NBA 90-99¢ Favorites Are Not Mispriced: Our Bias Hypothesis Fails at n=636

We test hypotheses and publish the results whether or not they support a position. This one does not.

Hypothesis: Are NBA markets closing at 90-99¢ correctly calibrated, or biased in one direction?

Verdict: NOT_SUPPORTED.

What we found

Across 636 graded outcomes drawn from 67 events, closing prices in the 90-99¢ band produced a record of 634-2. That is a realized win rate of 99.7% against a market-implied win rate of 98.8%. Blindly backing the band returned 0.9% ROI over those 636 outcomes.

The surface reading is that the market slightly underprices near-certain favorites. The statistical reading is that we cannot distinguish this from noise. The raw p-value on the calibration gap is 0.2507 — nowhere near any threshold at which we would act, and a value we would expect to see by chance roughly one time in four with no true edge at all.

The arithmetic of the band explains why. At 98.8% implied, the expected number of losers in 636 tries is around seven or eight. We observed two. The difference between "two losers" and "seven losers" feels enormous when you read the record as 634-2, but at this sample size it is a handful of coin flips. One additional upset in the sample moves the estimate materially. That is the definition of a result too fragile to trade.

The event count matters more than the outcome count. 636 outcomes came from only 67 events, so the effective independent sample is far smaller than 636 suggests. Outcomes within a single game share a game — the same blowout, the same injury, the same garbage-time collapse. Treating 636 correlated observations as 636 independent trials is the most common way a backtest lies to the person who ran it. We do not get to claim 636 degrees of freedom here, and the p-value already reflects a raw calculation that we would discount further, not less.

What it means

We are not adding a 90-99¢ calibration adjustment to the model. The 0.9% ROI is real in the sample and is not evidence of a repeatable edge; a positive number in a backtest is the beginning of an investigation, not the end of one.

There is also a practical point independent of the statistics. A 0.9% edge in a band where a single loss costs upwards of ten winners to recover is a strategy with severe negative skew. Even if the edge were real, the variance profile and the capital tied up per unit of return make it a poor use of a bankroll. Vigorish, limits, and fill quality at these prices would plausibly consume the entire margin before the edge could compound.

The honest summary: heavy NBA favorites closed approximately where they should have. Our model does not currently need a correction in this band, and we found nothing worth betting.

What would change our mind: a sustained gap between the 98.8% implied and 99.7% actual rates holding across several hundred additional independent events — enough that p drops well below 0.05 on an event-clustered basis rather than an outcome-count basis.

21+ where legal. Research only; not betting advice.