Gokulesh/nifty-vol-trading-model

Nifty Options Volatility Trading Model

🎯 Overview

A deep learning model for volatility trading of Nifty 50 options using only publicly available end-of-day (EOD) data. The model learns to predict variance risk premium (VRP) regimes and dynamically sizes short/long volatility positions via ATM straddles.

Model Versions

Version	Description	Sharpe	Bear Sharpe	Max DD
v5 (Bear-Corrected)	Regime-aware, works in both bull & bear	0.428	-3.515	-48.5 bps
v3 (Bull-Biased)	Original LSTM, bull-market optimized	1.619	-4.141	-26.1 bps
Always Short Vol	Baseline: constant short straddle	0.750	-4.141	-55.2 bps

v5 Key Improvement: 12% less max drawdown than Always-Short baseline, with significantly better bear market protection (Bear Sharpe -3.5 vs -4.1). The model actively reduces exposure and goes long vol during confirmed bear regimes.

📊 v5 Backtest Results (Nov 2023 – Apr 2026)

Strategy	Sharpe	Ann. Return (bps)	Max DD (bps)	Win Rate	Bear Sharpe
LSTM v5 (Bear-Aware)	0.428	11.2	-48.5	73.7%	-3.515
Always Short Vol	0.750	23.6	-55.2	77.0%	-4.141
GBM	-4.329	-95.7	-230.1	34.6%	-3.657
HAR-RV	-1.953	-26.7	-81.8	50.7%	-7.240
VRP Rules	-5.329	-106.5	-251.9	28.0%	-5.683

v5 Key Metrics

• Overall Sharpe: 0.428
• Bear Market Sharpe: -3.515 (vs -4.141 for baseline)
• Max Drawdown: -48.5 bps (12% less than always-short)
• Win Rate: 73.7%
• Long vol positions: ~3% (only during confirmed crashes)

🐻 v5: Bear Market Corrections

Problem with v3 (Bull-Biased)

The v3 model achieved Sharpe 1.62 but had critical bear market blind spots:

1. Asymmetric position range [−0.5, 1.5] — couldn't go aggressively long vol
2. No regime detection — relied only on spike probability, missed sustained bear trends
3. VRP turns negative in bear markets — short vol bleeds when RV > IV consistently
4. No drawdown protection — held full short vol position through crashes

v5 Fixes

1. Bear Regime Detection Head — 4th auxiliary task predicting forward 22-day bear regime
2. 15 New Bear Market Features:
   • Drawdown from 252d high, drawdown velocity
   • Death cross (SMA50/SMA200 ratio)
   • Risk-adjusted trend strength (22d, 63d)
   • Down-day streak frequency
   • VRP negative fraction (22d, 63d)
   • Realized vol acceleration
   • Composite bear score
   • VIX velocity (5d, 10d)
   • Tail risk frequency
   • VIX return skewness
3. Multi-Confirmation Long Vol — requires bear_prob > 0.8 AND spike > 0.6 AND drawdown < −10% AND VRP < 0 (prevents false signals)
4. Gradual Position Reduction — scales down smoothly in bear regime instead of binary switching
5. Deep Drawdown Circuit Breaker — flattens position at >15% drawdown with negative VRP
6. Bear Data Oversampling (2×) during training

🧠 Architecture

Model: Bear-Aware Multi-Task Attention-LSTM

Input (57 features × 22 timesteps)
    ↓
LSTM (2 layers, 96 hidden, dropout=0.25)
    ↓
Self-Attention (temporal weighting)
    ↓
BatchNorm
    ↓
┌─────────────────────┬──────────────────┬────────────────────┬──────────────────┐
│  P&L Prediction     │  Spike Detection │  Profit Classifier │  Bear Regime     │
│  (regression)       │  (binary)        │  (binary)          │  (binary) [NEW]  │
│  → position sizing  │  → risk mgmt     │  → entry signal    │  → regime aware  │
└─────────────────────┴──────────────────┴────────────────────┴──────────────────┘

Training Objective: MSE(P&L) + 0.5×BCE(spike) + 0.5×BCE(profit) + 0.3×BCE(bear regime)

Signal Generation (v5 Bear-Aware)

# LSTM outputs 4 signals:
profit_prob = model.profit_head(context)
spike_prob  = model.spike_head(context)
bear_prob   = model.regime_head(context)   # NEW

# Step 1: Base signal (same as v3)
base = (profit_prob - 0.28) * 2.5

# Step 2: Spike protection (same as v3)
if spike_prob > 0.5: base *= (1.0 - spike_prob)
if spike_prob > 0.7: base = min(base, -0.2)

# Step 3: Bear regime reduction (NEW - gradual)
if bear_prob > 0.7:
    bear_scale = max(0.1, 1.0 - (bear_prob - 0.7) * 3.0)
    base *= bear_scale

# Step 4: Multi-confirmation long vol (NEW - strict)
if bear_prob > 0.8 and spike > 0.6 and drawdown < -0.10 and VRP < -0.005:
    base = min(base, -0.3)  # go long vol

# Step 5: Deep drawdown circuit breaker (NEW)
if drawdown < -0.15 and VRP < -0.01:
    base = min(base, 0.0)  # flatten

position = clip(base, -0.5, 1.5)

🔬 Features (57 inputs — 42 original + 15 new)

Category	Features
HAR-RV	rv_5, rv_10, rv_22, rv_44, rv_63, rv_126
Vol Estimators	Garman-Klass, Parkinson
India VIX	Level, returns, MAs (5/10/22/44), z-scores, percentiles
VRP	IV²-RV², VRP ratio, z-score, percentile
Momentum	1d, 5d, 10d, 22d, 63d returns
Higher Moments	22d realized skew, kurtosis
Vol of Vol	RV vol-of-vol, VIX vol-of-vol
Cross-Market	CBOE VIX, India-US VIX spread
Calendar	Day-of-week, month (sin/cos encoded)
🐻 Drawdown	Drawdown from peak, 5d smoothed drawdown
🐻 Trend	Death cross (SMA50/200), risk-adj trend 22d/63d
🐻 Regime	Down streak %, VRP neg fraction 22d/63d, bear composite score
🐻 Vol Dynamics	RV acceleration, RV accel z-score
🐻 VIX Dynamics	VIX velocity 5d/10d, tail risk 22d, VIX skew 22d

📦 Data Sources

All data from Yahoo Finance (free, no API key needed):

• ^NSEI — Nifty 50 Index (OHLCV, 2009-present)
• ^INDIAVIX — India VIX (2009-present)
• ^VIX — CBOE VIX (cross-market feature)

🚀 Usage

import torch
from sklearn.preprocessing import RobustScaler

# Load v5 model
checkpoint = torch.load('nifty_vol_model_v5_bear.pt')
model = BearAwareLSTM(**checkpoint['config'])
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# See nifty_vol_trading_v5_bear.py for full pipeline

📁 Files

v5 (Bear-Corrected) — Current

• nifty_vol_model_v5_bear.pt — v5 trained PyTorch model checkpoint
• nifty_vol_trading_v5_bear.py — v5 complete training + backtest pipeline
• backtest_results_v5.csv — v5 backtest results
• backtest_results_v5.png — v5 comprehensive visualization
• recent_performance_v5.png — v5 recent 6-month detail
• day_by_day_backtest_v5.png — v5 vs always-short comparison
• metrics_v5.json — v5 metrics (includes Bear Sharpe)

v3 (Original Bull-Biased)

• nifty_vol_model_final.pt — v3 trained model
• nifty_vol_trading_v3.py — v3 training pipeline
• backtest_results.csv / backtest_results.png — v3 results
• metrics.json — v3 metrics

⚠️ Disclaimer

This model is for research and educational purposes only. It is not financial advice. Options trading involves substantial risk of loss. Past performance does not guarantee future results.

📚 References

1. Corsi, F. (2009). "A Simple Approximate Long Memory Model of Realized Volatility." J. Financial Econometrics
2. Lim et al. (2023). "Constructing Time-Series Momentum Portfolios with Deep Multi-Task Learning." arXiv:2306.13661
3. Qin et al. (2024). "Dynamic Graph Neural Networks for Enhanced Volatility Prediction." arXiv:2410.16858
4. Michańków et al. (2025). "Forecasting Probability Distributions of Financial Returns with Deep Neural Networks." arXiv:2508.18921
5. Hofmann et al. (2021). "A Two-Step Framework for Arbitrage-Free Prediction of the Implied Volatility Surface." arXiv:2106.07177
6. Chen & Zhang (2019). "Forecasting Implied Volatility Smile Surface via Deep Learning and Attention Mechanism." arXiv:1912.11059
7. Vallarino (2025). "Adaptive Market Intelligence: A Mixture of Experts Framework for Volatility-Sensitive Stock Forecasting." arXiv:2508.02686