Viral napkin trick: AI replicated a quant edge

A trader used Anthropic’s Claude to rebuild a prediction-market strategy that people on X described as a Jane Street-style “napkin model,” then claimed an 81% win rate and about $16,000 of profit and loss with cheap infrastructure instead of a full quant stack. The thread spread because the inputs were plain English prompts plus public data, not a research team with proprietary pipes. (x.com) Prediction markets are markets where a contract pays if an event happens, so a 63-cent price usually means the market is implying about a 63% chance. Traders try to buy when the true odds are higher than the market price and sell when the market price is too rich. (github.com) That sounds simple, but the hard part is noise. Jane Street’s own Kaggle competition page says profitable trading is difficult because signals are weak, features are correlated, and even a good decision can look like luck in the short run. (kaggle.com) Jane Street used that competition to show the shape of the problem. Kaggle says the contest ran from November 23, 2020 to August 23, 2021, and Jane Street said its real systems trade thousands of products across 200 venues while the contest only “far oversimplifies” the actual work. (kaggle.com) The viral part is not that a language model invented a magic formula from nothing. The viral part is that Claude could reportedly turn a rough idea into code, data cleaning, backtests, and iteration fast enough that one person could do work that used to need several specialists. (x.com, anthropic.com) That speed matters because the data layer is no longer the main bottleneck. Jon Becker’s open-source prediction-market-analysis project says it includes pre-collected Polymarket and Kalshi datasets, API indexers, and analysis scripts, and the compressed dataset alone is 36 gibibytes. (github.com) A few years ago, getting that much market history into a usable format was half the job. The repository says it already handles market metadata, trade history from application programming interfaces and blockchain sources, Parquet storage, and resumable collection, which means a model can start with shelves already stocked. (github.com) Anthropic has been pushing Claude in exactly this direction. Its May 22, 2025 Claude 4 launch said the models can use tools, work across files, and handle long-running coding tasks, which is the kind of workflow you need for research notebooks, feature engineering, and backtest loops. (anthropic.com) The important caveat is that a posted win rate is not the same thing as a durable edge. Kaggle’s Jane Street page warns that market volatility makes profitability uncertain and says it is hard to separate good luck from good decisions, which is even more true when a result comes from one viral backtest instead of audited live records. (kaggle.com) Still, the cost curve has clearly moved. When public repositories provide the market tape and a coding model can write the parser, test the hypothesis, and revise the strategy in one sitting, the scarce thing is less infrastructure and more judgment about which tiny pricing mistakes are real. (github.com, anthropic.com) That is why a napkin sketch turned into a story. The old moat was a building full of quants, data engineers, and expensive internal tools; the new version can start with one person, one model, and a GitHub repo that was updated four days ago. (x.com, github.com)