Polymarket · sports, finance, crypto
Trade only the deep books, only when the model disagrees with the price.
Zinger Books scans Polymarket for books you can actually fill, prices each market with a trained sequence model, and sizes a position only when the edge survives the spread and taker fee. The planned business model is a fee on profit, charged above a high-water mark.
- Resolved markets trained on
- 10,154
- Price observations
- 8,809,229
- Data window
- 2025-11-10 → 2026-10-20
- Model sizes built
- up to 514M params
- Holdout vs market (log loss)
- +0.10% (not significant)
- Holdout backtest P/L
- no trades cleared the 3¢ bar
How it works
Find depth
Rank live books by the dollars tradeable within 2¢ of the best price on both sides. Wide or thin books are dropped before the model sees them.
Price the market
ZingerLM, a causal transformer trained on resolved markets, reads the price path, time to start and in-play state, then outputs its own probability.
Clear the costs
A trade needs at least 3¢ of edge per share after the ask and Polymarket’s taker fee. Size is quarter-Kelly, capped at 3% of bankroll and 10% of visible depth.
Pay on profit
Users keep custody. A performance fee applies only to gains above their previous peak, after deposits and withdrawals. Losing weeks cost nothing.
Where it stands
The training pipeline, liquidity scanner, fee-aware sizing, profit-fee billing and paper trader all run. The model does not yet beat the market: on 2,031 markets that closed after training ended, its gain over the price is within noise, and it never finds 3¢ of edge after costs. Lowering the bar makes it trade and lose. The Research page has the full numbers.
- No real-money execution. Wallet delegation and per-user risk limits are not built.
- The 350M and 500M models are defined and verified, but need a rented GPU to train.
- Charging a share of profits on other people’s trading is regulated in most places. Legal review comes before onboarding.