Every upcoming NBA fixture on Polymarket and every forecastable market on it — the whole board, uncapped. Keyless, live, cached a few minutes. A market is tradeable when its spread is ≤ 5¢ and ≥ $100 sits within 2¢ of the ask.
No forecastable NBA markets are open right now — an off-season or between-slate window, not an error. The historical record below is unaffected.
Every resolved NBAmarket the favorite-longshot backtest could reconstruct point-in-time from Polymarket's price history: 5,945 markets across 544 fixtures. The question is whether the market itself is mispriced — buy every favorite and win. Mostly it isn't — with one flagged exception below that we don't yet trust.
Each bucket: what the market charged, versus how often it actually happened. The gap averages 1.51c — the price is the probability.
| market price | n | actually won | gap |
|---|---|---|---|
| 0.071 | 87 | 0.046 | −0.025 |
| 0.150 | 148 | 0.149 | −0.001 |
| 0.251 | 200 | 0.235 | −0.016 |
| 0.449 | 2,606 | 0.472 | +0.023 |
| 0.546 | 2,699 | 0.554 | +0.008 |
| 0.769 | 137 | 0.759 | −0.010 |
| 0.895 | 68 | 0.941 | +0.046 |
Buy every favorite priced ≥ 70c, net of ~2c cost, significance clustered by fixture. In NBA moneylines it comes to +7.33c per $1 (t = 3.38) — across all market types, +5.82c (t = 2.86). That clears the conventional t = 2 bar — the one place across five sports where it does. We flag it and distrust it in equal measure: it is in-sample, it is one of ~20 slices tested, a same-shaped soccer signal already died at larger sample, and favorites firming in the final hour would fake exactly this. A hypothesis to test out-of-sample as fresh fixtures settle — not an edge, not a bet.
| market type | bets | fixtures | net edge | t (clustered) |
|---|---|---|---|---|
| Moneyline | 110 | 100 | +7.33c | 3.38 |
| Totals | 87 | 14 | −1.84c | 0.02 |
| Both teams to score | — | — | — | — |
| Spread | 8 | 6 | −24.86c | -0.99 |
| All types | 205 | 116 | +5.82c | 2.86 |
“—” means fewer than 8 favorites of that type — too thin to test. A real edge would need net > 0 andt > ~2.5 in a type that isn't just near-certain totals, and would have to survive out-of-sample.
