The neural net was stuck at 264s. Diagnostics said rare pairs were the problem. The fix wasn't a better NN — it was two completely different models and an ensemble that cut MAE to 253s.
Zone-pair median — a dictionary lookup — beats XGBoost with zero ML. Here's how I built 26 features, iterated through 4 neural net versions, and hit a ceiling that no architecture change could break.
15 models. 4 disasters. A complete experiment log of training a 7B LLM to trade stocks using supervised fine-tuning and Group Relative Policy Optimization — what worked, what broke, and why.
Every shortcut I left in the environment, the agent found and exploited. Here's how I designed observations, rewards, grading, and a neural world model for training LLM trading agents on real Indian equity data.