pushing this all to git before cleaing it. putting this thing on my fucking Dell Poweredge bc god knows im not waiting 2 hours to run another ML test on this 12 core machine
This commit is contained in:
klein panic
472
src/Machine-Learning/LSTM-python/main.py
Normal file
472
src/Machine-Learning/LSTM-python/main.py
Normal file
@@ -0,0 +1,472 @@
|
||||
import os
|
||||
import sys
|
||||
import argparse
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
import matplotlib.pyplot as plt
|
||||
import seaborn as sns
|
||||
import logging
|
||||
from tabulate import tabulate
|
||||
|
||||
from sklearn.preprocessing import MinMaxScaler
|
||||
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
|
||||
from sklearn.model_selection import TimeSeriesSplit, GridSearchCV
|
||||
|
||||
import tensorflow as tf
|
||||
from tensorflow.keras.models import Sequential
|
||||
from tensorflow.keras.layers import LSTM, Dense, Dropout, Bidirectional
|
||||
from tensorflow.keras.optimizers import Adam, Nadam
|
||||
from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau
|
||||
from tensorflow.keras.losses import Huber
|
||||
|
||||
import xgboost as xgb
|
||||
import optuna
|
||||
from optuna.integration import KerasPruningCallback
|
||||
|
||||
# Reinforcement Learning
|
||||
import gym
|
||||
from gym import spaces
|
||||
from stable_baselines3 import DQN
|
||||
from stable_baselines3.common.vec_env import DummyVecEnv
|
||||
|
||||
# Suppress TensorFlow warnings
|
||||
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
|
||||
|
||||
# Configure logging
|
||||
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
|
||||
|
||||
##############################
|
||||
# 1. Data Loading & Indicators
|
||||
##############################
|
||||
def load_data(file_path):
|
||||
logging.info(f"Loading data from: {file_path}")
|
||||
try:
|
||||
data = pd.read_csv(file_path, parse_dates=['time'])
|
||||
except FileNotFoundError:
|
||||
logging.error(f"File not found: {file_path}")
|
||||
sys.exit(1)
|
||||
except pd.errors.ParserError as e:
|
||||
logging.error(f"Error parsing CSV file: {e}")
|
||||
sys.exit(1)
|
||||
except Exception as e:
|
||||
logging.error(f"Unexpected error: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
rename_mapping = {
|
||||
'time': 'Date',
|
||||
'open': 'Open',
|
||||
'high': 'High',
|
||||
'low': 'Low',
|
||||
'close': 'Close'
|
||||
}
|
||||
data.rename(columns=rename_mapping, inplace=True)
|
||||
|
||||
# Sort by Date
|
||||
data.sort_values('Date', inplace=True)
|
||||
data.reset_index(drop=True, inplace=True)
|
||||
logging.info(f"Data columns after renaming: {data.columns.tolist()}")
|
||||
logging.info("Data loaded and sorted successfully.")
|
||||
|
||||
return data
|
||||
|
||||
def compute_rsi(series, window=14):
|
||||
delta = series.diff()
|
||||
gain = delta.where(delta > 0, 0).rolling(window=window).mean()
|
||||
loss = -delta.where(delta < 0, 0).rolling(window=window).mean()
|
||||
RS = gain / loss
|
||||
return 100 - (100 / (1 + RS))
|
||||
|
||||
def compute_macd(series, span_short=12, span_long=26, span_signal=9):
|
||||
ema_short = series.ewm(span=span_short, adjust=False).mean()
|
||||
ema_long = series.ewm(span=span_long, adjust=False).mean()
|
||||
macd_line = ema_short - ema_long
|
||||
signal_line = macd_line.ewm(span=span_signal, adjust=False).mean()
|
||||
return macd_line - signal_line # MACD histogram
|
||||
|
||||
def compute_adx(df, window=14):
|
||||
"""
|
||||
Example ADX calculation (pseudo-real):
|
||||
You can implement a full ADX formula if you’d like.
|
||||
Here, we do a slightly more robust approach than rolling std.
|
||||
"""
|
||||
# True range
|
||||
df['H-L'] = df['High'] - df['Low']
|
||||
df['H-Cp'] = (df['High'] - df['Close'].shift(1)).abs()
|
||||
df['L-Cp'] = (df['Low'] - df['Close'].shift(1)).abs()
|
||||
tr = df[['H-L', 'H-Cp', 'L-Cp']].max(axis=1)
|
||||
tr_rolling = tr.rolling(window=window).mean()
|
||||
|
||||
# Simplistic to replicate ADX-like effect
|
||||
adx_placeholder = tr_rolling / df['Close']
|
||||
df.drop(['H-L','H-Cp','L-Cp'], axis=1, inplace=True)
|
||||
return adx_placeholder
|
||||
|
||||
def compute_obv(df):
|
||||
# On-Balance Volume
|
||||
signed_volume = (np.sign(df['Close'].diff()) * df['Volume']).fillna(0)
|
||||
return signed_volume.cumsum()
|
||||
|
||||
def compute_bollinger_bands(series, window=20, num_std=2):
|
||||
"""
|
||||
Bollinger Bands: middle=MA, upper=MA+2*std, lower=MA-2*std
|
||||
Return the band width or separate columns.
|
||||
"""
|
||||
sma = series.rolling(window=window).mean()
|
||||
std = series.rolling(window=window).std()
|
||||
upper = sma + num_std * std
|
||||
lower = sma - num_std * std
|
||||
bandwidth = (upper - lower) / sma # optional metric
|
||||
return upper, lower, bandwidth
|
||||
|
||||
def compute_mfi(df, window=14):
|
||||
"""
|
||||
Money Flow Index: uses typical price, volume, direction.
|
||||
For demonstration.
|
||||
"""
|
||||
typical_price = (df['High'] + df['Low'] + df['Close']) / 3
|
||||
money_flow = typical_price * df['Volume']
|
||||
# Positive or negative
|
||||
df_shift = typical_price.shift(1)
|
||||
flow_positive = money_flow.where(typical_price > df_shift, 0)
|
||||
flow_negative = money_flow.where(typical_price < df_shift, 0)
|
||||
# Sum over window
|
||||
pos_sum = flow_positive.rolling(window=window).sum()
|
||||
neg_sum = flow_negative.rolling(window=window).sum()
|
||||
# RSI-like formula
|
||||
mfi = 100 - (100 / (1 + pos_sum / (neg_sum + 1e-9)))
|
||||
return mfi
|
||||
|
||||
def calculate_technical_indicators(df):
|
||||
logging.info("Calculating technical indicators...")
|
||||
|
||||
df['RSI'] = compute_rsi(df['Close'], window=14)
|
||||
df['MACD'] = compute_macd(df['Close'])
|
||||
df['OBV'] = compute_obv(df)
|
||||
df['ADX'] = compute_adx(df)
|
||||
|
||||
# Bollinger
|
||||
upper_bb, lower_bb, bb_width = compute_bollinger_bands(df['Close'], window=20)
|
||||
df['BB_Upper'] = upper_bb
|
||||
df['BB_Lower'] = lower_bb
|
||||
df['BB_Width'] = bb_width
|
||||
|
||||
# MFI
|
||||
df['MFI'] = compute_mfi(df)
|
||||
|
||||
# Simple/EMA
|
||||
df['SMA_5'] = df['Close'].rolling(window=5).mean()
|
||||
df['SMA_10'] = df['Close'].rolling(window=10).mean()
|
||||
df['EMA_5'] = df['Close'].ewm(span=5, adjust=False).mean()
|
||||
df['EMA_10'] = df['Close'].ewm(span=10, adjust=False).mean()
|
||||
# STD
|
||||
df['STDDEV_5'] = df['Close'].rolling(window=5).std()
|
||||
|
||||
df.dropna(inplace=True)
|
||||
logging.info("Technical indicators calculated successfully.")
|
||||
return df
|
||||
|
||||
##############################
|
||||
# 2. Parse Arguments
|
||||
##############################
|
||||
def parse_arguments():
|
||||
parser = argparse.ArgumentParser(description='Train LSTM and DQN models for stock trading.')
|
||||
parser.add_argument('csv_path', type=str, help='Path to the CSV data file.')
|
||||
return parser.parse_args()
|
||||
|
||||
##############################
|
||||
# 3. Main
|
||||
##############################
|
||||
def main():
|
||||
args = parse_arguments()
|
||||
csv_path = args.csv_path
|
||||
|
||||
# 1) Load data
|
||||
data = load_data(csv_path)
|
||||
data = calculate_technical_indicators(data)
|
||||
|
||||
# 2) Build feature set
|
||||
# We deliberately EXCLUDE 'Close' from the features so the model doesn't trivially see it.
|
||||
# Instead, rely on advanced indicators + OHLC + Volume.
|
||||
feature_columns = [
|
||||
'Open', 'High', 'Low', 'Volume',
|
||||
'RSI', 'MACD', 'OBV', 'ADX',
|
||||
'BB_Upper', 'BB_Lower', 'BB_Width',
|
||||
'MFI', 'SMA_5', 'SMA_10', 'EMA_5', 'EMA_10', 'STDDEV_5'
|
||||
]
|
||||
target_column = 'Close' # still used for label/evaluation
|
||||
|
||||
# Keep only these columns + Date + target
|
||||
data = data[['Date'] + feature_columns + [target_column]].dropna()
|
||||
|
||||
# 3) Scale data
|
||||
from sklearn.preprocessing import MinMaxScaler
|
||||
scaler_features = MinMaxScaler()
|
||||
scaler_target = MinMaxScaler()
|
||||
|
||||
scaled_features = scaler_features.fit_transform(data[feature_columns])
|
||||
scaled_target = scaler_target.fit_transform(data[[target_column]]).flatten()
|
||||
|
||||
# 4) Create sequences for LSTM
|
||||
def create_sequences(features, target, window_size=15):
|
||||
X, y = [], []
|
||||
for i in range(len(features) - window_size):
|
||||
X.append(features[i:i+window_size])
|
||||
y.append(target[i+window_size])
|
||||
return np.array(X), np.array(y)
|
||||
|
||||
window_size = 15
|
||||
X, y = create_sequences(scaled_features, scaled_target, window_size)
|
||||
|
||||
# 5) Train/Val/Test Split
|
||||
train_size = int(len(X) * 0.7)
|
||||
val_size = int(len(X) * 0.15)
|
||||
test_size = len(X) - train_size - val_size
|
||||
|
||||
X_train, X_val, X_test = (
|
||||
X[:train_size],
|
||||
X[train_size:train_size+val_size],
|
||||
X[train_size+val_size:]
|
||||
)
|
||||
y_train, y_val, y_test = (
|
||||
y[:train_size],
|
||||
y[train_size:train_size+val_size],
|
||||
y[train_size+val_size:]
|
||||
)
|
||||
|
||||
logging.info(f"X_train: {X_train.shape}, X_val: {X_val.shape}, X_test: {X_test.shape}")
|
||||
logging.info(f"y_train: {y_train.shape}, y_val: {y_val.shape}, y_test: {y_test.shape}")
|
||||
|
||||
# 6) Device Config
|
||||
def configure_device():
|
||||
gpus = tf.config.list_physical_devices('GPU')
|
||||
if gpus:
|
||||
try:
|
||||
for gpu in gpus:
|
||||
tf.config.experimental.set_memory_growth(gpu, True)
|
||||
logging.info(f"{len(gpus)} GPU(s) detected and configured.")
|
||||
except RuntimeError as e:
|
||||
logging.error(e)
|
||||
else:
|
||||
logging.info("No GPU detected, using CPU.")
|
||||
configure_device()
|
||||
|
||||
# 7) Build LSTM
|
||||
from tensorflow.keras.regularizers import l2
|
||||
def build_lstm(input_shape, units=128, dropout=0.3, lr=1e-3):
|
||||
model = Sequential()
|
||||
# Example: 2 stacked LSTM layers
|
||||
model.add(Bidirectional(LSTM(units, return_sequences=True, kernel_regularizer=l2(1e-4)), input_shape=input_shape))
|
||||
model.add(Dropout(dropout))
|
||||
model.add(Bidirectional(LSTM(units, return_sequences=False, kernel_regularizer=l2(1e-4))))
|
||||
model.add(Dropout(dropout))
|
||||
model.add(Dense(1, activation='linear'))
|
||||
|
||||
optimizer = Adam(learning_rate=lr)
|
||||
model.compile(loss=Huber(), optimizer=optimizer, metrics=['mae'])
|
||||
return model
|
||||
|
||||
# 8) Train LSTM (you can still do Optuna if you like, omitted here for brevity)
|
||||
model_lstm = build_lstm((X_train.shape[1], X_train.shape[2]), units=128, dropout=0.3, lr=1e-3)
|
||||
|
||||
early_stop = EarlyStopping(patience=15, restore_best_weights=True)
|
||||
reduce_lr = ReduceLROnPlateau(factor=0.5, patience=5, min_lr=1e-6)
|
||||
|
||||
model_lstm.fit(
|
||||
X_train, y_train,
|
||||
validation_data=(X_val, y_val),
|
||||
epochs=100,
|
||||
batch_size=32,
|
||||
callbacks=[early_stop, reduce_lr],
|
||||
verbose=1
|
||||
)
|
||||
|
||||
# 9) Evaluate
|
||||
def evaluate_lstm(model, X_test, y_test):
|
||||
y_pred_scaled = model.predict(X_test).flatten()
|
||||
# If we forcibly clamp predictions to [0,1], do so, else skip:
|
||||
y_pred_scaled = np.clip(y_pred_scaled, 0, 1)
|
||||
y_pred = scaler_target.inverse_transform(y_pred_scaled.reshape(-1,1)).flatten()
|
||||
y_true = scaler_target.inverse_transform(y_test.reshape(-1,1)).flatten()
|
||||
|
||||
mse = mean_squared_error(y_true, y_pred)
|
||||
rmse = np.sqrt(mse)
|
||||
mae = mean_absolute_error(y_true, y_pred)
|
||||
r2 = r2_score(y_true, y_pred)
|
||||
|
||||
# Direction
|
||||
direction_true = np.sign(np.diff(y_true))
|
||||
direction_pred = np.sign(np.diff(y_pred))
|
||||
directional_acc = np.mean(direction_true == direction_pred)
|
||||
|
||||
logging.info(f"LSTM Test -> MSE={mse:.4f}, RMSE={rmse:.4f}, MAE={mae:.4f}, R2={r2:.4f}, DirAcc={directional_acc:.4f}")
|
||||
|
||||
# Quick Plot
|
||||
plt.figure(figsize=(12,6))
|
||||
plt.plot(y_true[:100], label='Actual')
|
||||
plt.plot(y_pred[:100], label='Predicted')
|
||||
plt.title("LSTM: Actual vs Predicted (first 100 test points)")
|
||||
plt.legend()
|
||||
plt.savefig("lstm_actual_vs_pred.png")
|
||||
plt.close()
|
||||
|
||||
evaluate_lstm(model_lstm, X_test, y_test)
|
||||
|
||||
# Save
|
||||
model_lstm.save("improved_lstm_model.keras")
|
||||
import joblib
|
||||
joblib.dump(scaler_features, "improved_scaler_features.pkl")
|
||||
joblib.dump(scaler_target, "improved_scaler_target.pkl")
|
||||
|
||||
##############################
|
||||
# 10) Reinforcement Learning
|
||||
##############################
|
||||
class StockTradingEnv(gym.Env):
|
||||
"""
|
||||
Improved RL Env that:
|
||||
- excludes raw 'Close' from observation
|
||||
- includes transaction cost (optional)
|
||||
- uses step-based PnL as reward
|
||||
"""
|
||||
metadata = {'render.modes': ['human']}
|
||||
def __init__(self, df, initial_balance=10000, transaction_cost=0.001):
|
||||
super().__init__()
|
||||
|
||||
self.df = df.reset_index(drop=True)
|
||||
self.initial_balance = initial_balance
|
||||
self.balance = initial_balance
|
||||
self.net_worth = initial_balance
|
||||
self.current_step = 0
|
||||
self.max_steps = len(df)
|
||||
|
||||
# Add transaction cost in decimal form (0.001 => 0.1%)
|
||||
self.transaction_cost = transaction_cost
|
||||
|
||||
self.shares_held = 0
|
||||
self.cost_basis = 0
|
||||
|
||||
# Suppose we exclude 'Close' from features to remove direct see of final price
|
||||
self.obs_columns = [
|
||||
'Open', 'High', 'Low', 'Volume',
|
||||
'RSI', 'MACD', 'OBV', 'ADX',
|
||||
'BB_Upper', 'BB_Lower', 'BB_Width',
|
||||
'MFI', 'SMA_5', 'SMA_10', 'EMA_5', 'EMA_10', 'STDDEV_5'
|
||||
]
|
||||
|
||||
# We'll normalize features with the same scaler used for LSTM. If you want EXACT same scaling:
|
||||
# you can pass the same 'scaler_features' object into this environment.
|
||||
self.scaler = MinMaxScaler().fit(df[self.obs_columns])
|
||||
# Or load from a pkl if you prefer: joblib.load("improved_scaler_features.pkl")
|
||||
|
||||
self.action_space = spaces.Discrete(3) # 0=Sell, 1=Hold, 2=Buy
|
||||
self.observation_space = spaces.Box(
|
||||
low=0.0, high=1.0,
|
||||
shape=(len(self.obs_columns) + 3,), # + balance, shares, cost_basis
|
||||
dtype=np.float32
|
||||
)
|
||||
|
||||
def reset(self):
|
||||
self.balance = self.initial_balance
|
||||
self.net_worth = self.initial_balance
|
||||
self.shares_held = 0
|
||||
self.cost_basis = 0
|
||||
self.current_step = 0
|
||||
return self._get_obs()
|
||||
|
||||
def step(self, action):
|
||||
# Current row
|
||||
row = self.df.iloc[self.current_step]
|
||||
current_price = row['Close']
|
||||
|
||||
prev_net_worth = self.net_worth
|
||||
|
||||
if action == 2: # Buy
|
||||
shares_bought = int(self.balance // current_price)
|
||||
if shares_bought > 0:
|
||||
cost = shares_bought * current_price
|
||||
fee = cost * self.transaction_cost
|
||||
self.balance -= (cost + fee)
|
||||
# Weighted average cost basis
|
||||
prev_shares = self.shares_held
|
||||
self.shares_held += shares_bought
|
||||
self.cost_basis = (
|
||||
(self.cost_basis * prev_shares) + (shares_bought * current_price)
|
||||
) / self.shares_held
|
||||
|
||||
elif action == 0: # Sell
|
||||
if self.shares_held > 0:
|
||||
revenue = self.shares_held * current_price
|
||||
fee = revenue * self.transaction_cost
|
||||
self.balance += (revenue - fee)
|
||||
self.shares_held = 0
|
||||
self.cost_basis = 0
|
||||
|
||||
# Recompute net worth
|
||||
self.net_worth = self.balance + self.shares_held * current_price
|
||||
self.current_step += 1
|
||||
done = (self.current_step >= self.max_steps - 1)
|
||||
|
||||
# *Step-based* reward => daily PnL
|
||||
reward = self.net_worth - prev_net_worth
|
||||
|
||||
obs = self._get_obs()
|
||||
return obs, reward, done, {}
|
||||
|
||||
def _get_obs(self):
|
||||
row = self.df.iloc[self.current_step][self.obs_columns]
|
||||
# Scale
|
||||
scaled = self.scaler.transform([row])[0]
|
||||
|
||||
additional = np.array([
|
||||
self.balance / self.initial_balance,
|
||||
self.shares_held / 100.0,
|
||||
self.cost_basis / (self.initial_balance+1e-9)
|
||||
], dtype=np.float32)
|
||||
|
||||
obs = np.concatenate([scaled, additional]).astype(np.float32)
|
||||
return obs
|
||||
|
||||
def render(self, mode='human'):
|
||||
profit = self.net_worth - self.initial_balance
|
||||
print(f"Step: {self.current_step}, "
|
||||
f"Balance: {self.balance:.2f}, "
|
||||
f"Shares: {self.shares_held}, "
|
||||
f"NetWorth: {self.net_worth:.2f}, "
|
||||
f"Profit: {profit:.2f}")
|
||||
|
||||
##############################
|
||||
# 11) Train DQN
|
||||
##############################
|
||||
def train_dqn(env):
|
||||
logging.info("Training DQN agent with improved environment...")
|
||||
model = DQN(
|
||||
'MlpPolicy',
|
||||
env,
|
||||
verbose=1,
|
||||
learning_rate=1e-3,
|
||||
buffer_size=50000,
|
||||
learning_starts=1000,
|
||||
batch_size=64,
|
||||
tau=0.99,
|
||||
gamma=0.99,
|
||||
train_freq=4,
|
||||
target_update_interval=1000,
|
||||
exploration_fraction=0.1,
|
||||
exploration_final_eps=0.02,
|
||||
tensorboard_log="./dqn_enhanced_tensorboard/"
|
||||
)
|
||||
model.learn(total_timesteps=50000)
|
||||
model.save("improved_dqn_agent")
|
||||
return model
|
||||
|
||||
# Initialize environment with the same data
|
||||
# *In a real scenario, you might feed a different dataset or do a train/test split
|
||||
# for the RL environment, too.
|
||||
rl_env = StockTradingEnv(data, initial_balance=10000, transaction_cost=0.001)
|
||||
vec_env = DummyVecEnv([lambda: rl_env])
|
||||
|
||||
dqn_model = train_dqn(vec_env)
|
||||
logging.info("Finished DQN training. You can test with a script like 'use_dqn.py' or do an internal test here.")
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
|
||||
Reference in New Issue
Block a user