
from google.colab import drive
drive.mount('/content/drive')

Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).


from IPython.utils import io
with io.capture_output() as captured:
  !pip install yfinance
  !pip install git+https://github.com/tensorflow/docs
  import tensorflow_docs as tfdocs
  import tensorflow_docs.modeling
  import sklearn.preprocessing
  import tensorflow as tf
  import os


os.chdir('/content/drive/MyDrive/Colab Notebooks/Crypto')

from data_loader import FinanceFetch
from data_process import DataPreProcess
from models_wrap import ModelsHandler
from models_dispatcher import ModelsDispatch

# see all the code here https://github.com/GabrieleGhisleni/Stock-Prediction-with-Multivariate-LSTM


ticket = "AAPL"
TicketHandler = FinanceFetch(ticket)
historical_data = TicketHandler.fetch('max', plot=True, year='2021')


TIMESTEPS = 20

dt = DataPreProcess(scaler=sklearn.preprocessing.StandardScaler, timesteps=TIMESTEPS)
scaled_data = dt.scale(data=historical_data.loc[:, ['Open','Close','Volume']], 
                       target=historical_data.Close, 
                       timestamp=historical_data.Date)
train, test = dt.test(scaled_data, N=100)

# creating windows
X_train, y_train = dt.create_windows(train, col_x=['Open','Close','Volume'], col_y=['Close'])
X_test, y_test= dt.create_windows(test, col_x=['Open','Close','Volume'], col_y=['Close'])

Shape of transormed data: x=(10212, 20, 3), y=(10212, 1)
Shape of transormed data: x=(80, 20, 3), y=(80, 1)


model_handler = ModelsDispatch(batch_size=32)
model = model_handler.light_dense_vanilla_LSTM(input_shape=(TIMESTEPS, 3), name='light dense vanilla_199-10lr_1') 
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=.0001), 
              loss='mse', 
              metrics=[tf.keras.metrics.MSE, tf.keras.metrics.MAE])

model_handler.draw_model(model, ticket=ticket)


EPOCH = 500
early_stop=tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=50)
history_model = model.fit(x=X_train, 
                          y=y_train, 
                          batch_size=32, 
                          epochs=EPOCH, 
                          shuffle=True, 
                          verbose=0,
                          callbacks=[early_stop, tfdocs.modeling.EpochDots()],
                          validation_split = 0.1)

Epoch: 0, loss:0.0099,  mean_absolute_error:0.0463,  mean_squared_error:0.0099,  val_loss:1.3241,  val_mean_absolute_error:0.7494,  val_mean_squared_error:1.3241,  
....................................................................................................
Epoch: 100, loss:0.0000,  mean_absolute_error:0.0027,  mean_squared_error:0.0000,  val_loss:0.0055,  val_mean_absolute_error:0.0475,  val_mean_squared_error:0.0055,  
.................


model_handler.plot_history_model(history_model)


best_model = model_handler.compare_eval(ticket).iloc[0]['model']
model = model_handler.load_model(ticket, best_model)


model_handler.plot_test_pred(model, X_test, y_test, dt.scaler_target)


res = model_handler.save_model_and_score(model, X_test, y_test, ticket=ticket, target_scalar=dt.scaler_target)

% Direction correct=0.6329113924050633, mse_test_mean=15.308696160605766

WARNING:absl:Found untraced functions such as lstm_cell_1_layer_call_fn, lstm_cell_1_layer_call_and_return_conditional_losses, lstm_cell_1_layer_call_fn, lstm_cell_1_layer_call_and_return_conditional_losses, lstm_cell_1_layer_call_and_return_conditional_losses while saving (showing 5 of 5). These functions will not be directly callable after loading.
INFO:tensorflow:Assets written to: ./models//AAPL/light dense vanilla 200-10lr_1.json/assets
WARNING:absl:<keras.layers.recurrent.LSTMCell object at 0x7f03d7bb3f50> 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.


all_X, all_y = dt.create_windows(scaled_data.iloc[-1000:], col_x=['Open','Close','Volume'], col_y=['Close'])
model_handler.plot_test_pred(model, all_X, all_y, dt.scaler_target, H=len(X_test), lw_1=1, lw_2=1)

Shape of transormed data: x=(980, 20, 3), y=(980, 1)


model_handler.compare_eval(ticket).sort_values('mse_mean_test')

	model	mse	mae	correct_dir	mse_mean_test	mse_std_test
7	light dense vanilla 200-10lr_1	0.0212117	0.126138	0.632911	15.308696	15.639879
18	light dense vanilla_201-10lr_1	0.0434222	0.191429	0.594937	31.374742	23.616368
13	light vanilla 205-15lr_1	0.0714554	0.250191	0.564246	51.308984	35.459333
19	light dense vanilla_199-10lr_1	0.0898177	0.286559	0.582278	64.799184	36.256184
12	light vanilla 185-7lr_1	0.103032	0.308008	0.564246	74.104104	41.757892
16	light vanilla 215-10lr_1	0.103675	0.309733	0.564246	74.537986	41.744812
3	heavy vanilla 200-100_10_1	0.116483	0.330142	0.582278	84.600379	43.307695
15	light vanilla 225-10lr_1	0.135721	0.357108	0.564246	97.623844	49.059649
2	heavy vanilla 100_10_1	0.219492	0.460594	0.582278	158.573406	60.272696
10	light vanilla 150-10lr_1	0.290634	0.529499	0.582278	208.431692	82.416254
14	light vanilla 220-10lr_1	0.358054	0.534491	0.564246	256.491362	212.639571
5	heavy vanilla 100-25r_10_1	0.375745	0.606192	0.582278	271.344044	84.345130
17	light dense vanilla_205-10lr_1	0.415319	0.568618	0.564246	297.133893	253.476319
11	light vanilla 200-20s_10s_1	0.664097	0.800117	0.564246	476.615989	180.834976
0	bidirectional_LSTM	0.726115	0.843575	0.582278	521.398706	157.864803
8	light vanilla 30-3lr_1	0.824454	0.895981	0.582278	588.180252	200.044536
4	heavy vanilla 200-100_50_25_10_1_dropped	2.04147	1.42339	0.582278	1468.744119	266.175212
1	heavy bidirectional 50-20_20_20_1	2.40076	1.54104	0.582278	1722.063373	371.600633
9	light vanilla 300-3lr_1	2.49422	1.57185	0.582278	1791.837785	350.851445
6	heavy vanilla 100-75s_10r_1	14.579	3.81233	0.582278	10474.441939	1181.262067

Stock prediction with Multivariate-LSTM