Polymarket Feature Engineering

Mid-based aggregates leak the venue's one-sidedness into the feature; flag-dropped curves fake liquidity; and minute-bucketed features destroy the reprint timing that carries the real signal on this record.

The honest engineering test: a feature that uses anything after the resolve second is a label, not a feature, and a feature that ignores the one-sided flag is a miss at 16.9% of frames.

Walk the 250ms rows per market, compute the feature series with a look-back window that ends before the resolve second, and store the flag columns beside the features so the model and the reviewer see the same texture.

Parquet exports (format) and Polars/DuckDB (libraries) keep the engineering at laptop scale even for the full corpus.

Feature engineering on this record is mostly discipline: the venue's real signal lives in depth and timing near the bound, and every other feature family is measured noise unless it beats the structural baseline on held-out resolved markets.

Polymarket features should be scored against the venue's structural baseline before anything else: executable prices bunch on the 0.99/0.995 grid 98.6% of the time, and a 0.99-predicting model and a 0.51-predicting model are different statistical beasts. A feature family that cannot beat a calendar-and-grid dummy on held-out resolved markets is measurement noise.

The honest build order is therefore baseline, then feature, then claim — the record makes every comparison reproducible, and the resolved outcome column provides the label.

• Data for ML
• Imbalance data
• Mid-price caution