account-aware-training
Installation
SKILL.md
Account-Aware RL Training (v2.4)
Experiment Overview
| Item | Details |
|---|---|
| Date | 2024-12-26 |
| Goal | Make RL model learn from account state (P&L, win rate, drawdown) |
| Environment | vectorized_env.py, inference_obs_builder.py, training notebook |
| Status | Success |
Context
Prior to v2.4, the RL model was "blind" to account performance. It received:
- 53 features: price action, technicals, regime probabilities, calendar effects
- No information about cumulative P&L, win rate, or drawdown
Problem: The model could generate signals that were individually good but led to excessive drawdowns at the account level. It had no incentive to trade conservatively after losses.
Solution: Add 3 account-level features + drawdown penalty in rewards.