Do Teams Favored by AI Become More Favored Before Tip-Off?
Three AI models forecast win chances before games; we check whether they anticipated pre-game odds moves and compare their accuracy. Part 1 uses the new NBA season; Part 2 looks back at 2026 MLB forecasts already locked before games on aisportsinfo.com.
- Result status
- Part 2 has preliminary results; Part 1 is collecting
- Last updated
- Start date
- Duration
- Day 2 of 206
- Research team
- Mysports.AI Research Team
- Plan registered
AI Predictions vs Pre-Game Odds: Key Findings
Range −0.13 to +0.14, 821 regular-season games. This is a look-back study: the AI forecasts come from our own site, aisportsinfo.com, and the AI could see ESPN’s win probabilities. A price of $0.60 ≈ a 60% chance ≈ odds of about 1.67; more favored = lower odds.
- Accuracy: no sign the AI beat the market. In the full sample the AI’s error was actually larger than the market’s at the same moment (Brier 0.249 vs 0.238; lower is better), but among the 416 games with third-party pre-game proof it is not conclusive after correction.
- Beyond the market odds, there is no sign the AI numbers added anything about the results.
- Of the three, Claude was closest to the market; ChatGPT and Gemini were overconfident.
- Part 1 (the new NBA season, registered before collecting) will re-test whether AI can get ahead of the market.
In plain words
Why study this?
AI predictions are more and more common, but few people have tested “can AI see the direction before the market does” using predictions recorded and timestamped in advance. The results tell our models whether AI opinions are worth using as an extra signal.
What do we want to know?
We want to know three things: Are the win chances AI gives accurate? Does AI guess ahead of time which way the odds will change before tip-off? And is looking at “market odds + AI” together more accurate than looking at the market odds alone?
How do we study it?
Polymarket is a prediction platform where people buy and sell shares in game results. The price is between $0 and $1; a price of $0.60 roughly means the people trading think, as a group, that this side has about a 60% chance to win (the crowd’s view, not necessarily the true chance). For every NBA game, 6 hours before tip-off, we ask three large language models (Claude, GPT, Gemini) to each give a win chance for the home team. The AIs can’t go online; they only see the official schedule and scores, injury reports and player stats at that moment, and no odds at all. Every prediction is kept with its content and time and timestamped, so anyone can check whether it was changed later. We then compare it with the Polymarket odds at the same moment and the last odds before tip-off.
How do we compare AI with the market?
- 016 hours before: freeze inputs
Provide schedules, player data and injury reports, but no market odds or internet access.
- 02Three independent AI forecasts
Claude, GPT and Gemini each estimate the home team’s chance, without assuming which is best.
- 03Preserve forecasts and times
Record and timestamp forecasts so their timing and later changes can be checked.
- 04Evaluate after the season
Compare the odds at the time, the closing odds and game outcomes; evaluate the direction of odds separately from forecasting accuracy.
This is a study workflow, not an AI performance chart. No invented win rates or profit curves are shown before analysis.
What have we found?
Automatic daily collection started on 2026/10/21. We will analyze only once, after the regular season ends (2027/4/11), so no one can peek midway or pick the results they like. There are no results yet.
Track record already public
Our other site, aisportsinfo.com, publishes three AI models’ pre-game MLB forecasts and checks them after each game. Through Oct 20 of the 2026 season, the three-model consensus picked 494 of 878 games correctly (56%)—Claude 57%, ChatGPT 56%, Gemini 54%—versus 53.8% for always picking the home team. These hit rates are already public, so they are not blind; what nobody has computed yet is how the AI views compare with Polymarket odds, which is what Part 2 answers.
What can’t we say yet?
The AIs only see the data we give them, so the results don’t show how AI would do when it can look things up online. The three AIs run on existing subscription plans; when usage limits run out, a few games will be missing, and any gaps will be recorded honestly. This study tests general AI models, not Mysports.AI’s own prediction models.
Terms used on this page
- Polymarket
- A prediction platform where people buy and sell shares in game results. Prices run from $0 to $1 and roughly show the chance the traders think an outcome has — not necessarily the true chance. A price converts directly to odds: odds = 1 ÷ price.
- Large language model (AI)
- AI such as ChatGPT, Claude and Gemini that can read text and answer questions.
- Timestamped record
- Keeping the content and time of every prediction and timestamping it, so anyone can check that the prediction was made before the game and wasn’t changed afterwards.
- Closing odds
- The last odds recorded before the game starts; they usually reflect the most news.
Detailed results
HypothesesFor deeper reading
- H1AI’s win chances will be less accurate than the market’s at the same momentPending
- H2The gap between AI and the market probably can’t predict which way the odds go laterPending
- H3“Market odds + AI” is probably not more accurate than looking at the market alonePending
- P2-H1(Part 2) The AI is as accurate as the market at the same moment → less accurate in the full sample, but not conclusive among games with pre-game proofInsufficient evidence
- P2-H2(Part 2) The AI–market gap predicts later moves toward the AI → a weak link in the full sample that disappears among games with pre-game proof and may come from stale price recordsInsufficient evidence
- P2-H3(Part 2) The AI adds predictive value beyond the market odds → no signInsufficient evidence
Further analysis
Further analysis done after the main results. Items marked “post-hoc” were added after seeing the data and are for reference only; items marked “pre-registered” or “Extension 1” had their rules written and timestamped before anything was calculated.
- The price records are fine; mostly the market itself rarely moves (a check added after seeing the results): against a public order-book archive, for about 520 games after Aug 18 we checked Polymarket's price at the moment the AIs answered, and in 518 it matched the exact middle of the bid and ask (a price of $0.60 ≈ a 60% implied chance ≈ decimal odds of about 1.67). Within 15 minutes after the AIs answered, that price changed in only about 4% of games; in the early US morning the bid and ask are just 1¢ apart and often do not move for tens of minutes.
- Re-running with real trade prices, there is still no sign that the market increasingly favours the team the AIs favour: the original small association shows up about as strongly when the AI is replaced by a flat 50% guess; using separate trades (541 games with enough trading), the AI's association could well be zero, and the AI-favoured team's implied win probability moved by only about +0.05 percentage points on average before the start — essentially no change.
- The AIs are over-confident. Pulling them part-way back to 50% clearly reduces the error for ChatGPT (Codex) and Gemini (e.g. turning an AI’s 70% into roughly 60%), but not for Claude; after the adjustment all three still trail the market at the same moment. When we blended AI numbers into the market price, practising on some weeks and testing on others, the weight chosen for the AI was 0.
- Part 1 (NBA) collection has started: from Oct 21, 2026, the three AIs are asked six hours before each tip-off (the first batch is the Oct 21 US-evening games). Part 1 results will not be looked at until the season ends in April 2027; until then we only count games, not accuracy.
Limits of this analysis
- All of this was added after seeing the Part 2 results and does not change its conclusions; the forecasts still come from our own site aisportsinfo, and the AIs could see ESPN's win probabilities.
- The order-book archive covers only games whose forecasts were locked after Aug 18, and we checked just two moments per game; the trade-price version includes only the 541 of 826 games with enough trading, which tends to leave out later games and games forecast earlier.
- The 'pull back toward 50%' and 'blend with the market' methods were chosen after seeing these data; they only count once re-tested on new games with a method fixed in advance.
AI Prediction Comparison: Methods
Research questions
- Are the win chances AI gives accurate?
- Does AI guess ahead of time which way the odds move before tip-off?
- Is “market odds + AI” more accurate than looking at the market alone?
Method summary
- The instructions given to the AI and the AI versions are written into the plan in advance, with a “digital fingerprint” of the files kept, so they can’t be secretly changed later. The AI runs in an empty folder and can’t go online or use tools; if any market-related word appears in the data, it isn’t sent.
- We analyze only once, after the regular season ends; if fewer than 600 games are usable, we only describe the results and draw no conclusions.
- Part 2: 2026 MLB forecasts already locked before games and pushed to GitHub by aisportsinfo.com (about 829 regular-season games), compared with Polymarket odds at the same moment and at close. We checked that the AI inputs contained no odds and almost no web access, but they did include ESPN’s win probabilities, so conclusions describe AI performance after seeing ESPN’s numbers.
Data and limitations
Data we use
- Pre-game win chances from three large language models (6 hours before each game, recorded and timestamped)
- Polymarket odds for single NBA games (shared with R1)
- Official NBA schedule, injury reports and player stats
- aisportsinfo.com: three AI models’ 2026 MLB pre-game forecasts and record (Part 2; same team as Mysports.AI)
Limitations
- The AI can’t go online, so the results don’t show how AI with internet access would do.
- We use existing subscription plans, so there will be gaps when the usage limits run out.
- Part 2 is a look-back study using data our own site produced; Claude’s logs before 9/20 were deleted by the system, so a version excluding those forecasts will be published alongside.
Latest progress
Latest：Part 2 extension (post-hoc): re-checked with order books and trade prices, still no sign the AIs are a step ahead of the market; Part 1 (NBA) has been collecting daily since Oct 21
Collecting：Win chances from three AIs, 6 hours before each game of the 2026-27 NBA regular season (recorded and timestamped)
Next step：Analyze Part 1 after the regular season ends in April 2027 to re-test whether AI can get ahead of the market
Study timeline
- Novelty checked; plan finalized and timestamped
- Part 2 (2026 MLB look-back): data audited, plan finalized and timestamped
- Collecting AI predictions every day, 6 hours before each game (until 2027/4/11)In progress
- Part 2 analysis and external code review done
- One analysis after the regular season ends, external code review, publishing results (by 2027/5/15 at the latest)
Update log
Version history
- p2-v1.0-results
Part 2 results; the analysis code was timestamped before running and corrected after two external reviews
- p2-v1.0
Part 2 plan finalized and timestamped (T0 = 2026-10-21 08:23 UTC) after two external reviews
- v1.0
Study plan finalized and timestamped (T0 = 2026-10-21 05:22 UTC); collection started
Study record
b35be5c62216364b0cf9ab6e454ff91b09f5583cc90c200e6b84332d0c7e99c2b186f633f44d48af9ea11e933298ba28b8db7a8d522b406c3de79ffffb42f464014c0f9debbd17614cc84daec96315175d3ce5daa537d78e4029fc6a68c4da547d881bc57a6e3802489783284c74dcbf5c59a7f50f2cc0a964ec0824ea1eb862This page is a research log for reference only and is not betting advice or a guarantee of profit.