Implement Regime Reversion Strategy and remove regime_detection.py

- Introduced `RegimeReversionStrategy` for ML-based regime detection and mean reversion trading.
- Added feature engineering and model training logic within the new strategy.
- Removed the deprecated `regime_detection.py` file to streamline the codebase.
- Updated the strategy factory to include the new regime strategy configuration.
This commit is contained in:
2026-01-13 21:55:34 +08:00
parent e6d69ed04d
commit 10bb371054
3 changed files with 294 additions and 384 deletions

View File

@@ -1,384 +0,0 @@
import sys
import os
from pathlib import Path
# Add project root to path
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import pandas as pd
import numpy as np
import ta
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, confusion_matrix
import plotly.graph_objects as go
from plotly.subplots import make_subplots
from engine.data_manager import DataManager
from engine.market import MarketType
def prepare_data(symbol_a="BTC-USDT", symbol_b="ETH-USDT", timeframe="1h", limit=None, start_date=None, end_date=None):
"""
Load and align data for two assets to create a pair.
"""
dm = DataManager()
print(f"Loading data for {symbol_a} and {symbol_b}...")
# Helper to load or download
def get_df(symbol):
try:
# Try load first
df = dm.load_data("okx", symbol, timeframe, MarketType.SPOT)
except Exception:
df = dm.download_data("okx", symbol, timeframe, market_type=MarketType.SPOT)
# If we have start/end dates, ensure we have enough data or re-download
if start_date:
mask_start = pd.Timestamp(start_date, tz='UTC')
if df.index.min() > mask_start:
print(f"Local data starts {df.index.min()}, need {mask_start}. Downloading...")
df = dm.download_data("okx", symbol, timeframe, start_date=start_date, end_date=end_date, market_type=MarketType.SPOT)
return df
df_a = get_df(symbol_a)
df_b = get_df(symbol_b)
# Filter by date if provided (to match CQ data range)
if start_date:
df_a = df_a[df_a.index >= pd.Timestamp(start_date, tz='UTC')]
df_b = df_b[df_b.index >= pd.Timestamp(start_date, tz='UTC')]
if end_date:
df_a = df_a[df_a.index <= pd.Timestamp(end_date, tz='UTC')]
df_b = df_b[df_b.index <= pd.Timestamp(end_date, tz='UTC')]
# Align DataFrames
print("Aligning data...")
common_index = df_a.index.intersection(df_b.index)
df_a = df_a.loc[common_index].copy()
df_b = df_b.loc[common_index].copy()
if limit:
df_a = df_a.tail(limit)
df_b = df_b.tail(limit)
return df_a, df_b
def load_cryptoquant_data(file_path: str) -> pd.DataFrame | None:
"""
Load CryptoQuant data and prepare it for merging.
"""
if not os.path.exists(file_path):
print(f"Warning: CQ data file {file_path} not found.")
return None
print(f"Loading CryptoQuant data from {file_path}...")
df = pd.read_csv(file_path, index_col='timestamp', parse_dates=True)
# CQ data is usually daily (UTC 00:00).
# Ensure index is timezone aware to match market data
if df.index.tz is None:
df.index = df.index.tz_localize('UTC')
return df
def calculate_features(df_a, df_b, cq_df=None, window=24):
"""
Calculate spread, z-score, and advanced regime features including CQ data.
"""
# 1. Price Ratio (Spread)
spread = df_b['close'] / df_a['close']
# 2. Rolling Statistics for Z-Score
rolling_mean = spread.rolling(window=window).mean()
rolling_std = spread.rolling(window=window).std()
z_score = (spread - rolling_mean) / rolling_std
# 3. Spread Momentum / Technicals
spread_rsi = ta.momentum.RSIIndicator(spread, window=14).rsi()
spread_roc = spread.pct_change(periods=5) * 100
# 4. Volume Dynamics
vol_ratio = df_b['volume'] / df_a['volume']
vol_ratio_ma = vol_ratio.rolling(window=12).mean()
# 5. Volatility Regime
ret_a = df_a['close'].pct_change()
ret_b = df_b['close'].pct_change()
vol_a = ret_a.rolling(window=window).std()
vol_b = ret_b.rolling(window=window).std()
vol_spread_ratio = vol_b / vol_a
# Create feature DataFrame
features = pd.DataFrame(index=spread.index)
features['spread'] = spread
features['z_score'] = z_score
features['spread_rsi'] = spread_rsi
features['spread_roc'] = spread_roc
features['vol_ratio'] = vol_ratio
features['vol_ratio_rel'] = vol_ratio / vol_ratio_ma
features['vol_diff_ratio'] = vol_spread_ratio
# 6. Merge CryptoQuant Data
if cq_df is not None:
print("Merging CryptoQuant features...")
# Forward fill daily data to hourly timestamps
# reindex features to match cq_df range or join
# Resample CQ to hourly (ffill)
# But easier: join features with cq_df using asof or reindex
cq_aligned = cq_df.reindex(features.index, method='ffill')
# Add derived CQ features
# Funding Diff: If ETH funding > BTC funding => ETH overheated
if 'btc_funding' in cq_aligned.columns and 'eth_funding' in cq_aligned.columns:
cq_aligned['funding_diff'] = cq_aligned['eth_funding'] - cq_aligned['btc_funding']
# Inflow Ratio: If ETH inflow >> BTC inflow => ETH dump incoming?
if 'btc_inflow' in cq_aligned.columns and 'eth_inflow' in cq_aligned.columns:
# Add small epsilon to avoid div by zero
cq_aligned['inflow_ratio'] = cq_aligned['eth_inflow'] / (cq_aligned['btc_inflow'] + 1)
features = features.join(cq_aligned)
# --- Refined Target Definition (Anytime Profit) ---
horizon = 6
threshold = 0.005 # 0.5% profit target
z_threshold = 1.0
# For Short Spread (Z > 1): Did it drop below target?
# We look for the MINIMUM spread in the next 'horizon' periods
future_min = features['spread'].rolling(window=horizon).min().shift(-horizon)
target_short = features['spread'] * (1 - threshold)
success_short = (features['z_score'] > z_threshold) & (future_min < target_short)
# For Long Spread (Z < -1): Did it rise above target?
# We look for the MAXIMUM spread in the next 'horizon' periods
future_max = features['spread'].rolling(window=horizon).max().shift(-horizon)
target_long = features['spread'] * (1 + threshold)
success_long = (features['z_score'] < -z_threshold) & (future_max > target_long)
conditions = [success_short, success_long]
features['target'] = np.select(conditions, [1, 1], default=0)
return features.dropna()
def train_regime_model(features):
"""
Train a Random Forest to predict mean reversion success.
"""
# Define excluded columns (targets, raw prices, intermediates)
exclude_cols = ['spread', 'horizon_ret', 'target', 'rolling_mean', 'rolling_std']
# Auto-select all other numeric columns as features
feature_cols = [c for c in features.columns if c not in exclude_cols]
# Handle NaN/Inf if any slipped through
X = features[feature_cols].replace([np.inf, -np.inf], np.nan).fillna(0)
y = features['target']
# Split Data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, shuffle=False)
print(f"\nTraining on {len(X_train)} samples, Testing on {len(X_test)} samples...")
print(f"Features used: {feature_cols}")
print(f"Class Balance (Target=1): {y.mean():.2%}")
# Model
model = RandomForestClassifier(
n_estimators=200,
max_depth=6,
min_samples_leaf=20,
class_weight='balanced_subsample',
random_state=42
)
model.fit(X_train, y_train)
# Evaluation
y_pred = model.predict(X_test)
y_prob = model.predict_proba(X_test)[:, 1]
print("\n--- Model Evaluation ---")
print(classification_report(y_test, y_pred))
# Feature Importance
importances = pd.Series(model.feature_importances_, index=feature_cols).sort_values(ascending=False)
print("\n--- Feature Importance ---")
print(importances)
return model, X_test, y_test, y_pred, y_prob
def plot_interactive_results(features, y_test, y_pred, y_prob):
"""
Create an interactive HTML plot using Plotly.
"""
print("\nGenerating interactive plot...")
test_idx = y_test.index
test_data = features.loc[test_idx].copy()
test_data['prob'] = y_prob
test_data['prediction'] = y_pred
test_data['actual'] = y_test
# Create Subplots
fig = make_subplots(
rows=3, cols=1,
shared_xaxes=True,
vertical_spacing=0.05,
row_heights=[0.5, 0.25, 0.25],
subplot_titles=('Spread & Signals', 'Exchange Inflows', 'Z-Score & Probability')
)
# Top: Spread
fig.add_trace(
go.Scatter(x=test_data.index, y=test_data['spread'], mode='lines', name='Spread', line=dict(color='gray')),
row=1, col=1
)
# Signals
# Separate Long and Short signals for clarity
# Logic: If Z-Score was High (>1), we were betting on a SHORT Spread (Reversion Down)
# If Z-Score was Low (< -1), we were betting on a LONG Spread (Reversion Up)
# Correct Short Signals (Green Triangle Down)
tp_short = test_data[(test_data['prediction'] == 1) & (test_data['actual'] == 1) & (test_data['z_score'] > 0)]
fig.add_trace(
go.Scatter(x=tp_short.index, y=tp_short['spread'], mode='markers', name='Win: Short Spread',
marker=dict(symbol='triangle-down', size=12, color='green')),
row=1, col=1
)
# Correct Long Signals (Green Triangle Up)
tp_long = test_data[(test_data['prediction'] == 1) & (test_data['actual'] == 1) & (test_data['z_score'] < 0)]
fig.add_trace(
go.Scatter(x=tp_long.index, y=tp_long['spread'], mode='markers', name='Win: Long Spread',
marker=dict(symbol='triangle-up', size=12, color='green')),
row=1, col=1
)
# False Short Signals (Red Triangle Down)
fp_short = test_data[(test_data['prediction'] == 1) & (test_data['actual'] == 0) & (test_data['z_score'] > 0)]
fig.add_trace(
go.Scatter(x=fp_short.index, y=fp_short['spread'], mode='markers', name='Loss: Short Spread',
marker=dict(symbol='triangle-down', size=10, color='red')),
row=1, col=1
)
# False Long Signals (Red Triangle Up)
fp_long = test_data[(test_data['prediction'] == 1) & (test_data['actual'] == 0) & (test_data['z_score'] < 0)]
fig.add_trace(
go.Scatter(x=fp_long.index, y=fp_long['spread'], mode='markers', name='Loss: Long Spread',
marker=dict(symbol='triangle-up', size=10, color='red')),
row=1, col=1
)
# Middle: Inflows (BTC vs ETH)
if 'btc_inflow' in test_data.columns:
fig.add_trace(
go.Bar(x=test_data.index, y=test_data['btc_inflow'], name='BTC Inflow', marker_color='orange', opacity=0.6),
row=2, col=1
)
if 'eth_inflow' in test_data.columns:
fig.add_trace(
go.Bar(x=test_data.index, y=test_data['eth_inflow'], name='ETH Inflow', marker_color='purple', opacity=0.6),
row=2, col=1
)
# Bottom: Z-Score
fig.add_trace(
go.Scatter(x=test_data.index, y=test_data['z_score'], mode='lines', name='Z-Score', line=dict(color='blue'), opacity=0.5),
row=3, col=1
)
fig.add_hline(y=2, line_dash="dash", line_color="red", row=3, col=1)
fig.add_hline(y=-2, line_dash="dash", line_color="green", row=3, col=1)
# Probability (Secondary Y for Row 3)
fig.add_trace(
go.Scatter(x=test_data.index, y=test_data['prob'], mode='lines', name='Prob', line=dict(color='cyan', width=1.5), yaxis='y4'),
row=3, col=1
)
fig.update_layout(
title='Regime Detection Analysis (with CryptoQuant)',
autosize=True,
height=None,
hovermode='x unified',
yaxis4=dict(title='Probability', overlaying='y3', side='right', range=[0, 1], showgrid=False),
template="plotly_dark",
margin=dict(l=10, r=10, t=40, b=10),
barmode='group'
)
# Update all x-axes to ensure spikes are visible everywhere
fig.update_xaxes(
showspikes=True,
spikemode='across',
spikesnap='cursor',
showline=False,
showgrid=True,
spikedash='dot',
spikecolor='white', # Make it bright to see
spikethickness=1,
)
fig.update_layout(
title='Regime Detection Analysis (with CryptoQuant)',
autosize=True,
height=None,
hovermode='x unified', # Keep unified hover for data reading
yaxis4=dict(title='Probability', overlaying='y3', side='right', range=[0, 1], showgrid=False),
template="plotly_dark",
margin=dict(l=10, r=10, t=40, b=10),
barmode='group'
)
output_path = "research/regime_results.html"
fig.write_html(
output_path,
config={'responsive': True, 'scrollZoom': True},
include_plotlyjs='cdn',
full_html=True,
default_height='100vh',
default_width='100%'
)
print(f"Interactive plot saved to {output_path}")
def main():
# 1. Load CQ Data first to determine valid date range
cq_path = "data/cq_training_data.csv"
cq_df = load_cryptoquant_data(cq_path)
start_date = None
end_date = None
if cq_df is not None and not cq_df.empty:
start_date = cq_df.index.min().strftime('%Y-%m-%d')
end_date = cq_df.index.max().strftime('%Y-%m-%d')
print(f"CryptoQuant Data Range: {start_date} to {end_date}")
# 2. Get Market Data (Aligned to CQ range)
df_btc, df_eth = prepare_data(
"BTC-USDT", "ETH-USDT",
timeframe="1h",
start_date=start_date,
end_date=end_date
)
# 3. Calculate Features
print("Calculating advanced regime features...")
data = calculate_features(df_btc, df_eth, cq_df=cq_df, window=24)
if data.empty:
print("Error: No overlapping data found between Price and CryptoQuant data.")
return
# 4. Train & Evaluate
model, X_test, y_test, y_pred, y_prob = train_regime_model(data)
# 5. Plot
plot_interactive_results(data, y_test, y_pred, y_prob)
if __name__ == "__main__":
main()

View File

@@ -36,6 +36,7 @@ def _build_registry() -> dict[str, StrategyConfig]:
# Import here to avoid circular imports
from strategies.examples import MaCrossStrategy, RsiStrategy
from strategies.supertrend import MetaSupertrendStrategy
from strategies.regime_strategy import RegimeReversionStrategy
return {
"rsi": StrategyConfig(
@@ -76,6 +77,19 @@ def _build_registry() -> dict[str, StrategyConfig]:
'period3': 12, 'multiplier3': 1.0
}
),
"regime": StrategyConfig(
strategy_class=RegimeReversionStrategy,
default_params={
'horizon': 96,
'z_window': 24,
'stop_loss': 0.06,
'take_profit': 0.05
},
grid_params={
'horizon': [72, 96, 120],
'stop_loss': [0.04, 0.06, 0.08]
}
)
}

View File

@@ -0,0 +1,280 @@
import pandas as pd
import numpy as np
import ta
import vectorbt as vbt
from sklearn.ensemble import RandomForestClassifier
from strategies.base import BaseStrategy
from engine.market import MarketType
from engine.data_manager import DataManager
from engine.logging_config import get_logger
logger = get_logger(__name__)
class RegimeReversionStrategy(BaseStrategy):
"""
ML-Based Regime Detection & Mean Reversion Strategy.
Logic:
1. Tracks the BTC/ETH Spread and its Z-Score (24h window).
2. Uses a Random Forest model to predict if an extreme Z-Score will revert profitably.
3. Features: Spread Technicals (RSI, ROC) + On-Chain Flows (Inflow, Funding).
4. Entry: When Model Probability > 0.5.
5. Exit: Z-Score reversion to 0 or SL/TP.
Walk-Forward Training:
- Trains on first `train_ratio` of data (default 70%)
- Generates signals only for remaining test period (30%)
- Eliminates look-ahead bias for realistic backtest results
"""
def __init__(self,
model_path: str = "data/regime_model.pkl",
horizon: int = 96, # 4 Days based on research
z_window: int = 24,
stop_loss: float = 0.06, # 6% to survive 2% avg MAE
take_profit: float = 0.05, # Swing target
train_ratio: float = 0.7 # Walk-forward: train on first 70%
):
super().__init__()
self.model_path = model_path
self.horizon = horizon
self.z_window = z_window
self.stop_loss = stop_loss
self.take_profit = take_profit
self.train_ratio = train_ratio
# Default Strategy Config
self.default_market_type = MarketType.PERPETUAL
self.default_leverage = 1
self.dm = DataManager()
self.model = None
self.feature_cols = None
self.train_end_idx = None # Will store the training cutoff point
def run(self, close, **kwargs):
"""
Execute the strategy logic.
We assume this strategy is run on ETH-USDT (the active asset).
We will fetch BTC-USDT internally to calculate the spread.
"""
# 1. Identify Context
# We need BTC data aligned with the incoming ETH 'close' series
start_date = close.index.min()
end_date = close.index.max()
logger.info("Fetching BTC context data...")
try:
# Load BTC data (Context) - Must match the timeframe of the backtest
# Research was done on 1h candles, so strategy should be run on 1h
df_btc = self.dm.load_data("okx", "BTC-USDT", "1h", MarketType.SPOT)
# Align BTC to ETH (close)
df_btc = df_btc.reindex(close.index, method='ffill')
btc_close = df_btc['close']
except Exception as e:
logger.error(f"Failed to load BTC context: {e}")
empty = self.create_empty_signals(close)
return empty, empty, empty, empty
# 2. Construct DataFrames for Feature Engineering
# We need volume/high/low for features, but 'run' signature primarily gives 'close'.
# kwargs might have high/low/volume if passed by Backtester.run_strategy
eth_vol = kwargs.get('volume')
if eth_vol is None:
logger.warning("Volume data missing. Feature calculation might fail.")
# Fallback or error handling
eth_vol = pd.Series(0, index=close.index)
# Construct dummy dfs for prepare_features
# We only really need Close and Volume for the current feature set
df_a = pd.DataFrame({'close': btc_close, 'volume': df_btc['volume']})
df_b = pd.DataFrame({'close': close, 'volume': eth_vol})
# 3. Load On-Chain Data (CryptoQuant)
# We use the saved CSV for training/inference
# In a live setting, this would query the API for recent data
cq_df = None
try:
cq_path = "data/cq_training_data.csv"
cq_df = pd.read_csv(cq_path, index_col='timestamp', parse_dates=True)
if cq_df.index.tz is None:
cq_df.index = cq_df.index.tz_localize('UTC')
except Exception:
logger.warning("CryptoQuant data not found. Running without on-chain features.")
# 4. Calculate Features
features = self.prepare_features(df_a, df_b, cq_df)
# 5. Walk-Forward Split
# Train on first `train_ratio` of data, test on remainder
n_samples = len(features)
train_size = int(n_samples * self.train_ratio)
train_features = features.iloc[:train_size]
test_features = features.iloc[train_size:]
train_end_date = train_features.index[-1]
test_start_date = test_features.index[0]
logger.info(
f"Walk-Forward Split: Train={len(train_features)} bars "
f"(until {train_end_date.strftime('%Y-%m-%d')}), "
f"Test={len(test_features)} bars "
f"(from {test_start_date.strftime('%Y-%m-%d')})"
)
# 6. Train Model on Training Period ONLY
if self.model is None:
logger.info("Training Regime Model on training period only...")
self.model, self.feature_cols = self.train_model(train_features)
# 7. Predict on TEST Period ONLY
# Use valid columns only
X_test = test_features[self.feature_cols].fillna(0)
X_test = X_test.replace([np.inf, -np.inf], 0)
# Predict Probabilities for test period
probs = self.model.predict_proba(X_test)[:, 1]
# 8. Generate Entry Signals (TEST period only)
# If Z > 1 (Spread High, ETH Expensive) -> Short ETH
# If Z < -1 (Spread Low, ETH Cheap) -> Long ETH
short_signal_test = (probs > 0.5) & (test_features['z_score'].values > 1.0)
long_signal_test = (probs > 0.5) & (test_features['z_score'].values < -1.0)
# Create full-length signal series (False for training period)
long_entries = pd.Series(False, index=close.index)
short_entries = pd.Series(False, index=close.index)
# Map test signals to their correct indices
test_idx = test_features.index
for i, idx in enumerate(test_idx):
if idx in close.index:
long_entries.loc[idx] = bool(long_signal_test[i])
short_entries.loc[idx] = bool(short_signal_test[i])
# 9. Generate Exits
# Exit when Z-Score crosses back through 0 (mean reversion complete)
z_reindexed = features['z_score'].reindex(close.index, fill_value=0)
# Exit Long when Z > 0, Exit Short when Z < 0
long_exits = z_reindexed > 0
short_exits = z_reindexed < 0
# Log signal counts for verification
n_long = long_entries.sum()
n_short = short_entries.sum()
logger.info(f"Generated {n_long} long signals, {n_short} short signals (test period only)")
return long_entries, long_exits, short_entries, short_exits
def prepare_features(self, df_btc, df_eth, cq_df=None):
"""Replicate research feature engineering"""
# Align
common = df_btc.index.intersection(df_eth.index)
df_a = df_btc.loc[common].copy()
df_b = df_eth.loc[common].copy()
# Spread
spread = df_b['close'] / df_a['close']
# Z-Score
rolling_mean = spread.rolling(window=self.z_window).mean()
rolling_std = spread.rolling(window=self.z_window).std()
z_score = (spread - rolling_mean) / rolling_std
# Technicals
spread_rsi = ta.momentum.RSIIndicator(spread, window=14).rsi()
spread_roc = spread.pct_change(periods=5) * 100
spread_change_1h = spread.pct_change(periods=1)
# Volume
vol_ratio = df_b['volume'] / df_a['volume']
vol_ratio_ma = vol_ratio.rolling(window=12).mean()
# Volatility
ret_a = df_a['close'].pct_change()
ret_b = df_b['close'].pct_change()
vol_a = ret_a.rolling(window=self.z_window).std()
vol_b = ret_b.rolling(window=self.z_window).std()
vol_spread_ratio = vol_b / vol_a
features = pd.DataFrame(index=spread.index)
features['spread'] = spread
features['z_score'] = z_score
features['spread_rsi'] = spread_rsi
features['spread_roc'] = spread_roc
features['spread_change_1h'] = spread_change_1h
features['vol_ratio'] = vol_ratio
features['vol_ratio_rel'] = vol_ratio / vol_ratio_ma
features['vol_diff_ratio'] = vol_spread_ratio
# CQ Merge
if cq_df is not None:
cq_aligned = cq_df.reindex(features.index, method='ffill')
if 'btc_funding' in cq_aligned.columns and 'eth_funding' in cq_aligned.columns:
cq_aligned['funding_diff'] = cq_aligned['eth_funding'] - cq_aligned['btc_funding']
if 'btc_inflow' in cq_aligned.columns and 'eth_inflow' in cq_aligned.columns:
cq_aligned['inflow_ratio'] = cq_aligned['eth_inflow'] / (cq_aligned['btc_inflow'] + 1)
features = features.join(cq_aligned)
return features.dropna()
def train_model(self, train_features):
"""
Train Random Forest on training data only.
This method receives ONLY the training subset of features,
ensuring no look-ahead bias. The model learns from historical
patterns and is then applied to unseen test data.
Args:
train_features: DataFrame containing features for training period only
"""
threshold = 0.005
horizon = self.horizon
# Define targets using ONLY training data
# For Short Spread (Z > 1): Did spread drop below target within horizon?
future_min = train_features['spread'].rolling(window=horizon).min().shift(-horizon)
target_short = train_features['spread'] * (1 - threshold)
success_short = (train_features['z_score'] > 1.0) & (future_min < target_short)
# For Long Spread (Z < -1): Did spread rise above target within horizon?
future_max = train_features['spread'].rolling(window=horizon).max().shift(-horizon)
target_long = train_features['spread'] * (1 + threshold)
success_long = (train_features['z_score'] < -1.0) & (future_max > target_long)
targets = np.select([success_short, success_long], [1, 1], default=0)
# Build model
model = RandomForestClassifier(
n_estimators=300, max_depth=5, min_samples_leaf=30,
class_weight={0: 1, 1: 3}, random_state=42
)
# Exclude non-feature columns
exclude = ['spread']
cols = [c for c in train_features.columns if c not in exclude]
# Clean features
X_train = train_features[cols].fillna(0)
X_train = X_train.replace([np.inf, -np.inf], 0)
# Remove rows with NaN targets (from rolling window at end of training period)
valid_mask = ~np.isnan(targets) & ~np.isinf(targets)
# Also check for rows where future data doesn't exist (shift created NaNs)
valid_mask = valid_mask & (future_min.notna().values) & (future_max.notna().values)
X_train_clean = X_train[valid_mask]
targets_clean = targets[valid_mask]
logger.info(f"Training on {len(X_train_clean)} valid samples (removed {len(X_train) - len(X_train_clean)} with incomplete future data)")
model.fit(X_train_clean, targets_clean)
return model, cols