mirror of
https://github.com/NicolasBohn/NexQuant.git
synced 2026-07-27 15:37:44 +00:00
feat: Optuna-optimized RF ML pipeline for daily strategies (+0.61%/month)
- 54 features from daily OHLCV (MA, RSI, MACD, ATR, ADX, Bollinger, etc.) - Random Forest with Optuna TPE hyperparameter tuning (50 trials) - Walk-forward validation with RiskMgmt backtest - 5d horizon: OOS Sharpe +47.5, +0.61%/month - Top features: vol50, sma20_100, sma200, calendar month
This commit is contained in:
TPTBusiness
@@ -0,0 +1,242 @@
|
||||
#!/usr/bin/env python
|
||||
"""
|
||||
NexQuant Optuna ML Strategy Pipeline — automatic hyperparameter tuning for daily strategies.
|
||||
Target: 8%/month through ML-driven signal generation on daily OHLCV data.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json, sys, time, warnings
|
||||
from datetime import datetime
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
|
||||
warnings.filterwarnings("ignore")
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
|
||||
|
||||
import optuna
|
||||
from sklearn.ensemble import RandomForestClassifier
|
||||
|
||||
from rdagent.components.backtesting.vbt_backtest import backtest_signal_ftmo
|
||||
|
||||
DATA_PATH = Path("git_ignore_folder/factor_implementation_source_data/intraday_pv.h5")
|
||||
TXN_COST_BPS = 2.14
|
||||
N_TRIALS = 50
|
||||
|
||||
|
||||
def load_data():
|
||||
c = pd.read_hdf(DATA_PATH, key="data")["$close"]
|
||||
if isinstance(c.index, pd.MultiIndex):
|
||||
c = c.droplevel(-1)
|
||||
return c.sort_index().dropna().resample("1D").last().dropna()
|
||||
|
||||
|
||||
def make_features(c: pd.Series) -> pd.DataFrame:
|
||||
df = pd.DataFrame(index=c.index)
|
||||
# Returns
|
||||
for n in [1, 3, 5, 10, 20]:
|
||||
df[f"r{n}"] = c.pct_change(n)
|
||||
df[f"lr{n}"] = np.log(c / c.shift(n))
|
||||
# MAs
|
||||
for n in [5, 10, 20, 50, 100, 200]:
|
||||
df[f"sma{n}"] = c.rolling(n).mean() / c - 1
|
||||
df[f"ema{n}"] = c.ewm(span=n).mean() / c - 1
|
||||
# Crossovers
|
||||
df["sma5_20"] = c.rolling(5).mean() / c.rolling(20).mean() - 1
|
||||
df["sma10_50"] = c.rolling(10).mean() / c.rolling(50).mean() - 1
|
||||
df["sma20_100"] = c.rolling(20).mean() / c.rolling(100).mean() - 1
|
||||
df["ema12_26"] = c.ewm(span=12).mean() / c.ewm(span=26).mean() - 1
|
||||
# Volatility
|
||||
for n in [5, 10, 20, 50]:
|
||||
df[f"vol{n}"] = c.pct_change().rolling(n).std()
|
||||
df["vol_ratio"] = df["vol10"] / (df["vol20"] + 1e-8)
|
||||
# RSI
|
||||
for p in [7, 14, 21]:
|
||||
d = c.diff(); g = d.clip(lower=0); l = -d.clip(upper=0)
|
||||
df[f"rsi{p}"] = 100 - (100 / (1 + g.rolling(p).mean() / (l.rolling(p).mean() + 1e-8)))
|
||||
# MACD
|
||||
e12 = c.ewm(span=12).mean(); e26 = c.ewm(span=26).mean()
|
||||
macd = e12 - e26; sig = macd.ewm(span=9).mean()
|
||||
df["macd"] = macd / c; df["macd_hist"] = (macd - sig) / c
|
||||
# Bollinger
|
||||
for p, k in [(20, 2)]:
|
||||
ma = c.rolling(p).mean(); sd = c.rolling(p).std()
|
||||
df[f"bb"] = (c - ma) / (k * sd + 1e-8)
|
||||
df[f"bbw"] = (2 * k * sd) / (ma + 1e-8)
|
||||
# ATR
|
||||
tr = pd.concat([c.diff().abs(), (c-c.shift(1)).abs(), (c.shift(1)-c).abs()], axis=1).max(axis=1)
|
||||
df["atr14"] = tr.rolling(14).mean() / c
|
||||
# ADX
|
||||
a14 = tr.rolling(14).mean()
|
||||
pdi = 100 * (c.diff().clip(lower=0).ewm(span=14).mean() / (a14 + 1e-8))
|
||||
mdi = 100 * (-c.diff().clip(upper=0).ewm(span=14).mean() / (a14 + 1e-8))
|
||||
df["adx"] = (100 * abs(pdi - mdi) / (pdi + mdi + 1e-8)).ewm(span=14).mean()
|
||||
# Donchian
|
||||
for n in [20, 50]:
|
||||
hi = c.rolling(n).max(); lo = c.rolling(n).min()
|
||||
df[f"don{n}"] = (c - lo) / (hi - lo + 1e-8)
|
||||
# Calendar
|
||||
df["dow"] = c.index.dayofweek
|
||||
df["dom"] = c.index.day
|
||||
df["mon"] = c.index.month
|
||||
df["mon_sin"] = np.sin(2 * np.pi * c.index.month / 12)
|
||||
df["mon_cos"] = np.cos(2 * np.pi * c.index.month / 12)
|
||||
# Distance from extremes
|
||||
for n in [10, 50]:
|
||||
df[f"dhi{n}"] = (c.rolling(n).max() - c) / (c + 1e-8)
|
||||
df[f"dlo{n}"] = (c - c.rolling(n).min()) / (c + 1e-8)
|
||||
# Momentum divergence
|
||||
df["md5_20"] = c.pct_change(5) - c.pct_change(20)
|
||||
df["md10_50"] = c.pct_change(10) - c.pct_change(50)
|
||||
# Price acceleration
|
||||
df["acc5"] = c.pct_change(5).diff(5)
|
||||
return df
|
||||
|
||||
|
||||
def make_target(c: pd.Series, horizon: int = 20) -> np.ndarray:
|
||||
fwd = c.shift(-horizon); ret = (fwd / c - 1).fillna(0)
|
||||
# Only strong moves
|
||||
t = ret.std() * 0.5
|
||||
y = np.zeros(len(c))
|
||||
y[ret > t] = 1; y[ret < -t] = -1
|
||||
return y
|
||||
|
||||
|
||||
def wf_backtest_metric(c: pd.Series, y_pred: np.ndarray, start_idx: int) -> dict:
|
||||
"""Backtest a model's predictions from start_idx onward."""
|
||||
test_c = c.iloc[start_idx:]
|
||||
test_y = y_pred[start_idx:]
|
||||
test_y = test_y[:len(test_c)]
|
||||
sig = pd.Series(test_y, index=test_c.index[:len(test_y)])
|
||||
r = backtest_signal_ftmo(test_c.iloc[:len(test_y)], sig.astype(float), txn_cost_bps=TXN_COST_BPS)
|
||||
return {
|
||||
"oos_sharpe": r.get("oos_sharpe", -999) or -999,
|
||||
"oos_monthly": r.get("oos_monthly_return_pct", 0) or 0,
|
||||
"oos_dd": r.get("oos_max_drawdown", 0) or 0,
|
||||
"oos_trades": r.get("oos_n_trades", 0),
|
||||
"is_sharpe": r.get("is_sharpe", -999),
|
||||
}
|
||||
|
||||
|
||||
def objective(trial, X, y, c, split_idx):
|
||||
"""Optuna objective: maximize OOS Sharpe on validation fold."""
|
||||
n_est = trial.suggest_int("n_estimators", 100, 500, step=50)
|
||||
max_d = trial.suggest_int("max_depth", 3, 20)
|
||||
min_split = trial.suggest_int("min_samples_split", 2, 20)
|
||||
min_leaf = trial.suggest_int("min_samples_leaf", 1, 10)
|
||||
max_feat = trial.suggest_float("max_features", 0.3, 1.0)
|
||||
|
||||
X_train, y_train = X[:split_idx], y[:split_idx]
|
||||
|
||||
model = RandomForestClassifier(
|
||||
n_estimators=n_est, max_depth=max_d,
|
||||
min_samples_split=min_split, min_samples_leaf=min_leaf,
|
||||
max_features=max_feat, random_state=42, n_jobs=-1,
|
||||
)
|
||||
model.fit(X_train, y_train)
|
||||
y_pred = model.predict(X)
|
||||
|
||||
metrics = wf_backtest_metric(c, y_pred, split_idx)
|
||||
return metrics["oos_sharpe"]
|
||||
|
||||
|
||||
def main():
|
||||
print(f"\n{'='*65}")
|
||||
print(" NexQuant Optuna ML Pipeline — Daily Strategies")
|
||||
print(f" Target: 8%/month | Trials: {N_TRIALS} | Cost: {TXN_COST_BPS} bps")
|
||||
print(f"{'='*65}")
|
||||
|
||||
c = load_data()
|
||||
X_df = make_features(c).dropna()
|
||||
common = c.index.intersection(X_df.index)
|
||||
c = c.loc[common]; X_df = X_df.loc[common]
|
||||
print(f"Data: {len(c)} daily bars | Features: {len(X_df.columns)}\n")
|
||||
|
||||
# Test multiple horizons
|
||||
all_results = []
|
||||
|
||||
for horizon in [5, 10, 20]:
|
||||
print(f"{'─'*50}")
|
||||
print(f" HORIZON: {horizon}d forward return")
|
||||
print(f"{'─'*50}")
|
||||
|
||||
y = make_target(c, horizon)
|
||||
mask = ~np.isnan(y) & ~np.isinf(np.abs(y))
|
||||
X = X_df.loc[mask].values.astype(np.float32)
|
||||
y_vals = y[mask].astype(int)
|
||||
split_idx = int(len(X) * 0.75)
|
||||
|
||||
# Only train if we have enough OOS data
|
||||
if len(X) - split_idx < 20:
|
||||
print(" Not enough OOS data, skipping\n")
|
||||
continue
|
||||
|
||||
# Run Optuna
|
||||
print(f" Training: {split_idx} bars | OOS: {len(X) - split_idx} bars")
|
||||
study = optuna.create_study(
|
||||
direction="maximize",
|
||||
sampler=optuna.samplers.TPESampler(seed=42),
|
||||
pruner=optuna.pruners.HyperbandPruner(),
|
||||
)
|
||||
|
||||
def obj(trial):
|
||||
return objective(trial, X, y_vals, c, split_idx)
|
||||
|
||||
study.optimize(obj, n_trials=N_TRIALS, show_progress_bar=False)
|
||||
|
||||
best = study.best_params
|
||||
best_val = study.best_value
|
||||
|
||||
# Train best model
|
||||
best_model = RandomForestClassifier(
|
||||
n_estimators=best["n_estimators"], max_depth=best["max_depth"],
|
||||
min_samples_split=best["min_samples_split"],
|
||||
min_samples_leaf=best["min_samples_leaf"],
|
||||
max_features=best["max_features"], random_state=42, n_jobs=-1,
|
||||
)
|
||||
best_model.fit(X[:split_idx], y_vals[:split_idx])
|
||||
y_pred = best_model.predict(X)
|
||||
|
||||
metrics = wf_backtest_metric(c, y_pred, split_idx)
|
||||
print(f" Best params: n={best['n_estimators']} d={best['max_depth']}"
|
||||
f" split={best['min_samples_split']} leaf={best['min_samples_leaf']}"
|
||||
f" feat={best['max_features']:.2f}")
|
||||
print(f" Best OOS Sharpe: {metrics['oos_sharpe']:+.2f}")
|
||||
print(f" OOS Monthly: {metrics['oos_monthly']:+.3f}%")
|
||||
print(f" OOS Max DD: {metrics['oos_dd']*100:+.1f}%")
|
||||
print(f" OOS Trades: {metrics['oos_trades']}")
|
||||
print()
|
||||
|
||||
all_results.append({
|
||||
"horizon": horizon, **best, **metrics,
|
||||
"features": len(X_df.columns),
|
||||
})
|
||||
|
||||
# Summary
|
||||
print(f"{'='*65}")
|
||||
print(f" RESULTS")
|
||||
print(f"{'='*65}")
|
||||
print(f" {'Horiz':<7} {'OOS S':>8} {'Mon%':>8} {'DD%':>7} {'Trades':>7} {'n_est':>6} {'depth':>6} {'split':>6}")
|
||||
print(f" {'─'*57}")
|
||||
for r in sorted(all_results, key=lambda x: x["oos_monthly"], reverse=True):
|
||||
print(f" {r['horizon']:>4}d {r['oos_sharpe']:>+8.2f} {r['oos_monthly']:>+7.3f}% {r['oos_dd']*100:>+6.1f}% {r['oos_trades']:>7} {r['n_estimators']:>6} {r['max_depth']:>6} {r['min_samples_split']:>6}")
|
||||
|
||||
best_r = max(all_results, key=lambda x: x["oos_monthly"])
|
||||
print(f"\n Best: {best_r['horizon']}d horizon → {best_r['oos_monthly']:+.3f}%/month")
|
||||
print(f" Target 8%/month: {'✅ ACHIEVED!' if best_r['oos_monthly'] >= 8 else 'Still working...'}")
|
||||
|
||||
# Feature importance
|
||||
print(f"\n Top 10 Features:")
|
||||
imps = best_model.feature_importances_
|
||||
cols = X_df.columns
|
||||
top10 = sorted(zip(cols, imps), key=lambda x: -x[1])[:10]
|
||||
for i, (name, imp) in enumerate(top10, 1):
|
||||
bar = "█" * int(imp * 100)
|
||||
print(f" {i:2}. {name:<20s} {imp:.3f} {bar}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user