Self-Improving Market Bot

What it is

A scaffolded recursive-improvement harness for a sector-rotation strategy: a Claude Code agent edits only train.py each iteration, and a deterministic harness keeps the change (git commit) or rolls it back. Acceptance needs a higher median CAGR across combinatorial purged cross-validation paths, beating the benchmark in at least half the time blocks, and a Probability of Backtest Overfitting under a ceiling.

The problem

Automated strategy search overfits unless the judge is deterministic and the agent cannot touch it.

How it works

1. The agent's only editable file is train.py; program.md steers it, and validation, CPCV and backtest code are off-limits
2. CPCV with purge and embargo produces a distribution of out-of-sample backtest paths
3. PBO is tracked across every strategy tried, and a sealed lockbox set is reserved for rare final checks
4. Run modes for the Claude agent, random parameter mutation, and manual edits; scaffold smoke-tested on synthetic data, real data not yet downloaded

Stack

Disciplines

For your business

Agent loops with deterministic acceptance gates, so AI proposals ship only when they pass independent validation.