Nifty RL Trading Results

Reinforcement Learning trading strategy using PPO

⚠️ Research Only — Not Production
This is a short / bounded training run (80,000 timesteps on CPU) for research purposes only. It is not a production model and not intended for live trading. The agent still shows a negative absolute return (-6.40%), though it outperforms buy-and-hold by +5.29 percentage points.

Adapted from: ZiadFrancis/Reinforcement_Trading_Part_2 (PPO + MultiDiscrete bracket environment)
Changes for Nifty: Indian equity index data, IST timezone, adjusted trading costs, calendar-based splits

Key Results Summary

PPO Agent

-6.40%

Total Return

554 trades • 42.78% win rate
Max DD: -13.27% • PF: 0.948
Final equity: 9,360

Buy & Hold

-11.69%

Total Return

1 trade (hold entire period)
Max DD: -15.72%
Final equity: 8,831

Excess Return

+5.29 pp

vs Buy & Hold

Agent outperforms baseline by
5.29 percentage points
(still negative absolute return)

Equity Curve — Test Period

📁 Result files: equity_curve.csv • trades.csv • metrics.json • equity.png • 📘 Plain-English explainer • 📈 v2 results (2M steps + walk-forward) • 📐 Intraday setups backtest · 📈 Swing / options setups · 🔴 Live Nifty bias dashboard · 🧭 Rule finder (opening 30-min move >150 & more)

Test Metrics (2025-07-10 to 2026-10-01)

Metric	PPO Agent	Buy & Hold
Total Return	-6.40%	-11.69%
Max Drawdown	-13.27%	-15.72%
Final Equity	9,360	8,831
Number of Trades	554	1
Win Rate	42.78%	—
Profit Factor	0.948	—
Excess vs B&H	+5.29 pp	—

Data Splits

Split	Start Date	End Date
Train	2022-02-22	2024-12-19
Validation	2025-01-10	2025-06-19
Test	2025-07-10	2026-10-01

Data: Nifty 50 1-minute bars (440,167 bars total from 2022-01-03 to 2026-10-01 IST)
Source: s3://analyticsdata123/options-ml-platform/data/timeframes/1m/nifty50.parquet (region ap-south-2)

Method Summary

Algorithm

RL Algorithm: PPO (Proximal Policy Optimization) via stable-baselines3
Training timesteps: 80,000 (short run, CPU)
Seed: 42
Initial equity: 10,000

Decision & Execution

Decision timeframe: H1 (hourly bars resampled from 1-minute data)
Execution timeframe: 1-minute (for intrabar TP/SL fills)
Action space: Gymnasium MultiDiscrete [direction, SL ATR bucket, TP R-multiple bucket]
- Stop-loss ATR multipliers: [1.0, 1.5, 2.0]
- Take-profit R-multiple buckets: [1.0, 1.5, 2.0, 3.0]
Fill order: Stop-loss first (SL-first), then take-profit

Features & Reward

Observations: 25 causal ATR-normalized features + 6 position features
Reward: Risk-normalized realized PnL + light MTM shaping − holding penalty

Trading Costs (Nifty index points)

Spread: 1.0
Slippage: 0.5
Commission: 2.0 per trade
Note: These are rough estimates for liquid index futures

Limitations & Next Steps

Limitations

Short training run (80k timesteps) — not sufficient for production validation
Single chronological split — no walk-forward or rolling window validation
CPU training only (slow convergence)
Rough cost estimates — actual costs may vary by broker and instrument
Negative absolute returns — agent loses money despite outperforming buy-and-hold
Test period may not be representative of other market regimes

Potential Improvements

Extend training (500k–2M timesteps on GPU)
Implement walk-forward validation across multiple periods
Tune hyperparameters (learning rate, batch size, network architecture)
Add regime awareness (trend/range detection)
Incorporate additional features (order flow, volatility regime, market breadth)
Test with actual broker costs and constraints
Validate on out-of-sample data from different years/regimes

Files & Code

This project is part of the s1a2t3y4/trading_algo monorepo under aiprojects/rl-trading-nifty/.

📂 Project structure:
README.md — Documentation
run_nifty_ppo.py — Main training & evaluation runner
nifty_data.py — Parquet loader for Nifty 1m bars
requirements.txt — Python dependencies
models/ppo_H1_sl1-1.5-2_tp1-1.5-2-3_80k_seed42.zip — Trained model checkpoint
results/ — Test results (metrics, equity, trades, charts)

Upstream source: github.com/ZiadFrancis/Reinforcement_Trading_Part_2

Legal & Disclaimers

This repository is for education and research purposes only. Nothing here is investment advice. Past backtest results do not guarantee future performance. The strategy shows a net loss on the test period and should not be used for live trading without extensive further validation, risk management, and paper trading.

Note: The matching VecNormalize pickle is not included in the repo. To reload the policy for inference, you will need to retrain or provide the vecnorm file separately.

Generated: Oct 4, 2026 • Test period: Jul 10, 2025 – Oct 1, 2026