pred=model_4_preds)
model_4_results{'mae': 923.6509,
'mape': 2.604802,
'mase': 0.99763596,
'mse': 2076426.5,
'rmse': 1440.9811}naive_results{'mae': 924.30414,
'mape': 2.6047938,
'mase': 0.99834156,
'mse': 2055079.2,
'rmse': 1433.5548}Model 5: LSTM model
HORIZON = 1 # predict next day
WINDOW_SIZE = 7 # use previous week worth of data# Create windowed dataset
full_windows, full_labels = make_windows(np.array(prices), window_size=WINDOW_SIZE, horizon=HORIZON)
len(full_windows), len(full_labels)(2541, 2541)# Create train/test splits
train_windows, test_windows, train_labels, test_labels = make_train_test_splits(full_windows, full_labels)
len(train_windows), len(test_windows), len(train_labels), len(test_labels)(2032, 509, 2032, 509)# Create LSTM model
tf.random.set_seed(42)
# Let's build an LSTM model with the Functional API
inputs = layers.Input(shape=(WINDOW_SIZE))
x = layers.Lambda(lambda x: tf.expand_dims(x, axis=1))(inputs) # expand input dimension to be compatible with LSTM
# print(x.shape)
# x = layers.LSTM(128, activation="relu", return_sequences=True)(x) # this layer will error if the inputs are not the right shape
x = layers.LSTM(128, activation="relu")(x) # using the tanh loss function results in a massive error
# print(x.shape)
# Add another optional dense layer (you could add more of these to see if they improve model performance)
# x = layers.Dense(32, activation="relu")(x)
output = layers.Dense(HORIZON)(x)
model_5 = tf.keras.Model(inputs=inputs, outputs=output, name="model_5_lstm")
model_5.compile(loss='mae',
metrics=['mae'],
optimizer=tf.keras.optimizers.Adam())
model_5.fit(train_windows, train_labels,
epochs=100,
batch_size=128,
verbose=0,
validation_data=(test_windows,test_labels),
callbacks=[create_model_checkpoint(model_name=model_5.name)])WARNING:tensorflow:Layer lstm will not use cuDNN kernels since it doesn't meet the criteria. It will use a generic GPU kernel as fallback when running on GPU.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
INFO:tensorflow:Assets written to: model_experiments/model_5_lstm/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f9e7807d210> has the same name 'LSTMCell' as a built-in Keras object. Consider renaming <class 'keras.layers.recurrent.LSTMCell'> to avoid naming conflicts when loading with `tf.keras.models.load_model`. If renaming is not possible, pass the object in the `custom_objects` parameter of the load function.
<keras.callbacks.History at 0x7f9e77d83950>Evaluate
model_5 = tf.keras.models.load_model("model_experiments/model_5_lstm/")
model_5.evaluate(test_windows, test_labels)WARNING:tensorflow:Layer lstm will not use cuDNN kernels since it doesn't meet the criteria. It will use a generic GPU kernel as fallback when running on GPU.
WARNING:tensorflow:Layer lstm will not use cuDNN kernels since it doesn't meet the criteria. It will use a generic GPU kernel as fallback when running on GPU.
16/16 [==============================] - 0s 4ms/step - loss: 998.5007 - mae: 998.5007
[998.500732421875, 998.500732421875]# Get forecast predictions
model_5_preds = make_preds(model_5,
input_data=test_windows)test_labels.shape, model_5_preds.shape((509, 1), TensorShape([509]))model_5_results = evaluate_preds(y_true=tf.squeeze(test_labels), y_pred=model_5_preds)
model_5_results{'mae': 998.50073,
'mape': 2.8254218,
'mase': 1.0784813,
'mse': 2332680.2,
'rmse': 1527.3114}Model 6: Multivariate model
I would love to add Elon Musk tweeting as an additional variable but I think that would be better for dodgecoin
Maybe could add number of tweets about bitcoin
However, bitcoin halving seems like a strong variable to test
Sidenote: here's how to do regression with multivariate time series data: https://www.analyticsvidhya.com/blog/2018/09/multivariate-time-series-guide-forecasting-modeling-python-codes/
# Block reward values
block_reward_1 = 50 # 3 January 2009 (2009-01-03) - this block reward isn't in our dataset (it starts from 01 October 2013)
block_reward_2 = 25 # 28 November 2012
block_reward_3 = 12.5 # 9 July 2016
block_reward_4 = 6.25 # 11 May 2020
# Block reward dates (datetime form of the above date stamps)
block_reward_2_datetime = np.datetime64("2012-11-28")
block_reward_3_datetime = np.datetime64("2016-07-09")
block_reward_4_datetime = np.datetime64("2020-05-11")bitcoin_prices.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
# Get date indexes for when to add in different block dates
block_reward_2_days = (block_reward_3_datetime - bitcoin_prices.index[0]).days
block_reward_3_days = (block_reward_4_datetime - bitcoin_prices.index[0]).days
block_reward_2_days, block_reward_3_days(613, 2015)# Add block_reward column
bitcoin_prices_block = bitcoin_prices.copy()
bitcoin_prices_block["block_reward"] = None
# Set values of block_reward column (it's the last column hence -1 indexing on iloc)
bitcoin_prices_block.iloc[:block_reward_2_days, -1] = block_reward_2
bitcoin_prices_block.iloc[block_reward_2_days:block_reward_3_days, -1] = block_reward_3
bitcoin_prices_block.iloc[block_reward_3_days:, -1] = block_reward_4
bitcoin_prices_block.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
# Plot the block reward/price over time
# Note: Because of the different scales of our values we'll scale them to be between 0 and 1.
from sklearn.preprocessing import minmax_scale
scaled_price_block_df = pd.DataFrame(minmax_scale(bitcoin_prices_block[["Price", "block_reward"]]), # we need to scale the data first
columns=bitcoin_prices_block.columns,
index=bitcoin_prices_block.index)
scaled_price_block_df.plot(figsize=(10, 7));Preparing data for multivariate model
Our current functions won't work because we have two variables
Let's use pandas.DataFrame.shift()
# Setup dataset hyperparameters
HORIZON = 1
WINDOW_SIZE = 7# Make a copy of the Bitcoin historical data with block reward feature
bitcoin_prices_windowed = bitcoin_prices_block.copy()
# Add windowed columns
for i in range(WINDOW_SIZE): # Shift values for each step in WINDOW_SIZE
bitcoin_prices_windowed[f"Price+{i+1}"] = bitcoin_prices_windowed["Price"].shift(periods=i+1)
bitcoin_prices_windowed.head(10).dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
Yay now we can have data like this:
[0, 1, 2, 3, 4, 5, 6, block_reward] -> [7]
[1 ,2 ,3, 4, 5, 6, 7, block_reward] -> [8]# Create X (windows) and y(horizons) features
X = bitcoin_prices_windowed.dropna().drop("Price", axis=1).astype(np.float32)
y = bitcoin_prices_windowed.dropna()["Price"].astype(np.float32)
X.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
This looks like a basic setup for a regression model!
# Make train and test sets
split_size = int(len(X) * 0.8)
X_train, y_train = X[:split_size], y[:split_size]
X_test, y_test = X[split_size:], y[split_size:]
len(X_train), len(y_train), len(X_test), len(y_test)(2032, 2032, 509, 509)Creating model
tf.random.set_seed(42)
# Make multivariate time series model
model_6 = tf.keras.Sequential([
layers.Dense(128, activation="relu"),
layers.Dense(128, activation="relu"), # adding an extra layer here should lead to beating the naive model
layers.Dense(HORIZON)
], name="model_6_dense_multivariate")
# Compile
model_6.compile(loss="mae",
optimizer=tf.keras.optimizers.Adam())
# Fit
model_6.fit(X_train, y_train,
epochs=100,
batch_size=128,
verbose=0, # only print 1 line per epoch
validation_data=(X_test, y_test),
callbacks=[create_model_checkpoint(model_name=model_6.name)])INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
INFO:tensorflow:Assets written to: model_experiments/model_6_dense_multivariate/assets
<keras.callbacks.History at 0x7f9e77455b90>Evaluate
model_6 = tf.keras.models.load_model("model_experiments/model_6_dense_multivariate/")
model_6.evaluate(X_test, y_test)16/16 [==============================] - 0s 2ms/step - loss: 922.8585
922.8585205078125# Get forecast predictions
model_6_preds = make_preds(model_6,
input_data=X_test)y_test.shape, model_6_preds.shape((509,), TensorShape([509]))model_6_results = evaluate_preds(y_true=y_test, y_pred=model_6_preds)
model_6_results{'mae': 922.85846,
'mape': 2.6030893,
'mase': 0.9967801,
'mse': 2067919.0,
'rmse': 1438.0261}naive_results{'mae': 924.30414,
'mape': 2.6047938,
'mase': 0.99834156,
'mse': 2055079.2,
'rmse': 1433.5548}Model 7: N-BEATS algorithm
Let's build the biggest model based on 2020's state-of-the-art N-BEATS model
Let's see how it performs!
Building and testing the 'Block Input' Layer
It doesn't exist in TF so we need to create it: https://www.tensorflow.org/guide/keras/custom_layers_and_models
# Create NBeatBlock custom layer
class NBeatsBlock(tf.keras.layers.Layer):
def __init__(self,
input_size: int,
theta_size: int,
horizon: int,
n_neurons: int,
n_layers: int,
**kwargs): # **kwargs has the tf.keras.layers arguments such as name
super().__init__(**kwargs)
self.input_size = input_size # WINDOW_SIZE
self.theta_size = theta_size
self.horizon = horizon
self.n_neurons = n_neurons
self.n_layers = n_layers
# Block contains stack of 4 fully connected layers each has ReLU activation
self.hidden = [tf.keras.layers.Dense(n_neurons, activation='relu') for _ in range(n_layers)]
# Output of block is theta layer with linear activation
self.theta_layer = tf.keras.layers.Dense(theta_size, activation='linear', name="theta")
def call(self, inputs):
x = inputs
for layer in self.hidden:
x = layer(x)
theta = self.theta_layer(x)
# Output the backcast and forecast from the theta
backcast, forecast = theta[:, :self.input_size], theta[:, -self.horizon]
return backcast, forecastLet's test this class
# Set up dummy NBeatsBlock layer to represent inputs and outputs
dummy_nbeats_block_layer = NBeatsBlock(input_size=WINDOW_SIZE,
theta_size=WINDOW_SIZE+HORIZON, # backcast + forecast
n_neurons = 128,
horizon=HORIZON,
n_layers=4)# Create dummy inputs (have to be same size as input_size)
dummy_inputs = tf.expand_dims(tf.range(WINDOW_SIZE) + 1, axis=0)
dummy_inputs<tf.Tensor: shape=(1, 7), dtype=int32, numpy=array([[1, 2, 3, 4, 5, 6, 7]], dtype=int32)># Pass dummy inputs to dummy NBeatsBlock layer
backcast, forecast = dummy_nbeats_block_layer(dummy_inputs)
# These are the activation outputs of the theta layer (they'll be random due to no training of the model)
print(f"Backcast: {tf.squeeze(backcast.numpy())}")
print(f"Forecast: {tf.squeeze(forecast.numpy())}")Backcast: [-0.4634872 0.10944451 -0.4335534 0.93602407 -0.35703534 -0.17927621
-0.34048355]
Forecast: -0.4397900700569153So far so good
Data Pipeline
Using tf.data: https://www.tensorflow.org/guide/data_performance
HORIZON = 1
WINDOW_SIZE = 7# Create NBEATS data inputs (N-BEATS works with univariate time series)
bitcoin_prices.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
# Add windowed columns
bitcoin_prices_nbeats = bitcoin_prices.copy()
for i in range(WINDOW_SIZE):
bitcoin_prices_nbeats[f"Price+{i+1}"] = bitcoin_prices_nbeats["Price"].shift(periods=i+1)
bitcoin_prices_nbeats.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
# Make features and labels
X = bitcoin_prices_nbeats.dropna().drop("Price", axis=1)
y = bitcoin_prices_nbeats.dropna()["Price"]
# Make train and test sets
split_size = int(len(X) * 0.8)
X_train, y_train = X[:split_size], y[:split_size]
X_test, y_test = X[split_size:], y[split_size:]
len(X_train), len(y_train), len(X_test), len(y_test)(2032, 2032, 509, 509)# Time to use tf.data API
train_features_dataset = tf.data.Dataset.from_tensor_slices(X_train)
train_labels_dataset = tf.data.Dataset.from_tensor_slices(y_train)
test_features_dataset = tf.data.Dataset.from_tensor_slices(X_test)
test_labels_dataset = tf.data.Dataset.from_tensor_slices(y_test)
# Combine labels and features
train_dataset = tf.data.Dataset.zip((train_features_dataset, train_labels_dataset))
test_dataset = tf.data.Dataset.zip((test_features_dataset, test_labels_dataset))
# Batch and prefetch
BATCH_SIZE = 1024 # see paper
train_dataset = train_dataset.batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
test_dataset = test_dataset.batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
train_dataset, test_dataset(<PrefetchDataset shapes: ((None, 7), (None,)), types: (tf.float64, tf.float64)>,
<PrefetchDataset shapes: ((None, 7), (None,)), types: (tf.float64, tf.float64)>)Hyperparameters for the model
See table 18 in the paper (we are doing N-BEATS-G)
N_EPOCHS = 5000
N_NEURONS = 512
N_LAYERS = 4 # they used 4 layers in figure 1
N_STACKS = 30
INPUT_SIZE = WINDOW_SIZE * HORIZON
THETA_SIZE = INPUT_SIZE + HORIZON # backcast + forecast
INPUT_SIZE, THETA_SIZE(7, 8)Preparing for Residual Connections (see the diagram with the stacks)
The N-BEATS algorithm uses double residual stacking (3.2) because the architecture is pretty deep (vanishing gradient issues in deep models)
See https://paperswithcode.com/method/resnet
# Make tensors
tensor_1 = tf.range(10) + 10
tensor_2 = tf.range(10)subtracted = layers.subtract([tensor_1, tensor_2])
added = layers.add([tensor_1, tensor_2])
print(f"Input tensors: {tensor_1.numpy()} & {tensor_2.numpy()}")
print(f"Subtracted: {subtracted.numpy()}")
print(f"Added: {added.numpy()}")Input tensors: [10 11 12 13 14 15 16 17 18 19] & [0 1 2 3 4 5 6 7 8 9]
Subtracted: [10 10 10 10 10 10 10 10 10 10]
Added: [10 12 14 16 18 20 22 24 26 28]Building, Compiling, and Fitting N-BEATS
Setup an instance of the N-BEATS block layer (this will be first block, and then will set up a loop to create stacks)
Create an input layer for the N-BEATS stack (using Functional API)
Make the initial backcast and forecast for the model using (1.)
Use for loop to create stacks of block layers
Use the
NBeatsBlockclass within the for loop in (4.) to create blocks which return backcasts and block-level forecastsCreate the double residual stacking using subtract and add layers
Put the model inputs and outputs together in
tf.keras.Model()Compile the model with MAE loss (the paper uses multiple losses but we'll use MAE to be consistent) and Adam optimizer (paper used it)
Fit the N-BEATS for 5000 epochs and use some callbacks:
Early Stopping
Reduce LR on plateau
# 1.
nbeats_initial_layer = NBeatsBlock(input_size=WINDOW_SIZE,
theta_size=THETA_SIZE, # backcast + forecast
n_neurons = N_NEURONS,
horizon=HORIZON,
n_layers=N_LAYERS,
name='InitialBlock')
nbeats_layer = NBeatsBlock(input_size=WINDOW_SIZE,
theta_size=THETA_SIZE, # backcast + forecast
n_neurons = N_NEURONS,
horizon=HORIZON,
n_layers=N_LAYERS,
name='Block')# Rough draft
# tf.random.set_seed(42)
# from tensorflow.keras import layers
# # 2.
# inputs = layers.Input(shape=(INPUT_SIZE), name='input layer')
# # 3.
# backcast, forecast = nbeats_initial_layer(inputs)
# stack = [nbeats_layer for _ in range(N_STACKS-1)]
# all_forecast = forecast
# for block in stack:
# inputs = tf.subtract(inputs, backcast, name='subtract')
# backcast, forecast = block(inputs)
# all_forecast = tf.add(all_forecast, forecast)
%%time
tf.random.set_seed(42)
# 1. Setup N-BEATS Block layer
nbeats_block_layer = NBeatsBlock(input_size=INPUT_SIZE,
theta_size=THETA_SIZE,
horizon=HORIZON,
n_neurons=N_NEURONS,
n_layers=N_LAYERS,
name="InitialBlock")
# 2. Create input to stacks
stack_input = layers.Input(shape=(INPUT_SIZE), name="stack_input")
# 3. Create initial backcast and forecast input (backwards predictions are referred to as residuals in the paper)
backcast, forecast = nbeats_block_layer(stack_input)
# Add in subtraction residual link, thank you to: https://github.com/mrdbourke/tensorflow-deep-learning/discussions/174
residuals = layers.subtract([stack_input, backcast], name=f"subtract_00")
# 4. Create stacks of blocks
for i, _ in enumerate(range(N_STACKS-1)): # first stack is already created in (3)
# 5. Use the NBeatsBlock to calculate the backcast as well as block forecast
backcast, block_forecast = NBeatsBlock(
input_size=INPUT_SIZE,
theta_size=THETA_SIZE,
horizon=HORIZON,
n_neurons=N_NEURONS,
n_layers=N_LAYERS,
name=f"NBeatsBlock_{i}"
)(residuals) # pass it in residuals (the backcast)
# 6. Create the double residual stacking
residuals = layers.subtract([residuals, backcast], name=f"subtract_{i}")
forecast = layers.add([forecast, block_forecast], name=f"add_{i}")
# 7. Put the stack model together
model_7 = tf.keras.Model(inputs=stack_input,
outputs=forecast,
name="model_7_N-BEATS")
# 8. Compile with MAE loss and Adam optimizer
model_7.compile(loss="mae",
optimizer=tf.keras.optimizers.Adam(0.001),
metrics=["mae", "mse"])
# 9. Fit the model with EarlyStopping and ReduceLROnPlateau callbacks
model_7.fit(train_dataset,
epochs=N_EPOCHS,
validation_data=test_dataset,
verbose=0, # prevent large amounts of training outputs
# callbacks=[create_model_checkpoint(model_name=stack_model.name)] # saving model every epoch consumes far too much time
callbacks=[tf.keras.callbacks.EarlyStopping(monitor="val_loss", patience=200, restore_best_weights=True),
tf.keras.callbacks.ReduceLROnPlateau(monitor="val_loss", patience=100, verbose=1)])Epoch 00246: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00346: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
CPU times: user 1min 50s, sys: 2.65 s, total: 1min 53s
Wall time: 2min 23s# Make predictions with N-BEATS model
model_7_preds = make_preds(model_7, test_dataset)
# Evaluate N-BEATS model predictions
model_7_results = evaluate_preds(y_true=y_test,
y_pred=model_7_preds)
model_7_results{'mae': 952.4994,
'mape': 2.6837056,
'mase': 1.0287952,
'mse': 2230253.8,
'rmse': 1493.4036}# compare to model_1_results
model_1_results{'mae': 933.08374,
'mape': 2.6302116,
'mase': 1.0078244,
'mse': 2106266.2,
'rmse': 1451.2981}# Plot the N-BEATS model and inspect the architecture
from tensorflow.keras.utils import plot_model
plot_model(model_7)Model 8: Creating an Ensemble
An ensemble leverages the wisdom of the crowd effect
It combines many different models to predict a common goal
HORIZON, WINDOW_SIZE(1, 7)## Constructing and fitting multiple models
def get_ensemble_models(horizon=HORIZON,
train_data=train_dataset,
test_data=test_dataset,
num_iter=10,
num_epochs=1000,
loss_fn=['mae','mse','mape']):
"""
Returns a list of num_iter models each trained on MAE, MSE, and MAPE loss
For example, if num_iter = 10 it'll return 30 models
"""
ensemble_models = []
for i in range(num_iter):
# Build and fit a new model with different loss function
for loss_function in loss_fn:
print(f"Optimizing model by reducing: {loss_function} for {num_epochs}, model number {i}")
# construct a simple model (similar to model_1)
model = tf.keras.Sequential([
# using he_normal for prediction intervals later
layers.Dense(128, kernel_initializer='he_normal', activation='relu'),
layers.Dense(128, kernel_initializer='he_normal', activation='relu'),
layers.Dense(HORIZON)
])
# Compile simple model
model.compile(loss=loss_function,
optimizer=tf.keras.optimizers.Adam(),
metrics=['mae','mse'])
# Fit the current model
model.fit(train_data,
epochs=num_epochs,
verbose=0,
validation_data=test_data,
callbacks=[tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=200,
restore_best_weights=True),
tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
patience=100,
verbose=1)])
ensemble_models.append(model)
return ensemble_models%%time
# Get list of trained ensemble models
ensemble_models = get_ensemble_models(num_iter=5,
num_epochs=1000)Optimizing model by reducing: mae for 1000, model number 0
Optimizing model by reducing: mse for 1000, model number 0
Epoch 00641: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00741: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Optimizing model by reducing: mape for 1000, model number 0
Epoch 00129: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Optimizing model by reducing: mae for 1000, model number 1
Optimizing model by reducing: mse for 1000, model number 1
Epoch 00424: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00524: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Optimizing model by reducing: mape for 1000, model number 1
Epoch 00539: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00661: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Optimizing model by reducing: mae for 1000, model number 2
Epoch 00726: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Optimizing model by reducing: mse for 1000, model number 2
Epoch 00900: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 01000: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Optimizing model by reducing: mape for 1000, model number 2
Epoch 00497: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00597: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Optimizing model by reducing: mae for 1000, model number 3
Epoch 00640: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Optimizing model by reducing: mse for 1000, model number 3
Epoch 00385: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00485: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Optimizing model by reducing: mape for 1000, model number 3
Epoch 00316: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Epoch 00429: ReduceLROnPlateau reducing learning rate to 1.0000000474974514e-05.
Epoch 00529: ReduceLROnPlateau reducing learning rate to 1.0000000656873453e-06.
Optimizing model by reducing: mae for 1000, model number 4
Epoch 00601: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
Optimizing model by reducing: mse for 1000, model number 4
Optimizing model by reducing: mape for 1000, model number 4
Epoch 00774: ReduceLROnPlateau reducing learning rate to 0.00010000000474974513.
CPU times: user 7min 30s, sys: 50.3 s, total: 8min 20s
Wall time: 9min 14slen(ensemble_models)15ensemble_models[<keras.engine.sequential.Sequential at 0x7f9e78c60610>,
<keras.engine.sequential.Sequential at 0x7f9e78b33a10>,
<keras.engine.sequential.Sequential at 0x7f9e75b77e50>,
<keras.engine.sequential.Sequential at 0x7f9e752cb3d0>,
<keras.engine.sequential.Sequential at 0x7f9e79152190>,
<keras.engine.sequential.Sequential at 0x7f9dff702e10>,
<keras.engine.sequential.Sequential at 0x7f9dff58ced0>,
<keras.engine.sequential.Sequential at 0x7f9dff39dc90>,
<keras.engine.sequential.Sequential at 0x7f9dff1f3290>,
<keras.engine.sequential.Sequential at 0x7f9df98539d0>,
<keras.engine.sequential.Sequential at 0x7f9df9805810>,
<keras.engine.sequential.Sequential at 0x7f9df9554f50>,
<keras.engine.sequential.Sequential at 0x7f9df93dedd0>,
<keras.engine.sequential.Sequential at 0x7f9df9253e10>,
<keras.engine.sequential.Sequential at 0x7f9df90ee4d0>]Make predictions with ensemble model
# Create a function to return a list of predictions
def make_ensemble_preds(ensemble_models,data):
ensemble_preds = []
for model in ensemble_models:
preds = model.predict(data)
ensemble_preds.append(preds)
return tf.constant(tf.squeeze(ensemble_preds))%%time
ensemble_preds = make_ensemble_preds(ensemble_models=ensemble_models,
data=test_dataset)WARNING:tensorflow:5 out of the last 20 calls to <function Model.make_predict_function.<locals>.predict_function at 0x7f9e78b42dd0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.
WARNING:tensorflow:5 out of the last 20 calls to <function Model.make_predict_function.<locals>.predict_function at 0x7f9e78b42dd0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.
WARNING:tensorflow:6 out of the last 21 calls to <function Model.make_predict_function.<locals>.predict_function at 0x7f9e752818c0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.
WARNING:tensorflow:6 out of the last 21 calls to <function Model.make_predict_function.<locals>.predict_function at 0x7f9e752818c0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings could be due to (1) creating @tf.function repeatedly in a loop, (2) passing tensors with different shapes, (3) passing Python objects instead of tensors. For (1), please define your @tf.function outside of the loop. For (2), @tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3), please refer to https://www.tensorflow.org/guide/function#controlling_retracing and https://www.tensorflow.org/api_docs/python/tf/function for more details.
CPU times: user 1.09 s, sys: 55.1 ms, total: 1.14 s
Wall time: 1.13 s# Evaluate ensemble model predictions
ensemble_results = evaluate_preds(y_true=y_test,
y_pred=ensemble_preds)
ensemble_results{'mae': 941.8538,
'mape': 2.6583095,
'mase': 1.017297,
'mse': 2135938.8,
'rmse': 1461.3636}# Or we can do this
ensemble_results = evaluate_preds(y_true=y_test,
y_pred=tf.reduce_mean(ensemble_preds, axis=0))
ensemble_results{'mae': 924.18054,
'mape': 2.6091447,
'mase': 0.99820805,
'mse': 2087776.1,
'rmse': 1444.914}# Or we can do this - median performed best here!
ensemble_results = evaluate_preds(y_true=y_test,
y_pred=np.median(ensemble_preds, axis=0))
ensemble_results{'mae': 922.4596,
'mape': 2.603472,
'mase': 0.9963493,
'mse': 2083582.8,
'rmse': 1443.462}# Performed better than model_1!
model_1_results{'mae': 933.08374,
'mape': 2.6302116,
'mase': 1.0078244,
'mse': 2106266.2,
'rmse': 1451.2981}Plotting the prediction intervals
One benefit of using an ensemble is the ability to get prediction intervals
Bootstrap method:
Take the predictions from a number of randomly initialized models (ensemble models)
Measure the standard deviation of the predictions
Mulitiply the standard deviation by 1.96 (assuming the distribution is Gaussian/Normal)
To get the prediction interval upper and lower bounds, add and subtract the value obtained in (3.) to the mean/median of the predictions made in (1.)
See: https://eng.uber.com/neural-networks-uncertainty-estimation/
# Find upper and lower bounds of ensemble predictions
def get_upper_lower(preds): # 1. Take the predictions of multiple randomly initialized deep learning neural networks
# 2. Measure the standard deviation of the predictions
std = tf.math.reduce_std(preds, axis=0)
# 3. Multiply the standard deviation by 1.96
interval = 1.96 * std # https://en.wikipedia.org/wiki/1.96
# 4. Get the prediction interval upper and lower bounds
preds_mean = tf.reduce_mean(preds, axis=0)
lower, upper = preds_mean - interval, preds_mean + interval
return lower, upper
# Get the upper and lower bounds of the 95%
lower, upper = get_upper_lower(preds=ensemble_preds)# Get the median values of our ensemble preds
ensemble_median = np.median(ensemble_preds, axis=0)
# Plot the median of our ensemble preds along with the prediction intervals (where the predictions fall between)
offset=475
plt.figure(figsize=(10, 7))
plt.plot(X_test.index[offset:], y_test[offset:], "g", label="Test Data")
plt.plot(X_test.index[offset:], ensemble_median[offset:], "k-", label="Ensemble Median")
plt.xlabel("Date")
plt.ylabel("BTC Price")
plt.fill_between(X_test.index[offset:],
(lower)[offset:],
(upper)[offset:], label="Prediction Intervals")
plt.legend(loc="upper left", fontsize=14);So far all the models predictions are lagging behind the test data, essentially doing the same thing as the naive model - predicting the prev timestep as the next timestep
Potential Problems:
Overfitting
Model might be copying the naive model
Data isn't predictable (I believe this is the main problem)
Note: the prediction intervals were made assuming the model's data is normally distributed
Model 9 - predicting into the future by training on full data
bitcoin_prices_windowed.head().dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
HORIZON = 1
WINDOW_SIZE = 7# Train model on entire data to make predictions for the next day
X_all = bitcoin_prices_windowed.dropna().drop(['Price', 'block_reward'], axis=1).to_numpy()
y_all = bitcoin_prices_windowed.dropna()['Price'].to_numpy()
len(X_all), len(y_all)(2541, 2541)# Turn data into efficient running tensors
features_dataset_all = tf.data.Dataset.from_tensor_slices(X_all)
labels_dataset_all = tf.data.Dataset.from_tensor_slices(y_all)
dataset_all = tf.data.Dataset.zip((features_dataset_all, labels_dataset_all))
BATCH_SIZE = 1024
dataset_all = dataset_all.batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
dataset_all<PrefetchDataset shapes: ((None, 7), (None,)), types: (tf.float64, tf.float64)>tf.random.set_seed(42)
# create simple model
model_9 = tf.keras.Sequential([
layers.Dense(128, activation='relu'),
layers.Dense(128, activation='relu'),
layers.Dense(HORIZON)
], name='model_9_future_model')
model_9.compile(loss=tf.keras.losses.mae,
optimizer=tf.keras.optimizers.Adam())
model_9.fit(dataset_all,
epochs = 100,
verbose=0)<keras.callbacks.History at 0x7f9e7a326810>Make predictions into the future
# How many timesteps to predict into the future?
INTO_FUTURE = 14To make predictions into the future we want a function which:
Takes as an input:
a list of values (Bitcoin historical)
a trained model
a window into the future to predict (
INTO_FUTURE)window size a model was trained (
WINDOW_SIZE)
Creates an empty list for future forecasts which we will fill up and extracts the last
WINDOW SIZEvalues from the input valuesLoop
INTO_FUTUREtimes making a prediction onWINDOW_SIZEsequences which update to remove the first value and append the latest prediction
def make_forecast(data, model, into_future, window_size):
forecasts = []
inputs = data[-window_size:]
for i in range(0, into_future):
pred = model.predict(tf.data.Dataset.from_tensor_slices(inputs.reshape((1,7))).batch(BATCH_SIZE))
print(f"Predicting on: \n {inputs} -> Prediction: {pred[0][0]}\n")
forecasts.append(pred[0][0])
inputs = np.append(inputs, pred)
inputs = inputs[-window_size:]
return forecastsmake_forecast(y_all, model_9, INTO_FUTURE, WINDOW_SIZE)Predicting on:
[63901.07308021 65972.86312976 62603.57506991 60689.23826486
61124.34712562 60936.15085142 63004.38111467] -> Prediction: 64438.8203125
Predicting on:
[65972.86312976 62603.57506991 60689.23826486 61124.34712562
60936.15085142 63004.38111467 64438.8203125 ] -> Prediction: 63827.90625
Predicting on:
[62603.57506991 60689.23826486 61124.34712562 60936.15085142
63004.38111467 64438.8203125 63827.90625 ] -> Prediction: 61270.8984375
Predicting on:
[60689.23826486 61124.34712562 60936.15085142 63004.38111467
64438.8203125 63827.90625 61270.8984375 ] -> Prediction: 60892.890625
Predicting on:
[61124.34712562 60936.15085142 63004.38111467 64438.8203125
63827.90625 61270.8984375 60892.890625 ] -> Prediction: 61533.6796875
Predicting on:
[60936.15085142 63004.38111467 64438.8203125 63827.90625
61270.8984375 60892.890625 61533.6796875 ] -> Prediction: 62313.30859375
Predicting on:
[63004.38111467 64438.8203125 63827.90625 61270.8984375
60892.890625 61533.6796875 62313.30859375] -> Prediction: 63516.15234375
Predicting on:
[64438.8203125 63827.90625 61270.8984375 60892.890625
61533.6796875 62313.30859375 63516.15234375] -> Prediction: 63653.12890625
Predicting on:
[63827.90625 61270.8984375 60892.890625 61533.6796875
62313.30859375 63516.15234375 63653.12890625] -> Prediction: 62261.9921875
Predicting on:
[61270.8984375 60892.890625 61533.6796875 62313.30859375
63516.15234375 63653.12890625 62261.9921875 ] -> Prediction: 61014.31640625
Predicting on:
[60892.890625 61533.6796875 62313.30859375 63516.15234375
63653.12890625 62261.9921875 61014.31640625] -> Prediction: 61434.3984375
Predicting on:
[61533.6796875 62313.30859375 63516.15234375 63653.12890625
62261.9921875 61014.31640625 61434.3984375 ] -> Prediction: 62239.453125
Predicting on:
[62313.30859375 63516.15234375 63653.12890625 62261.9921875
61014.31640625 61434.3984375 62239.453125 ] -> Prediction: 62905.96484375
Predicting on:
[63516.15234375 63653.12890625 62261.9921875 61014.31640625
61434.3984375 62239.453125 62905.96484375] -> Prediction: 63236.1484375
[64438.82,
63827.906,
61270.9,
60892.89,
61533.68,
62313.31,
63516.152,
63653.13,
62261.992,
61014.316,
61434.4,
62239.453,
62905.965,
63236.15]future_forecast = make_forecast(y_all, model_9, INTO_FUTURE, WINDOW_SIZE)Predicting on:
[63901.07308021 65972.86312976 62603.57506991 60689.23826486
61124.34712562 60936.15085142 63004.38111467] -> Prediction: 64438.8203125
Predicting on:
[65972.86312976 62603.57506991 60689.23826486 61124.34712562
60936.15085142 63004.38111467 64438.8203125 ] -> Prediction: 63827.90625
Predicting on:
[62603.57506991 60689.23826486 61124.34712562 60936.15085142
63004.38111467 64438.8203125 63827.90625 ] -> Prediction: 61270.8984375
Predicting on:
[60689.23826486 61124.34712562 60936.15085142 63004.38111467
64438.8203125 63827.90625 61270.8984375 ] -> Prediction: 60892.890625
Predicting on:
[61124.34712562 60936.15085142 63004.38111467 64438.8203125
63827.90625 61270.8984375 60892.890625 ] -> Prediction: 61533.6796875
Predicting on:
[60936.15085142 63004.38111467 64438.8203125 63827.90625
61270.8984375 60892.890625 61533.6796875 ] -> Prediction: 62313.30859375
Predicting on:
[63004.38111467 64438.8203125 63827.90625 61270.8984375
60892.890625 61533.6796875 62313.30859375] -> Prediction: 63516.15234375
Predicting on:
[64438.8203125 63827.90625 61270.8984375 60892.890625
61533.6796875 62313.30859375 63516.15234375] -> Prediction: 63653.12890625
Predicting on:
[63827.90625 61270.8984375 60892.890625 61533.6796875
62313.30859375 63516.15234375 63653.12890625] -> Prediction: 62261.9921875
Predicting on:
[61270.8984375 60892.890625 61533.6796875 62313.30859375
63516.15234375 63653.12890625 62261.9921875 ] -> Prediction: 61014.31640625
Predicting on:
[60892.890625 61533.6796875 62313.30859375 63516.15234375
63653.12890625 62261.9921875 61014.31640625] -> Prediction: 61434.3984375
Predicting on:
[61533.6796875 62313.30859375 63516.15234375 63653.12890625
62261.9921875 61014.31640625 61434.3984375 ] -> Prediction: 62239.453125
Predicting on:
[62313.30859375 63516.15234375 63653.12890625 62261.9921875
61014.31640625 61434.3984375 62239.453125 ] -> Prediction: 62905.96484375
Predicting on:
[63516.15234375 63653.12890625 62261.9921875 61014.31640625
61434.3984375 62239.453125 62905.96484375] -> Prediction: 63236.1484375Plotting the predictions into the future
def get_future_dates(start_date, into_future, offset=1):
"""
Returns array of datetime values from ranging from start_date to start_date+horizon.
start_date: date to start range (np.datetime64)
into_future: number of days to add onto start date for range (int)
offset: number of days to offset start_date by (default 1)
"""
start_date = start_date + np.timedelta64(offset, "D") # specify start date, "D" stands for day
end_date = start_date + np.timedelta64(into_future, "D") # specify end date
return np.arange(start_date, end_date, dtype="datetime64[D]") # return a date range between start date and end date# Last timestep of timesteps (currently in np.datetime64 format)
last_timestep = bitcoin_prices.index[-1]
last_timestepTimestamp('2021-10-25 00:00:00')# Get next two weeks of timesteps
next_time_steps = get_future_dates(start_date=last_timestep,
into_future=INTO_FUTURE)
next_time_stepsarray(['2021-10-26', '2021-10-27', '2021-10-28', '2021-10-29',
'2021-10-30', '2021-10-31', '2021-11-01', '2021-11-02',
'2021-11-03', '2021-11-04', '2021-11-05', '2021-11-06',
'2021-11-07', '2021-11-08'], dtype='datetime64[D]')# Insert last timestep/final price so the graph doesn't look messed
next_time_steps = np.insert(next_time_steps, 0, last_timestep)
future_forecast = np.insert(future_forecast, 0, prices[-1])
next_time_steps, future_forecast(array(['2021-10-25', '2021-10-26', '2021-10-27', '2021-10-28',
'2021-10-29', '2021-10-30', '2021-10-31', '2021-11-01',
'2021-11-02', '2021-11-03', '2021-11-04', '2021-11-05',
'2021-11-06', '2021-11-07', '2021-11-08'], dtype='datetime64[D]'),
array([63004.383, 64438.82 , 63827.906, 61270.9 , 60892.89 , 61533.68 ,
62313.31 , 63516.152, 63653.13 , 62261.992, 61014.316, 61434.4 ,
62239.453, 62905.965, 63236.15 ], dtype=float32))# Plot future price predictions of Bitcoin
plt.figure(figsize=(10, 7))
plot_time_series(bitcoin_prices.index, prices, start=2500, format="-", label="Actual BTC Price")
plot_time_series(next_time_steps, future_forecast, format="-", label="Predicted BTC Price")Model 10 - turkey data
Showing why forecasting is BS
One single unpredictable datapoint can ruin everything
# Let's introduce a Turkey problem to our BTC data (price BTC falls 100x in one day)
btc_price_turkey = prices.copy()
btc_price_turkey[-1] = btc_price_turkey[-1] / 100# Manufacture an extra price on the end (to showcase the Turkey problem)
btc_price_turkey[-10:][60744.9849380446,
60965.4654336171,
61707.4813118213,
63901.0730802098,
65972.8631297649,
62603.5750699116,
60689.2382648626,
61124.3471256221,
60936.1508514192,
630.043811146675]# Get the timesteps for the turkey problem
btc_timesteps_turkey = np.array(bitcoin_prices.index)
btc_timesteps_turkey[-10:]array(['2021-10-16T00:00:00.000000000', '2021-10-17T00:00:00.000000000',
'2021-10-18T00:00:00.000000000', '2021-10-19T00:00:00.000000000',
'2021-10-20T00:00:00.000000000', '2021-10-21T00:00:00.000000000',
'2021-10-22T00:00:00.000000000', '2021-10-23T00:00:00.000000000',
'2021-10-24T00:00:00.000000000', '2021-10-25T00:00:00.000000000'],
dtype='datetime64[ns]')plt.figure(figsize=(10, 7))
plot_time_series(timesteps=btc_timesteps_turkey,
values=btc_price_turkey,
format="-",
label="BTC Price + Turkey Problem",
start=2500)# Create train and test sets for turkey problem data
full_windows, full_labels = make_windows(np.array(btc_price_turkey), window_size=WINDOW_SIZE, horizon=HORIZON)
len(full_windows), len(full_labels)
X_train, X_test, y_train, y_test = make_train_test_splits(full_windows, full_labels)
len(X_train), len(X_test), len(y_train), len(y_test)(2032, 509, 2032, 509)# Clone model 1 architecture for turkey model and fit the turkey model on the turkey data
turkey_model = tf.keras.models.clone_model(model_1)
turkey_model._name = "Turkey_Model"
turkey_model.compile(loss="mae",
optimizer=tf.keras.optimizers.Adam())
turkey_model.fit(X_train, y_train,
epochs=100,
verbose=0,
validation_data=(X_test, y_test),
callbacks=[create_model_checkpoint(turkey_model.name)])INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
INFO:tensorflow:Assets written to: model_experiments/Turkey_Model/assets
<keras.callbacks.History at 0x7f9dfcec6850># Load best model and evaluate on test data
turkey_model = tf.keras.models.load_model("model_experiments/Turkey_Model/")
turkey_model.evaluate(X_test, y_test)16/16 [==============================] - 0s 3ms/step - loss: 1034.4946
1034.49462890625# Make predictions with Turkey model
turkey_preds = make_preds(turkey_model, X_test)
turkey_preds[:10]<tf.Tensor: shape=(10,), dtype=float32, numpy=
array([9660.122, 9827.668, 9672.227, 9759.875, 9710.475, 9807.82 ,
9819.334, 9879.259, 9335.341, 9451.662], dtype=float32)># Evaluate turkey preds
turkey_results = evaluate_preds(y_true=y_test,
y_pred=turkey_preds)
turkey_results{'mae': 20278.744,
'mape': 108.450264,
'mase': 19.489779,
'mse': 633565600.0,
'rmse': 24717.24}And with just one value change, our error metrics go through the roof.
To make sure, let's remind ourselves of how model_1 went on unmodified Bitcoin data (no turkey problem).
# Comparing to model 1
model_1_results{'mae': 933.08374,
'mape': 2.6302116,
'mase': 1.0078244,
'mse': 2106266.2,
'rmse': 1451.2981}plt.figure(figsize=(10, 7))
# plot_time_series(timesteps=btc_timesteps_turkey[:split_size], values=btc_price_turkey[:split_size], label="Train Data")
offset=300
plot_time_series(timesteps=btc_timesteps_turkey[-len(X_test):],
values=btc_price_turkey[-len(y_test):],
format="-",
label="Turkey Test Data", start=offset)
plot_time_series(timesteps=btc_timesteps_turkey[-len(X_test):],
values=turkey_preds,
label="Turkey Preds",
start=offset);Highly unlikely price movements (based on historical movements), upward or downward will likely never be part of a forecast.
However, as we've seen, despite their unlikeliness, these events can happen and will have huuuuuuuuge impacts to the performance of our models.
Comparing the models
# Compare different model results (w = window, h = horizon, e.g. w=7 means a window size of 7)
model_results = pd.DataFrame({"naive_model": naive_results,
"model_1_dense_w7_h1": model_1_results,
"model_2_dense_w30_h1": model_2_results,
"model_3_dense_w30_h7": model_3_results,
"model_4_CONV1D": model_4_results,
"model_5_LSTM": model_5_results,
"model_6_multivariate": model_6_results,
"model_8_NBEATs": model_7_results,
"model_9_ensemble": ensemble_results,
"model_10_turkey": turkey_results}).T
model_results.head(10).dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; }
# Sort model results by MAE and plot them
model_results[["mae"]].sort_values(by="mae").plot(figsize=(10, 7), kind="bar");Last updated
