AcyLa/volatility-predictive-model

High-Volatility Asset Predictive Model

Ensemble ML pipeline (LightGBM + LSTM with attention) covering 19 of the most volatile instruments available on Yahoo Finance.

Trained on 15+ years of historical data with 194+ engineered features per asset, including Nadaraya-Watson envelope estimation and multi-window Bollinger Band systems. Assets span leveraged ETFs (3x), VIX products, crypto miners, biotech, and high-beta tech stocks with annualized volatility ranging from 31% to 185%.

---

Covered Assets

Leveraged ETFs (3x daily leverage)

Ticker	Name	Vol	Bars	Date Range
TQQQ	ProShares UltraPro QQQ (3x Nasdaq-100)	61%	4,171	2010–2026
SQQQ	ProShares UltraPro Short QQQ (-3x)	61%	4,171	2010–2026
UPRO	ProShares UltraPro S&P500 (3x)	51%	4,330	2009–2026
SPXU	ProShares UltraPro Short S&P500 (-3x)	51%	4,330	2009–2026
SOXL	Direxion Semiconductor Bull 3x	92%	4,152	2010–2026
TNA	Direxion Small Cap Bull 3x	70%	4,479	2008–2026

VIX & Volatility Products

Ticker	Name	Vol	Bars
UVXY	ProShares Ultra VIX Short-Term (1.5x)	118%	3,756
VIXY	ProShares VIX Short-Term Futures	70%	3,945

Crypto Miners (Bitcoin proxy)

Ticker	Name	Vol	Bars
MSTR	MicroStrategy (Bitcoin proxy)	76%	7,107
MARA	Marathon Digital (BTC mining)	162%	3,609
RIOT	Riot Platforms (BTC mining)	112%	2,628
CLSK	CleanSpark (BTC mining)	185%	2,467

Biotech (binary event-driven)

Ticker	Name	Vol	Bars
XBI	SPDR S&P Biotech ETF	31%	5,182

High-Beta Tech

Ticker	Name	Vol	Bars
NVDA	NVIDIA (Semiconductor/AI)	59%	6,952
AMD	Advanced Micro Devices	59%	11,717
TSLA	Tesla (High-beta EV)	57%	4,076
PLTR	Palantir (Data/AI)	70%	1,494
COIN	Coinbase (Crypto exchange)	85%	1,360
APP	AppLovin (Adtech/AI)	77%	1,359

---

Architecture

Data Download (Yahoo Finance) → Feature Engineering (194+ indicators) → Label Generation (multi-horizon) → LightGBM + LSTM Ensemble → Prediction API

Feature Categories (194+ 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 QQQ, VIXY, and GLD benchmarks (15+ features)

Model Architecture

• LightGBM — Gradient-boosted trees, 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, 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 (>+4%), SELL (<-4%), HOLD (between) — higher threshold for vol assets
• Triple-Barrier — López de Prado method with 6% barriers (wider for high-vol assets)

---

Training Results (Quick Split — 70/15/15, no data leakage)

Asset	Vol	1d Acc	5d Acc	21d Acc	Best Sharpe
SPXU	51%	56.3%	60.3%	72.4%	5.59
SQQQ	61%	57.3%	59.6%	71.3%	5.41
UPRO	51%	56.6%	59.6%	68.9%	4.98
UVXY	118%	57.0%	61.8%	69.4%	4.29
TQQQ	61%	57.5%	58.9%	62.4%	4.32
VIXY	70%	52.5%	59.6%	65.4%	2.10
NVDA	59%	53.5%	57.5%	60.8%	5.18
TNA	70%	53.5%	51.7%	61.0%	2.70
RIOT	112%	49.4%	56.5%	57.3%	3.75
SOXL	92%	55.2%	58.4%	55.8%	4.04
TSLA	57%	51.3%	48.3%	53.8%	3.64
AMD	59%	51.8%	56.7%	51.7%	4.53
CLSK	185%	50.7%	52.3%	50.0%	2.84
MSTR	76%	53.1%	53.8%	54.0%	0.90
MARA	162%	45.5%	47.9%	49.5%	0.34
XBI	31%	49.0%	48.3%	47.9%	0.76
PLTR	70%	51.1%	45.1%	41.8%	1.39
COIN	85%	43.3%	56.1%	40.2%	1.29
APP	77%	47.6%	40.2%	34.8%	-1.04

Key Findings

Most predictable (21-day horizon):

• Inverse leveraged ETFs (SPXU, SQQQ) — 72% directional accuracy, Sharpe 5.4-5.6. This reflects the well-known volatility decay of leveraged products — they trend downward over time, making direction easier to predict.
• Leveraged ETFs (UPRO, TQQQ) — 62-69% accuracy at 21d. Same effect in reverse.
• VIX products (UVXY, VIXY) — 65-69% at 21d. VIX futures are in contango most of the time, causing systematic downward drift.
• NVDA — 61% at 21d, Sharpe 5.18. Strong momentum trends in semiconductor/AI cycle.

Least predictable (near random):

• Crypto miners (MARA, CLSK) — 50% at all horizons. Extreme volatility (162-185%) makes them noise-like.
• Recent IPOs (APP, PLTR, COIN) — 35-42% at 21d. Limited history + regime changes.
• Biotech (XBI) — 48% at all horizons. Binary FDA events are fundamentally unpredictable from technicals.

Volatility decay insight: The inverse leveraged ETFs (SPXU, SQQQ) are MORE predictable than their bullish counterparts (UPRO, TQQQ) because volatility decay creates a systematic downward drift. This is a structural alpha source, not a model artifact.

---

Usage

Install

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

Predict — Single Asset

python predict.py --ticker TQQQ
# Output:
#   TQQQ — ProShares UltraPro QQQ (3x Nasdaq-100)
#   Consensus: BUY (1/3)
#   h= 1d: HOLD conf=50% pred=+0.0085
#   h= 5d: HOLD conf=50% pred=+0.0089
#   h=21d:  BUY conf=100% pred=+0.0497

Predict — All Assets

python predict.py --all

Fast Mode (LightGBM only, no LSTM)

python predict.py --all --no-lstm

Python API

from predict import AssetPredictor
predictor = AssetPredictor("TQQQ")
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 TQQQ

# All assets (LightGBM only)
python batch_train.py

# All assets (LSTM)
python batch_train_lstm.py

---

File Structure

vol_predictor/
├── config.py              # All hyperparameters (19 volatile assets, features, model params)
├── data_loader.py         # Yahoo Finance download with parquet caching
├── nadaraya_watson.py     # Vectorized NW envelope estimator
├── features.py            # 194+ feature engineering (NW, BB, trend, momentum, etc.)
├── labels.py              # Multi-horizon regression + classification + triple-barrier (6% barriers)
├── models.py              # LightGBMPredictor + LSTMPredictor + EnsemblePredictor
├── backtest.py            # Walk-forward CV with purging + quick split (70/15/15)
├── batch_train.py         # Batch LightGBM training for all assets
├── batch_train_lstm.py    # Batch LSTM training for all assets
├── predict.py             # Prediction API + CLI
├── vol_models/            # Trained models (per-ticker subdirectories)
│   ├── TQQQ/
│   │   ├── lgbm_h1.joblib    # LightGBM (1-day)
│   │   ├── lgbm_h5.joblib    # LightGBM (5-day)
│   │   ├── lgbm_h21.joblib   # LightGBM (21-day)
│   │   ├── lstm_h1.pt        # LSTM (1-day)
│   │   ├── lstm_h5.pt        # LSTM (5-day)
│   │   ├── lstm_h21.pt       # LSTM (21-day)
│   │   ├── weights_h*.joblib # Ensemble weights
│   │   ├── feature_cols.joblib
│   │   └── meta.json
│   ├── SQQQ/ ...
│   └── ... (19 assets total)
└── vol_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: 5-15 years depending on asset (1980-2026 for AMD, 2010-2026 for TQQQ)
• Caching: Parquet files in vol_data_cache/

Key Design Decisions

1. No target leakage — Target columns (Target_*) excluded from features. Verified via feature importance check.
2. Three-way split — 70% train / 15% validation / 15% test (chronological, no overlap)
3. Wider classification threshold — 4% (vs 2% for normal assets) to account for higher volatility
4. Wider triple-barrier — 6% barriers (vs 3%) appropriate for assets with 50-185% annual volatility
5. Cross-asset features — Correlation with QQQ (tech), VIXY (volatility), and GLD (gold) as market regime indicators
6. Nadaraya-Watson vectorized — Uses sliding_window_view for O(n) computation

Honest Assessment

• Leveraged/inverse ETFs and VIX products show strong predictability at 21-day horizons (62-72%) — this is structural alpha from volatility decay, not a magical model
• High-beta tech (NVDA, AMD, TSLA) shows moderate predictability (53-61% at 21d) — momentum effects
• Crypto miners and recent IPOs are near-random (35-50%) — too volatile/noisy for technical analysis alone
• Biotech is fundamentally event-driven — technicals cannot predict FDA approvals
• 1-day predictions are 50-57% for most assets — consistent with efficient markets at short horizons
• These models provide signal, not certainty. High-volatility assets can gap 5-10% in a single session. Always use strict risk management (stop-losses, position sizing, max drawdown limits)
• For production: retrain monthly, use walk-forward backtesting, add options implied volatility and sentiment features

Dependencies

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

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