Why Most RL Trading Projects Fail in Production
The four failure modes that kill 90% of AI trading systems: curve fitting, unrealistic simulation, slow iteration, and missing risk models. How we designed around each one.
The Hidden Cost of Ignoring Transaction Fees in Indian Markets
Indian equity transaction costs are material for delivery trades. With daily rebalancing, turnover costs compound to a significant annual drag. Any backtest that ignores this is fiction. Here's the full breakdown.
Walk-Forward Validation: The Only Honest Way to Evaluate Trading Strategies
Why train/test splits are insufficient for financial time series, how rolling-window validation works, and why we require consistency across 5+ windows before promoting an agent.
Tuning RL Algorithms for Low-SNR Financial Data
Standard RL hyperparameters tuned on robotics or game benchmarks often fail on financial time series. We discuss the importance of domain-specific tuning for entropy targets, learning rates, and reward scaling.
From PyTorch to JAX: Why We Train on Two Backends
JAX gives us XLA-compiled speed on GPU. PyTorch gives us rapid iteration on desktop native. Here's how we maintain parity between two entirely different training stacks.