AI Autoresearch Strategy Generator
- Source
- x.com
- Author
- ryan 🌊
- Date
Context
Automated strategy search still needs tests that distinguish improvement from luck. In ryan's account of a simulated prediction-market contest, coding agents repeatedly modified strategies and evaluated them. After seeing results vary across simulated price paths, the author tested candidates against several sets of paths instead of optimizing only for one.
The author reports $52.03 on the hackathon leaderboard, which reused one fixed batch of simulations, and $42.32 in the final , the median of three runs on fresh random seeds. The score multiplies executed trade size by the favorable or unfavorable gap between price and the simulator's underlying probability, averaged across simulations. Those figures describe simulated pricing advantage, not live-trading returns. The code and evaluation scripts are published.
- The winning market maker's first rule is refusal: skip the tightest-spread regime entirely, because the informed arbitrageur sweeps your quote before retail arrives. Not trading there is the profitable move.
- Order size scales with the modeled probability that an arbitrageur hits you. Quote large at extreme prices and wide spreads where that's unlikely, small near 50% where it's dangerous.
- An empty book is the best moment, not a reason to pause. When the competitor's orders get consumed, the strategy quotes at tick 1 or 99 rather than cancelling, capturing the widest edge available.
- The loop ran 8 to 20 parallel Claude Code sessions, each assigned one hypothesis or search space: read the current best strategy, build variants, evaluate, report. The human's job shrank to folding winners back in.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
“I won a hackathon by @paradigm, but I have no idea what the winning strategy was.”
“Every step of the solution was made by AI.”
“At peak, I had 20 agents running simultaneously, each sweeping different parameter spaces or testing structural changes. This massively parallelized the search — equivalent to weeks of manual experimentation compressed into hours.”
“Early on, I optimized only on seed=0 and hit +$44 locally. But when I tested on other seeds, performance varied wildly (+$34 to +$70).”
Checking sign-in…
Loading comments…