R23Published

In NBA Games, Do Teams the Live Odds Give a 90% Chance Really Win 90% of the Time?

During a game, when a team’s Polymarket odds first drop to 1.11 or lower (a 90%+ chance), how often does that team end up winning, and how far is that from what the odds said?

Result status
Preliminary results
Last updated
Start date
Duration
Day 1 of 1
Plan registered
Illustration of ten dots in a row, nine green and one yellow, showing that a 90% chance still loses about once in ten
Image:Illustration (made by Mysports.AI)

Do NBA Teams at 90% In-Game Odds Win 90%? Key Findings

Finding so far
How much higher (+) or lower (−) the actual win rate was than the chance implied by the odds, for teams whose odds first reached 1.11 or lower:+0.3pts

694 games, 2025-26 regular season; actual win rate 90.8% vs 90.5% implied; 95% range −1.9 to +2.4, which includes 0 = no difference can be seen (not the same as exactly equal). Only games that reached 90% within 105 minutes of tip-off; we do not have each game’s end time, so post-game trades cannot be fully ruled out.

  • Share that lost in the end: 9.2%, versus 9.5% according to the odds.
  • Other settings give the same answer, no difference can be seen: a 120-minute window (864 games, −0.1), estimating the end of the game from prices (1,059 games, +0.5), no confirmation rule (709 games, +0.3), prices shifted by a fixed 0.005 (+0.7), excluding games with several markets (687 games, +0.2). Suspected post-game price jumps: 0 games (our pre-set flag caught none; this does not prove there were no post-game trades).
  • Other thresholds (descriptive, not tested, no multiple-comparison correction, only a lead): teams that first reached 1.25 or lower (80%) won 82.7% vs 82.0% implied (946 games); those that reached 1.05 or lower (95%) won 97.3% vs 95.1% (510 games), an estimate pointing the same way as the “too cautious” finding abroad.
  • The 66 teams that were already 90%+ favorites before tip-off won 97.0% (92.6% implied), also pointing the same way as studies abroad; the 628 that only got there during the game won 90.1% (90.3% implied) (descriptive, not tested, only a lead).
  • In these games, the later a team first reached 90%, the lower its actual win rate: 92.2% within 48 minutes of tip-off (332 games), 91.5% at 48–72 minutes (82), 90.1% at 72–96 minutes (202), 85.9% at 96–105 minutes (78); the implied chance was 90.3%–90.6% in every group. This is descriptive with few games and does not show that late-game 90% odds are too optimistic.

How to use this yourself

  1. During a game, look at the “who will win” market (moneyline) for that NBA game on Polymarket. Polymarket’s help page explains where the price next to each team comes from. Open source ↗

  2. The price next to each team converts to odds: a price of $0.90 = about a 90% chance = odds of about 1.11 (odds = 1 ÷ price). The number on screen is usually the midpoint of the best bid and ask, while this study used actual trade prices, so they can differ slightly.

  3. When a team first reaches 0.90 or more (odds of 1.11 or lower), use this page’s numbers: on past average such teams won 90.8%, losing about 1 game in 11, close to what the odds implied. So do not read “1.11” as “locked”: there is still about a 10% chance of losing.

  4. There is also no basis for adjusting 90% up or down yourself: no tilt toward too cautious or too optimistic can be seen (+0.3 percentage points, range −1.9 to +2.4). Games that first reached 90% very late (96–105 minutes after tip-off) won only 85.9%, and teams that were already heavy favorites before tip-off won 97.0%; both are descriptive with few games, so do not treat them as signals.

These results come from 694 games in Polymarket’s 2025-26 NBA regular season, looking only at the first moment a team reached 90% within 105 minutes of tip-off. They cannot be applied directly to Taiwan Sports Lottery or other bookmakers (bookmaker odds include a margin, so the chance converted directly from them comes out too high), and they do not cover teams that first reach 90% in the last ten-plus minutes; we do not have each game’s end time, so post-game trades cannot be fully ruled out. This is a past average with no guarantee for any game; this page is not betting advice and does not teach in-game betting.

In plain words

Why study this?

Helps you read in-game odds: when the odds show a 90% chance, how common it is for that team to lose in the end.

What do we want to know?

During an NBA game, a team’s odds on Polymarket drop to 1.11 or lower (a 90%+ chance) for the first time. Does that team really end up winning about 90% of the time?

How do we study it?

We used public trade-by-trade data for 1,067 games in the 2025-26 regular season. In each game we looked at one moment only: after tip-off, the first time a team traded at a price of 0.90 or more, confirmed (to rule out a single odd trade, another trade at 0.90+ had to happen in a different second within 60 seconds). We recorded the price in that second and checked whether the team won. To avoid trades after the game ended, when the result was already known, we only looked at the first 105 minutes after tip-off (most NBA games take over 2 hours from tip-off to the final buzzer). The plan and the analysis code were timestamped before we looked at any results.

What have we found?

In 694 games, one team reached 90% within 105 minutes of tip-off. Those teams won 90.8% of the time, versus 90.5% implied by the odds: a gap of +0.3 percentage points, range −1.9 to +2.4, which includes 0. No difference can be seen, and the gap is likely within 3 points. Using 120 minutes, estimating the end of the game from prices, dropping the confirmation rule or shifting prices by a fixed 0.005 all gave the same answer: no difference can be seen. Only games that reached 90% within 105 minutes of tip-off count; those that first got there in the last ten-plus minutes do not; we do not have each game’s end time, so post-game trades cannot be fully ruled out.

What does this mean?

Teams whose in-game odds first (confirmed) show a 90% chance lose about 1 game in 11, close to what the odds implied (this does not prove they are exactly the same). So “1.11 means it’s locked” is wrong: there is still about a 10% chance of losing, which is exactly what the odds were telling you. Studies on another platform (Kalshi) found high prices late in games to be too cautious; we did not see that here, but the two look at different moments (we take only the first time each game reaches 90%, up to 105 minutes after tip-off, excluding the last ten-plus minutes), and we did not test why they differ.

What can’t we say yet?

What we cannot tell yet: we do not have each game’s end time, so looking only within 105 minutes of tip-off reduces, but cannot fully rule out, post-game trades, and teams that first reach 90% in the last ten-plus minutes are not covered. At the 95% threshold (97.3% vs 95.1%) and for teams that were already big favorites before tip-off (97.0% vs 92.6%), teams won more often than the odds said, in the same direction as studies abroad, but these are descriptions without multiple-comparison correction, so we cannot say “odds of 95% or more are too cautious”. There is no comparable trade-level data for the new season to test it again.

Terms used on this page

Odds
Decimal odds converted from a Polymarket price: odds = 1 ÷ price. A price of $0.90 = a 90% implied chance = odds of about 1.11.
First time (confirmed)
To rule out a single odd trade, one trade at 0.90+ is not enough; another trade at 0.90+ must happen in a different second within 60 seconds, and we record the price in that later second.
Percentage point
The unit for the difference between two percentages: 90% to 91% is 1 percentage point.
Range
A 95% interval: based on these games, the gap most likely falls in this range. We only say there is a difference when the range does not include 0.

Detailed results

Teams whose odds first reached 1.11 or lower: actual win rate minus implied chance (percentage points)

2025-26 regular season. The first row is the pre-registered main result; the next 5 rows are pre-listed alternative calculations; the last 3 rows change the starting point or threshold and are pre-listed descriptions (descriptive only, a different question). Positive = teams won more often than the odds said.

Main result (105 minutes after tip-off, 60-second confirmation)pts0.3
120-minute windowpts-0.1
Game end estimated from pricespts0.5
No confirmation rulepts0.3
Prices shifted by a fixed 0.005pts0.7
Excluding games with several marketspts0.2
Starting from the scheduled tip-off timepts1.0
80% threshold (odds 1.25)pts0.7
95% threshold (odds 1.05)pts2.1

The main result’s range is −1.9 to +2.4 percentage points, which includes 0 = no difference can be seen. Only games that reached 90% within 105 minutes of tip-off are included, and post-game trades cannot be fully ruled out. The other rows are for reference only. This is research, not betting advice.

HypothesesFor deeper reading
  • H1Odds of 90% are too cautious (teams win more than the odds say) → no difference can be seen: they won 90.8% vs 90.5% implied, a gap of +0.3 percentage points (range −1.9 to +2.4); the gap is likely within 3 points. This does not mean they are exactly the same (only games that reached 90% within 105 minutes of tip-off; post-game trades cannot be fully ruled out)Insufficient evidence
  • D1Teams whose odds first reached 1.05 or lower (95%) won 97.3% vs 95.1% implied (descriptive, not tested)Exploratory
  • D2The 66 teams that were already 90%+ favorites before tip-off won 97.0% vs 92.6% implied once confirmed at 90% during the game (descriptive, not tested, few games)Exploratory

NBA In-Game 90% Odds: Research Methods

Research questions

  • When a team’s odds first (confirmed) drop to 1.11 or lower during a game, how many percentage points apart are its actual win rate and the chance implied by the odds at that moment?
  • Does it look the same at odds of 1.25 (80%) or 1.05 (95%), or for teams that were already heavy favorites before tip-off? (descriptive, not tested)

Method summary

  • There is one main result: one event per game, and the average of “won or not (win = 1, loss = 0) minus the price at that moment”, in percentage points. If the odds always equalled the true chance at each moment, this average should be 0.
  • We use only the first time each game reaches 90%: counting every minute would count one-sided games dozens of times and change the question. We use the actual trade price at that moment rather than 0.90, because prices often jump past 0.90.
  • “During the game” runs from the tip-off time in ESPN’s records to 105 minutes later; tip-off is usually about 11 minutes after the scheduled time, and that gap is still pre-game.

Data and limitations

Data we use (all public and free)

Limitations

  • Retrospective, not blind: game results are public; we had not analysed NBA in-game odds before, but an earlier study read the same data files when preparing trade data (using pre-game periods only).
  • We do not have each game’s end time: looking only at the first 105 minutes reduces, but cannot fully rule out, trades after the game ended; it also means games that first reach 90% only in the last ten-plus minutes are not included.
  • Trade prices include the bid-ask spread and we cannot see quotes; the number Polymarket shows on screen is usually the midpoint of the best bid and ask, which can differ slightly from trade prices.
  • Only the 2025-26 regular season and Polymarket’s first-generation contracts; 21 days are missing from the data files, so games on those days are not included. Only games settled normally (one winner, one loser). No comparable trade-level data exists for the new season.
  • We look only at the first moment each game reaches 90%; this does not mean 90% odds are equally accurate at every moment. The 78 games that first reached 90% 96–105 minutes after tip-off won only 85.9% (descriptive, not tested), and the very end of games is not covered.

Latest progress

Latest:Main results are in: teams whose in-game odds first reached 1.11 or lower won 90.8% of the time versus 90.5% implied; no difference can be seen (only games that reached 90% within 105 minutes of tip-off). The plan and analysis code were timestamped first, and results were reviewed by an AI code-review tool

Next step:Study completed; no comparable trade-level data exists for the new season, so there is no forward test; we will reconsider if new public trade-level data appears

Study timeline

  1. Data gate passed (games and trades counted only, no prices or results looked at)
  2. Plan finalized and timestamped after one round of AI-tool code review; analysis code tested on synthetic data and timestamped before running
  3. Main results in; one round of AI-tool code review

Update log

Version history

  1. v1.0-results

    Main result: no difference can be seen (actual win rate 90.8% vs 90.5% implied, +0.3 percentage points, range −1.9 to +2.4)

  2. v1.0

    Plan finalized and timestamped (T0 = 2026-11-19 02:20 UTC); retrospective, not blind; no forward test because no comparable trade-level data exists for the new season

Study record

Plan fingerprint (SHA-256)
Plain-language plan8e91fb7b8e37d22ce3344d7096bda6134efb41597de183dbddf0baa8eee20f5a
Technical appendixcc437e58fac8f8be081af7fa8fa912b9ae404bc90e47d583235502c6332cbe58
Data gate (thresholds and count results)d5aa516fd98fd79e3f6db20145c08b980c8c93f0b40ad31716f33dd9d5065799
Analysis code timestamp (before running)fcd96b0602cc688b0219d3b50fa6572130d07a3abac0dddcfa5ec77bb1d0a779
Second analysis code timestamp (before results; data check only)d0c13274f40c30b31f0673f8356dc131e5e60b353df0b3eaf6fcd1fc22112c31
Timestamped into the public Bitcoin blockchain record with OpenTimestamps on Nov 19, 2026
Plan changes:8 (each recorded in the deviation log, DEVIATIONS.md; not public for now)
Code and data:Not public for now (contains third-party data sources and internal information)

This page is a research log for reference only and is not betting advice or a guarantee of profit.