Where AI actually helps in systematic strategy work.

AI is a strong candidate-generator and a sharp red-team partner for systematic strategy work. It is a weak portfolio-construction agent. The operator still owns the live-vs-backtest variance call, the kill-switch call, and the portfolio-weights call.

Where should a retail systematic trader plug AI into the build process?

This is the operational map I'd lay out for any retail systematic trader trying to work out where to plug AI into their build process.

Where does AI genuinely help?

Candidate generation. Hand a strong model your feature inventory, your risk budget, and instrument universe with a tight prompt. It produces 5 to 15 ideas in an afternoon. Two or three will be worth backtesting. Better hit rate than reading old papers because the model synthesises across more material than a human can hold in head.

I use Claude Cowork to run an automated scheduled task every morning that independently scrapes different sources for highly cited trading frameworks, cross-references them against ideas from respected quantitative authors, and then synthesises the viable candidates.

Without an inch of manual intervention, the workflow takes those concepts, applies my own custom-built code libraries to fragment the entries into a multi-position split, hooks in my volatility filters, and drops the compiled file straight into my active strategy funnel directory. This is the real power of AI.

Most of the code it auto-generates is not market-ready, and a large portion of the strategies fail initial screening. The edge is not in any single output.

The edge is the compounding speed of the process itself.

Risk-rule red-teaming. Hand it your sizing logic, stops, drawdown circuit-breakers and ask for failure modes. Output is sharper than most prop-desk risk reviews because the model lacks the "I built this so it must be defensible" bias.

Code generation for backtest infrastructure. Walk-forward scaffolds, Monte Carlo loops, parameter sweeps, custom reporting. 80-90% first-pass, debug the edge cases. A weekend's plumbing becomes an evening.

Surfacing factor exposures. Hand the model your monthly returns and a factor list. Ask for a regression specification and interpretation. Useful first pass for catching hidden exposures.

Where does AI break down?

Portfolio construction across multiple strategies. The model cannot see your live book or real-time correlation matrix during the next vol expansion. It cannot tell you which strategies are correlated through a single hidden factor that only shows up under stress. Portfolio construction is where the operator still owns the call, every time.

Strategy selection from a candidate pool. The model generates; the operator picks. Selection requires judgement about regime forward, capacity, broker-side execution, and existing book.

Live-vs-backtest variance interpretation. The model can suggest hypotheses. It cannot decide whether deviation is signal or noise.

What is the three-check framework for trusting AI output?

Stat-source check. Find the source for any specific stat.

Instrument-specificity check. Strip the generality.

Falsifiability check. If you can falsify, do so.

For the XAQP build, I use Claude across all four "where it helps" categories. I do not use it for portfolio construction, strategy selection, or live variance calls.