multi-asset-predictive-model

Multi-Asset Predictive Model

Ensemble ML pipeline (LightGBM + LSTM with attention) covering 14 ETFs, commodities, and international market assets.

Trained on 15+ years of Yahoo Finance historical data with 202+ engineered features per asset, including Nadaraya-Watson envelope estimation and multi-window Bollinger Band systems.

---

Covered Assets

US ETFs & Commodities

Non-US Market ETFs

---

Architecture

Data Download (Yahoo Finance) → Feature Engineering (202+ indicators) → Label Generation (multi-horizon) → LightGBM + LSTM Ensemble → Prediction API

Feature Categories (202+ per asset)

1. Nadaraya-Watson Envelope — Vectorized kernel regression for trend smoothing + overbought/oversold envelopes (10 features)
2. Bollinger Bands (multi-window) — 3 windows (20/50/100) × 3 std devs (2.0/2.5/3.0) = 45 BB features, including squeeze detection and band-walk signals
3. Trend — EMA (9/21/50/100/200), SMA (20/50/100/200), MACD, ADX, Ichimoku Cloud (30+ features)
4. Momentum — RSI (7/14/21), Stochastic, Williams %R, ROC, CCI, Awesome Oscillator (20+ features)
5. Volatility — ATR (14/21/50), Keltner Channel, Donchian Channel, Historical Volatility (20+ features)
6. Volume — OBV, MFI, CMF, VWAP (10/20/50), Volume Ratio (10+ features)
7. Price Structure — Multi-horizon returns, candlestick patterns, drawdown, HH/LL detection (40+ features)
8. Calendar — Day of week, month, quarter-end effects (10 features)
9. Cross-Asset Correlation — Beta and rolling correlation with SPY and GLD benchmarks (15+ features)

Model Architecture

• LightGBM — Gradient-boosted trees for tabular features. 300 estimators, 127 leaves, feature_fraction=0.7
• LSTM with Attention — 2-layer LSTM (hidden=128) with attention pooling, LayerNorm, GELU head. Sequence length=60
• Ensemble — Weighted average of LightGBM + LSTM, weights auto-tuned on validation set

Prediction Horizons

• 1-day (next session)
• 5-day (1 week)
• 21-day (1 month)

Labeling

• Regression — Future return: close[t+h] / close[t] - 1
• Classification — Directional: BUY (>+2%), SELL (<-2%), HOLD (between)
• Triple-Barrier — López de Prado method with upper/lower barriers and time stop

---

Training Results (Quick Split — 70/15/15)

Key findings:

• US equity ETFs (SPY, QQQ) show the strongest predictability at 21-day horizon (69-72% directional accuracy, Sharpe 6-7)
• Developed market ETFs (EFA, EWU, EWJ, EWG) also show strong 21-day signals (52-69%)
• Emerging markets (EEM, EWZ) show moderate predictability at 5-21 day horizons
• Commodities (GLD, SLV, USO) and India/China (INDA, KWEB) are harder to predict — closer to 50% (random) at most horizons
• 1-day predictions are near 50% for most assets — consistent with efficient market hypothesis at short horizons

---

Usage

Install

pip install yfinance pandas numpy scikit-learn lightgbm ta torch pyarrow joblib scipy huggingface_hub

Predict — Single Asset

python predict.py --ticker QQQ
# Output:
#   QQQ — Invesco QQQ (Nasdaq 100)
#   Consensus: BUY (1/3)
#   h= 1d: HOLD conf=50% pred=-0.0077
#   h= 5d: HOLD conf=50% pred=+0.0042
#   h=21d:  BUY conf=65% pred=+0.0130

Predict — All Assets

python predict.py --all
# Output:
#   QQQ: BUY (1/3)
#   SPY: HOLD (3/3)
#   GLD: HOLD (3/3)
#   ...

Fast Mode (LightGBM only, no LSTM)

python predict.py --ticker SPY --no-lstm

Python API

from predict import AssetPredictor
predictor = AssetPredictor("QQQ")
results = predictor.get_multi_horizon()
print(results['consensus'])  # "BUY (2/3)"
for h in [1, 5, 21]:
    print(f"{h}d: {results[h]['signal']} conf={results[h]['confidence']:.0%}")

Retrain Models

# Single asset
python train.py --ticker QQQ

# All assets
python train.py --all

# With walk-forward backtest
python train.py --all --backtest

# LightGBM only (faster)
python train.py --all --no-lstm

---

File Structure

multi_asset_predictor/
├── config.py              # All hyperparameters (assets, features, model params)
├── data_loader.py         # Yahoo Finance download with parquet caching
├── nadaraya_watson.py     # Vectorized NW envelope estimator
├── features.py            # 202+ feature engineering (NW, BB, trend, momentum, etc.)
├── labels.py              # Multi-horizon regression + classification + triple-barrier
├── models.py              # LightGBMPredictor + LSTMPredictor + EnsemblePredictor
├── backtest.py            # Walk-forward CV with purging (López de Prado)
├── train.py               # CLI training pipeline
├── batch_train.py         # Batch LightGBM training for all assets
├── batch_train_lstm.py    # Batch LSTM training for all assets
├── predict.py             # Prediction API + CLI
├── asset_models/          # Trained models (per-ticker subdirectories)
│   ├── QQQ/
│   │   ├── lgbm_h1.joblib    # LightGBM model (1-day horizon)
│   │   ├── lgbm_h5.joblib    # LightGBM model (5-day horizon)
│   │   ├── lgbm_h21.joblib   # LightGBM model (21-day horizon)
│   │   ├── lstm_h1.pt        # LSTM model (1-day horizon)
│   │   ├── lstm_h5.pt        # LSTM model (5-day horizon)
│   │   ├── lstm_h21.pt       # LSTM model (21-day horizon)
│   │   ├── weights_h*.joblib # Ensemble weights per horizon
│   │   ├── feature_cols.joblib  # Feature column names
│   │   └── meta.json         # Asset metadata
│   ├── SPY/ ...
│   └── ... (14 assets total)
└── asset_results/         # Evaluation metrics (JSON per asset)

---

Data Provenance

• Source: Yahoo Finance (yfinance API)
• Data: OHLCV (Open/High/Low/Close/Volume), auto-adjusted
• Interval: Daily (1d)
• History: 15-30 years depending on asset (1993-2026 for SPY, 1999-2026 for QQQ)
• Caching: Parquet files in asset_data_cache/

Key Design Decisions

1. No target leakage — Target columns (Target_*) are excluded from features. Verified via feature importance check.
2. Three-way split — 70% train / 15% validation / 15% test (chronological, no overlap)
3. Walk-forward ready — Full walk-forward with purging available via --backtest flag
4. Cross-asset features — Beta and rolling correlation with SPY and GLD benchmarks
5. Nadaraya-Watson vectorized — Uses sliding_window_view for O(n) computation (vs O(n²) naive)
6. Multi-window Bollinger — 9 Bollinger Band variants per asset + squeeze + band-walk signals

Honest Assessment

• 50-55% directional accuracy at 1-day horizon is statistically meaningful (markets are efficient at short horizons)
• 55-72% at 21-day horizon for trending assets (SPY, QQQ, EFA, EWU, EWJ) represents a genuine edge
• Commodities and some emerging markets are harder to predict — closer to random
• These models provide signal, not certainty. Always use risk management (stop-losses, position sizing)
• For production: retrain monthly, use walk-forward backtesting, add sentiment/on-chain features

Dependencies

yfinance pandas numpy scikit-learn lightgbm ta torch pyarrow joblib scipy huggingface_hub

<!-- ml-intern-provenance -->

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern