Unlike traditional genetic algorithms that optimize parameter vectors, Cambr evolves entire strategy codebases. Code is the genome. An AI agent “tunes” a strategy by writing different code, not by tweaking numbers. This means the search space includes entirely novel approaches — not just variations of a predefined template.
The name — Competitive Algorithmic Mutation Bot Reactor — is a nod to the Cambrian explosion, the moment in evolutionary history when biological diversity detonated. Cambr applies the same principle to trading strategies.
Why
Part curiosity, part stress-test. Algorithmic trading is an interesting domain because the feedback loop is merciless — overfitting is easy, edge is fragile, and the market doesn’t care about elegant code. I wanted to see how far generative AI could push the envelope when given a well-defined contract and a fitness function, and what breaks first.
How it works
The evolution loop:
- Load a population of strategy modules
- Evaluate fitness using walk-forward backtesting — in-sample/out-of-sample splits to detect overfitting
- Select top performers based on a composite risk-adjusted score
- Evolve — AI agents generate children via mutation (tweak one parent), crossover (combine two parents), or genesis (invent from scratch)
- Validate and cull — children must compile, pass contract checks, and clear minimum viability thresholds
- Repeat
Strategies implement a minimal contract: pure functions from market data to entry/exit signals. No internal state, no side effects, deterministic. This keeps the search space tractable for AI generation while making overfitting easy to detect.
Status
The full pipeline works end-to-end: population initialization, fitness evaluation, selection, and multi-strategy evolution — all automated. Currently preparing for paper trading to validate strategy performance against live market data.
Cambr is a private project. The competitive nature of algorithmic trading means publishing the full system would undermine its edge. That said, the framework contains components — the backtest engine, strategy contract, fitness evaluation — that aren’t competitive on their own. I plan to open-source those pieces selectively once the boundaries are well-defined.