Polymarket Data for Machine Learning

The outcome column is the supervised label, and the honest label design uses only data known before the resolve second. The classic leak is temporal: a model trained on a market's full history and scored on its final minute has seen its own answer.

The archive's flag columns are the second trap: dropped one-sided rows vanish the venue's structure, and a model that only ever sees two-sided frames learns a market that does not exist.

Score on resolved markets with the clock honored: train/validation splits by time, never random, and evaluate on out-of-sample windows the model never saw. Calibration matters on a venue whose prices bunch at the bound — a 0.99-predicting model and a 0.51-predicting model are different beasts.

The honest report includes the one-sided share and the grid, so a reviewer can judge whether the model was graded on the venue's real texture.

ML on this record is forecasting structure, not predicting the reference print: backtests confirm the venue's own structural baseline (one-sided 16.9%, bound resting 42k median) and a model that beats it on held-out resolved markets has a claim — a model that overfits the final minute has only a graph.

• Feature engineering
• Research methods
• Python data guide