Method and results
Does the model beat the price?
The baseline is the market itself. ZingerLM predicts a correction to the market’s log-odds, so an untrained model equals the price, and any edge has to show up on markets it never saw.
Data
Resolved Polymarket markets with at least $50,000 volume, collected from public APIs: each market’s YES price every 10 minutes for up to 14 days before close, including in-play trading. The split is by close time. Every holdout market closed after the last training market.
- Markets
- 10,154
- crypto
- 2,614
- finance
- 774
- sports
- 6,766
- Observations
- 8,809,229
- Train / val / test markets
- 7,107 / 1,016 / 2,031
- Test closes
- 2026-09-15 → 2026-10-20
Model sizes
| Preset | Parameters | Width × layers | Status |
|---|---|---|---|
| nano | 1.7M | 128 × 8 | Trained; results below |
| small | 25.8M | 512 × 8 | Defined |
| 350m | 354.3M | 1024 × 28 | Forward and backward pass verified; GPU training script ready |
| 500m | 514.0M | 1280 × 26 | Forward and backward pass verified; GPU training script ready |
At about 8.8M observations, the dataset is far smaller than a 350M model would normally need. The larger presets are worth training only as the market archive grows, and only if they beat the small model on the same holdout.
Holdout: model vs market
| 2,031 test markets | ZingerLM | Market | Change |
|---|---|---|---|
| Log loss | 0.6099 | 0.6105 | +0.10% |
| Brier score | 0.2142 | 0.2144 | +0.12% |
95% bootstrap interval for the per-market log-loss gain: [-0.0001, +0.0013]. The interval includes zero, so this holdout does not show a reliable edge. Preset nano, 1,500 steps, CPU.
Holdout backtest
One entry per test market when edge after fees reaches 3¢, quarter-Kelly, capped at 3% of a $10,000 bankroll, held to settlement, with an assumed 1¢ half-spread. The control run uses the market price as the model and should make no trades.
| Segment | Trades | Win rate | Staked | P/L | ROI on stake |
|---|---|---|---|---|---|
| All test markets | 0 | – | – | – | – |
| Control (market as model) | 0 | – | – | – | – |
- P/L 95% bootstrap interval
- $0 to $0
- Max drawdown
- 0.00%
- Performance fees the product would have earned (20%, weekly, high-water mark)
- $0
- User P/L after fees
- $0
What happens if the bar is lowered
The model’s distance from the market price is almost the same everywhere: median 0.95¢, maximum 1.17¢. It has learned a small uniform correction, not market-specific signal. Holdout backtest re-run at lower entry thresholds and spreads. Control (market price as model) made 0 trades in every row.
| Min edge | Half-spread | Trades | P/L | ROI on stake | P/L 95% interval |
|---|---|---|---|---|---|
| 0.0¢ | 1.0¢ | 70 | −$18.73 | -15.04% | −$46.82 to $11.65 |
| 0.0¢ | 0.5¢ | 154 | −$327.82 | -9.30% | −$902.45 to $239.61 |
| 0.5¢ | 0.5¢ | 133 | −$389.87 | -11.24% | −$974.57 to $213.22 |
| 1.0¢ | 0.5¢ | 0 | – | – | – |
| 3.0¢ | 1.0¢ | 0 | – | – | – |
Limitations
- Prices are 10-minute samples, not executable quotes; fills assume a fixed half-spread.
- Depth is assumed sufficient up to the stake cap; real books may not be.
- One entry per market, held to settlement; correlated markets in one event are not netted.
- Out-of-sample by close time only; results are a research estimate, not a performance record.