Learning Polymarket Taker Trade Direction from the On-Chain Tape

Classical trade-direction classifiers, the tick rule and bulk-volume classification, perform near randomly on prediction markets, reaching roughly 50% accuracy on the full Polymarket-v1 archive (Qin and Yang, 2026), because order flow there exhibits positive direction autocorrelation and long same-price runs that violate the mean-reversion logic those rules assume. Prior work documents this failure with classical econometrics but does not build a predictive model that exploits the structure. We ask whether a learned classifier can infer taker direction from the trade tape alone, validated against on-chain ground-truth direction. On a representative liquid month, we engineer a fully causal, leakage-audited feature set and compare gradient-boosted trees and a 1D convolutional sequence model against a price-bin-corrected tick rule. The learned models improve direction accuracy by roughly 16 percentage points over the corrected baseline and stay flat across price levels, where the classical rule is price-dependent. An ablation shows the signal comes from the recent magnitude and short-window trajectory of price changes, not maker identity or explicit run-length features, and both models are well calibrated. Substituting predicted directions into standard microstructure measures recovers ground-truth Order Flow Imbalance and VPIN markedly better than the classical rule. We report relative improvements on a single-month panel and make no full-panel accuracy claims.