Nifty 50 RL trader: v2 results

PPO, large_batch (lr 0.0003, batch 256, net [256, 256]), 2,000,000 steps, 8 parallel simulators, NVIDIA GeForce RTX 5090 Laptop GPU. Date-aware Indian futures costs (STT and lot size by date), 27 features. Run on 5 Oct 2026, 10:49 to 11:16 IST.

80k baseline results · Plain-English explainer

Bottom line: no trading edge. On the sealed test the agent lost 19.11% while Nifty buy-and-hold lost 11.69%, and only 1 of 4 walk-forward folds made money.

Sealed test: 2025-07-10 to 2026-10-01

Model	Return	Max DD	Trades	Win rate	Profit factor	Sharpe	Excess vs B&H
Best-val checkpoint (official)	−19.11%	−21.39%	567	41.45%	0.80	−1.82	−7.43 pp
Last-step model (2M)	−22.77%	−24.58%	952	40.76%	0.87	−1.40	−11.08 pp
Buy-and-hold Nifty	−11.69%	−15.72%

Equity chart, full 2M run (in-sample and out-of-sample)

Walk-forward: 4 folds × 200k steps

Each fold trains, validates and tests on its own consecutive block of history. Test scores use that fold's best-validation checkpoint.

Fold	Test window	Val return	Test return	Test max DD	Trades	B&H return	Excess (pp)	Equity
1	2023-01-23 to 2023-04-20	+2.39%	+1.12%	−3.69%	85	−2.73%	+3.85	chart
2	2024-03-19 to 2024-06-14	+4.77%	−0.43%	−3.11%	93	+7.13%	−7.56	chart
3	2025-05-19 to 2025-08-06	+6.91%	−4.04%	−5.12%	81	−1.73%	−2.32	chart
4	2026-07-13 to 2026-10-01	+4.55%	−4.27%	−6.59%	70	−7.39%	+3.12	chart

Files

full_training_2000k_results.json · walk_forward_results.json · full_2000k_equity.html · fold 1 · fold 2 · fold 3 · fold 4

All numbers are copied from the result JSONs written by the run. Old-cost runs (the stopped ~1.46M-step run and the 4k smoke tests) are not shown and are not results.