Are Underdogs Overrated?
How far do Polymarket odds at different price levels stray from actual results? We check the NBA, MLB, NHL and soccer.
- Result status
- Preliminary results
- Last updated
- Start date
- Duration
- Day 2 of 7
- Research team
- Mysports.AI Research Team
- Plan registered
Are Underdog Odds Overrated? Key Findings
Range −1.6 to +0.9 percentage points—indistinguishable from zero, so the conclusion is “no sign.” 4,346 underdogs across the NBA, MLB, NHL and soccer, all using odds 5 minutes before the start. Example: an underdog at an implied 25% (odds 4.00) actually wins about 24.8%.
- No sign in any sport separately either; but MLB and NHL games are mostly near 50/50, so the data are thin and no sign does not mean no bias.
- The bias does not grow further from the start (1, 4, 6, 24 hours before), unlike some foreign studies that find underdogs more overrated further from settlement.
- Worth watching in the NBA extremes: heavy underdogs below an implied 10% won about 5 points less, and heavy favorites above 90% about 5 points more. This is descriptive with small samples, left for the new season.
- No sign does not mean the market is perfectly accurate: we can only say the average over- or under-pricing of underdogs is roughly between −1.6 and +0.9 points.
In plain words
Why study this?
Horse racing and many bookmakers show a “favorite–longshot bias”: underdogs overrated, favorites underrated. If Polymarket has it too, a model using its odds should correct by price level first; it is also the baseline studies like R9 (popular teams) must net out.
What do we want to know?
Horse racing and many bookmakers show a well-known pattern: underdogs are often overrated and favorites underrated, perhaps because people like small stakes for big payouts. We ask whether Polymarket sports odds show the same pattern, across the NBA, MLB, NHL and soccer.
How do we study it?
We take the odds 5 minutes before each start, count each game once, and compare with the outcome. The main number is the calibration slope: 1 means no overall deviation; above 1 means the odds are not extreme enough—in two-team games, underdogs are overrated. For example, with a slope of 1.1, an underdog at an implied 25% wins only about 23%. Soccer has draws, so the three prices are first scaled to sum to 100%.
What have we found?
No sign that underdogs are overrated. Across the four sports, underdogs had an average implied chance of 25.5% and won 25.1%, just 0.4 point apart; the main test’s calibration slope is 1.01, showing no systematic deviation by price level. The same holds by sport and at different times before the start.
Why does it matter?
Five minutes before the start, Polymarket odds show no overall underdog–favorite bias, so a model using them needs no blanket correction; the NBA’s extreme underdogs and favorites are worth watching in the new season.
What can’t we say yet?
Many studies exist with mixed conclusions, and their periods overlap our 2025-26 data, so this is not a blind test; the new-season re-test is. We also do not answer whether betting on it would make money.
Terms used on this page
- Polymarket
- A prediction platform for trading game outcomes; prices run from $0 to $1. A price of $0.60 roughly means traders give that side about a 60% chance—not necessarily the true probability. Price converts directly to odds: odds = 1 ÷ price.
- Underdog / favorite
- In a game, the side with higher odds (less favored) is the underdog; the side with lower odds is the favorite.
- Underdog overrated
- The market gives the underdog a higher win chance than it actually achieves—its odds are too low.
- Calibration slope
- A single number describing how implied win chances relate to actual outcomes overall; 1 means no overall deviation.
- Percentage point
- The difference between two probabilities. For example, 50% to 51% is an increase of 1 percentage point.
Detailed results
Underdogs: implied versus actual win chance
One underdog per game (implied win chance below 40%); Polymarket odds 5 minutes before the start.
Per the pre-registered plan | NBA 2024-25 and 2025-26, MLB 2026, NHL 2025-26, soccer big-five leagues and Champions League 2025-26 | MLB and NHL have fewer underdogs and are not listed.
HypothesesFor deeper reading
- H1Across the four sports, odds 5 minutes before the start deviate systematically by price level → no sign (calibration slope 1.01, range 0.94–1.08)Insufficient evidence
- H2The NBA shows a bias → no signInsufficient evidence
- H3MLB shows a bias → no sign (thin data)Insufficient evidence
- H4The NHL shows a bias → no sign (thin data)Insufficient evidence
- H5Soccer shows a bias → no signInsufficient evidence
- H6The bias 4 hours before differs from 5 minutes before → no signInsufficient evidence
- H7For the same soccer matches, Polymarket’s bias differs from Pinnacle’s → holds per plan, but reverses with another common margin-removal method, so neither can be called more accurateExploratory
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.
- (Pre-registered extension) MLB and NHL games were too few to tell much in the main study, so we added more than 4,000 games (3,146 MLB games from the full 2025 season and other 2026 dates, and 902 NHL games from 2024-25), timestamping the plan before collecting any data. We still could not detect overpriced underdogs: MLB underdogs won 1.3 percentage points more often than their implied chance and NHL underdogs 0.5 points less, both indistinguishable from zero. Combining old and new games and converting through our model, an underdog with a 25% implied chance (odds of 4.00) is estimated to win about 23-30% of the time in MLB and about 22-30% in the NHL.
- (Post-hoc exploration) Comparing Polymarket's recorded odds history with MLB order-book records from summer 2026, we found the records were almost always the midpoint between the best bid and the best ask at that moment, not the last traded price (conversion: a price of $0.60 = about a 60% chance = odds of about 1.67). Between two hours and one hour before the first pitch, the odds often stayed exactly the same for the whole hour, yet every one of those hours had trades (a median of 34); the best bid and ask simply stayed on the same tick. "No change" does not mean nobody was trading.
- (Post-hoc exploration) NBA favorites with an implied chance of 90% or more (odds of about 1.11 or less) lost only 3 of 113 games over two seasons, while about 9 losses would be expected if the odds were exactly right, which looks like overpriced long shots. But we picked this cell after seeing the results, and most of it comes from the end of the 2025-26 season. So we fixed and timestamped the test in advance and will check it after the 2026-27 and 2027-28 seasons, in May 2028.
- (Post-hoc exploration) An outside analysis reported that Polymarket favorites are overpriced, but it only looked at games with at least $1 million traded. When we keep only high-volume games, we also see favorites winning less often (favorites with an 80-90% implied chance won only about 78%), but the gap disappears when all games are included. Trading volume is only known after the game ends, so selecting games by it may favor upsets and close games.
Limits of this analysis
- The three points marked post-hoc were added after we saw the main result. They are descriptions, not pre-registered tests, and they do not change the main finding that overpriced underdogs could not be detected. Only the extra MLB and NHL seasons were a formal test registered and timestamped before any data was collected.
- The extra MLB and NHL seasons are not a blind test either: outside studies may have covered the same periods, and the combined numbers include games we had already seen.
- Order-book records were only available for MLB in summer 2026 (698 games). The finding that the odds history is the bid-ask midpoint was only verified for these games; for other sports we only looked at the prices themselves.
- The new-season check of big NBA favorites has limited data: even if big favorites truly win 5 percentage points more often, two seasons together give only about a 60% chance of detecting it.
- No fees or bid-ask spreads were deducted, and none of this means that following these patterns would make money.
Favorite and Underdog Odds: Research Methods
Method summary
- Odds 5 minutes before each start, one count per game; games are not selected by volume.
- One primary test (all four sports combined); the six other comparisons are corrected together for multiple testing.
- We also look at 1, 4, 6 and 24 hours before the start.
- The new season will be re-tested with rules fixed now.
Data and limitations
Data we use (all public and free)
- Polymarket single-game odds: NBA 2024-25 and 2025-26, NHL 2025-26, MLB 2026, soccer big-five leagues and Champions League 2025-26
- Official scores (NBA, NHL, MLB, API-Football) to verify results
- Comparison: Pinnacle (early 2026-27 soccer)
Limitations
- Existing studies overlap the 2025-26 data, so this is not blind.
- A few soccer matches with extra time cannot be cross-checked and are listed separately.
- No historical bid/ask or fees, and no betting simulation.
- The Pinnacle comparison covers only about 420 early 2026-27 soccer matches.
Latest progress
Latest：Extension 1 (pre-registered): 4,000+ more MLB and NHL games still show no sign that underdogs are overrated; NBA heavy favourites are registered for a new-season re-test
Next step：Re-test with the same rules on new-season games (NBA, NHL, soccer) from 2026-10-23 to 2027-02-28
Study timeline
- Novelty and data checked; plan finalized and timestamped after two external code reviews
- Analysis done; corrected after two external code reviews
- Re-test on new-season games (2026/10/23–2027/2/28)In progress
Update log
Version history
- v1.0-results
Main results; 18 external review points handled over two rounds; one code bug fixed (8 postponed MLB games), conclusion unchanged
- v1.0
Plan finalized and timestamped (T0 = 2026-10-21 11:01 UTC) after two external code reviews
Study record
89b428e4a990270753e92be560250ff5cb44f4724779351cef00b49b5fe4a1a487ddf14669f461a953c1873de870a3117e80f1f8ba7e77ffed0c6b6f3c784cffThis page is a research log for reference only and is not betting advice or a guarantee of profit.