Experiment Objective

Test whether scaled training length and tuned hyperparameters for low-SNR financial data produce generalizable policies on NSE daily data that clear our promotion gates.

Setup (High-Level)

• Universe: NIFTY 50 stocks by liquidity
• Granularity: Daily bars
• Algorithm: Actor-critic with dual value networks
• Backend: JAX with XLA compilation
• Training length: 5M steps
• Parallel environments: 64
• Episode length: 252 trading days
• Action space: Discrete multi-slot portfolio

Key Results

Summary

• Win rate 65%: Above threshold. The agent is reliably profitable across test episodes.
• Test Sharpe ~0.5: Positive risk-adjusted return on held-out data.
• Training speed 9700+ steps/sec: JAX XLA optimization delivered efficient throughput.

Known Limitations

• Val/test gap persists: Significant gap between validation and test performance.
• Test Sharpe below target: While positive, this is not yet competitive with sophisticated benchmarks on risk-adjusted terms.
• Single seed: This run used one random seed. Multiple seeds required for statistical confidence.
• No walk-forward yet: These results are from a single train/val/test split. Walk-forward validation may reveal additional gaps.

Next Steps

This run clears the training gate but does not yet justify live deployment. Our next experiments focus on:

• Walk-forward validation: Multi-window rolling test to prove generalization
• Multi-seed runs: 5 random seeds for statistical confidence
• Adversarial stress tests: Flash crash, volatility spike, trend reversal
• Longer training: Test whether 10M steps provides further improvement