84 lines
2.9 KiB
Python
84 lines
2.9 KiB
Python
|
import numpy as np
|
||
|
|
import tensorflow as tf
|
||
|
|
from tensorflow.keras import layers, models, callbacks
|
||
|
|
|
||
|
|
def build_lstm_model(input_shape):
|
||
|
|
"""
|
||
|
|
Builds a robust LSTM classification network with regularizers
|
||
|
|
to handle noisy financial time-series data.
|
||
|
|
"""
|
||
|
|
model = models.Sequential([
|
||
|
|
# 1. First LSTM Layer: Returns sequences to feed into the next layer
|
||
|
|
layers.Input(shape=input_shape),
|
||
|
|
layers.LSTM(units=64, return_sequences=True,
|
||
|
|
kernel_regularizer=tf.keras.regularizers.l2(0.001)),
|
||
|
|
layers.Dropout(0.3), # High dropout to reduce overfitting
|
||
|
|
|
||
|
|
# 2. Second LSTM Layer: Returns only the last step's hidden state
|
||
|
|
layers.LSTM(units=32, return_sequences=False,
|
||
|
|
kernel_regularizer=tf.keras.regularizers.l2(0.001)),
|
||
|
|
layers.Dropout(0.3),
|
||
|
|
|
||
|
|
# 3. Dense Hidden Layer for abstract feature synthesis
|
||
|
|
layers.Dense(units=16, activation='relu'),
|
||
|
|
layers.BatchNormalization(), # Stabilizes training weights
|
||
|
|
|
||
|
|
# 4. Output Layer: Single neuron outputting a probability (0.0 to 1.0)
|
||
|
|
layers.Dense(units=1, activation='sigmoid')
|
||
|
|
])
|
||
|
|
|
||
|
|
# Compile with Binary Crossentropy for the binary UP/DOWN classification task
|
||
|
|
model.compile(
|
||
|
|
optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005),
|
||
|
|
loss='binary_crossentropy',
|
||
|
|
metrics=['accuracy', tf.keras.metrics.AUC(name='auc')]
|
||
|
|
)
|
||
|
|
|
||
|
|
return model
|
||
|
|
|
||
|
|
def train_trading_model(X, y):
|
||
|
|
"""Handles data splitting, model creation, and training with early stopping."""
|
||
|
|
|
||
|
|
# 1. Chronological Train/Test Split (Crucial for Time-Series! Do NOT random split)
|
||
|
|
# Using the last 20% of chronological data for evaluation to avoid data leakage
|
||
|
|
split_idx = int(len(X) * 0.8)
|
||
|
|
X_train, X_val = X[:split_idx], X[split_idx:]
|
||
|
|
y_train, y_val = y[:split_idx], y[split_idx:]
|
||
|
|
|
||
|
|
# 2. Extract dimensions for input shape: (Time Steps, Features)
|
||
|
|
input_shape = (X.shape[1], X.shape[2])
|
||
|
|
model = build_lstm_model(input_shape)
|
||
|
|
model.summary()
|
||
|
|
|
||
|
|
# 3. Set up Training Callbacks
|
||
|
|
# Early Stopping prevents the network from memorizing noise when val loss stalls
|
||
|
|
custom_callbacks = [
|
||
|
|
callbacks.EarlyStopping(
|
||
|
|
monitor='val_loss',
|
||
|
|
patience=5,
|
||
|
|
restore_best_weights=True
|
||
|
|
),
|
||
|
|
callbacks.ReduceLROnPlateau(
|
||
|
|
monitor='val_loss',
|
||
|
|
factor=0.5,
|
||
|
|
patience=3
|
||
|
|
)
|
||
|
|
]
|
||
|
|
|
||
|
|
# 4. Fit the Model
|
||
|
|
print("\nStarting model training...")
|
||
|
|
history = model.fit(
|
||
|
|
X_train, y_train,
|
||
|
|
validation_data=(X_val, y_val),
|
||
|
|
epochs=30,
|
||
|
|
batch_size=64,
|
||
|
|
callbacks=custom_callbacks,
|
||
|
|
verbose=1
|
||
|
|
)
|
||
|
|
|
||
|
|
return model, history
|
||
|
|
|
||
|
|
# --- PIPELINE INTEGRATION EXAMPLE ---
|
||
|
|
# trained_model, training_history = train_trading_model(X, y)
|
||
|
|
# trained_model.save("portfolio_lstm_bot.h5")
|