"There is a clearly demonstrated bias baked into Polymarket. Almost every bucket in the chart below resolves to "yes" at a lower rate than expected. Market participants seem to overestimate the likelihood of most events, especially the lower probability buckets. that they overpay for things happening. It's one of those things everyone in the space seems to accept as true." - Alex Mccolough

We decided to check. And started with a simple question:

Do people buy Yes more than they buy No?

What we found led us down a path from resolution rates to trade microstructure to a question about the editorial role the platform itself plays, and quickly realising the nuance of that question matters more than trying to answer it outright.

A note on what this is: This is a small, self-directed exploration of Polymarket as the most topically diversified binary prediction market venue...not an academic paper. It is a starting point, and deserves more continued depth from the Prediction community.

Two substantial studies have covered this territory with far more rigour and scale Becker's Jan '26 72M-trade Kalshi analysis and Deleep et al.'s 7M-observation UC Berkeley study from last week.

We include a brief comparison at the end. Ours isn't large, it's a few angles they didn't cover, full transparency, and a question about platform design that neither gets into. The Jupyter notebooks and data pipeline are HERE.

These findings encourage challenge.

Base Rates: Single-market events resolve NO 59% of the time

We pulled resolution data for 7,292 resolved single-market events on Polymarket - this is retro view and doing it in reverse chronological order is what we might do next. Single-market means one tradable question per event — "Will TikTok be banned?" not "Presidential Election 2024" (which has 17 candidate markets inside one event).

We excluded multi-market events because they introduce structural confounds (in a 17-candidate election, 16 markets resolve NO by construction). Why 7292? We ran out of patience with the query and decided it was enough. This gives us our base rate for resolution and allows to map Yes/No purchasing to the actual result.

The result: 41% YES, 59% NO. A naive NO bet wins roughly 6 in 10 times.

But this isn't uniform. The category breakdown is where it gets interesting:

The difference between categories is by our estimation, a question design effect. Polymarket predominantly frames markets around whether specific events will occur, and the default outcome in the real world is that specific things don't happen by specific dates. A naive no bet on the lower end wins even more frequently.

Hypothesis: what we tested and how

We wanted to know whether the behavioural YES bias reported in the literature and in industry commentary shows up in on-chain trade data. To do that we cut the ~7k markets to 88 resolved (top 500 by volume, filtered to single-market), across 28,793 trades, fetching both YES and NO token `OrderFilled` events from the Goldsky subgraph.

Three pre-registered hypotheses:

• H1: Traders prefer buying YES tokens over NO | YES buy proportion > 55%, p < 0.01 |

• H2: Small and large traders differ in YES/NO preference | YES buy % differs by > 10pp across size buckets |

• H3: "YES bias" is actually longshot bias channelled through question framing | YES preference disappears when controlling for token price |

Important caveat on our sample: Our 88 markets were drawn from the highest-volume events and showed 51% YES resolution - we took the first 200 trades, 10 percentage points above the population baseline of 41%. This volume-sorting selection bias means our sample over-represents YES-resolving events. There is a potential temporal bias here also, however it does not skew the data for the hypothesis test.

H1: Traders buy YES 51% of the time

Of 22,078 buy trades, 51% targeted YES tokens. Statistically significant (p=0.0001) but a small effect just above 50%, well below our 55% threshold.

In the context of the 41% population base rate, this is a 10-percentage-point gap between buying behaviour and resolution outcomes. Traders on these markets bought YES slightly more than NO, even though YES usually resolves substantially less than NO.

But the order book enforces balance. YES + NO prices sum to ~$1. When YES buying pushes the price up, NO becomes cheaper, attracting NO buyers. Most individual markets show YES buy rates between 47% and 55%. The constraint is structural - measuring bias through trade counts is likely wrong - H1 invalidated or just the wrong approach?

Side note a: We confirmed this holds on both maker-side (limit orders, 51.3%) and taker-side (market orders, 51.5%).

Observation: the ~77% BUY rate we observe on both YES and NO tokens seems to not resemble bias but position accumulation. Polymarket reports ~60% of traders hold positions to resolution, generating BUY trades with no matching SELL. This applies equally to both token types.

The conjecture view here, is buying and holding til death/victory just a symptom of all types of markets, or are these sorts of rates unique to prediction markets? some comparison to other market structures would be interesting.

H2: The trade-size gradient is real - but it's about price, not preference

Small trades (<$20) are 56% YES. Large trades ($500+) are 37% YES. A 19-percentage-point gap that holds within individual markets (63 of 88 show this pattern, p < 0.0001). This exceeds our 10pp threshold.

The immediate interpretation could be that small retail traders are biased toward YES, while larger traders know better.

However, the actual mechanism is different. When we control for token price level, the gradient essentially disappears. Within each price bucket, small and large trades show similar YES/NO ratios.

This chart is the key visual. Within each price bucket, the bars are nearly the same height.

What's happening here is that small trades cluster at low prices where YES tokens are cheap ($0.05–$0.20). Large trades cluster at high prices where NO tokens are available at favourable prices. Traders at every price level buy whichever token is cheaper.

H3: "YES bias" may be a question framing artefact

This is where it comes together. Our 88 markets have a median YES price of 34%. Forty-one markets have average YES price below 30%. Only 4 have YES price above 70%.

This reflects how Polymarket frames questions. The platform's most popular single-market events are "Will [unlikely/dramatic thing] happen?" - Trump resigning, Bitcoin hitting $250k, aliens confirmed. Dramatic long-shot questions generate volume. The editorial incentive is clear - these markets attract engagement and volume to the venue.

This creates a structural observation that in most markets, YES = the long-shot token. Any long-shot preference will therefore show up as YES preference in aggregate data.

We tested whether the preference is about the YES label or about cheapness:

Yes or No - traders don't care, just as long as its a bargain

When we bin all buys by what the trader paid per token (regardless of YES/NO):

• At 1–20c: ~90% of buys are YES tokens - but YES is the cheap long-shot here

• At 40–60c: dead 50/50 - neither token is cheap

• At 80–100c: ~90% of buys are NO tokens - NO is the cheap long-shot here

The curve is smooth, symmetric, and entirely explained by price. In the few markets where YES is the favourite (price > 60%), traders buy cheap NO at 50–55%, not expensive YES.

Traders don't prefer YES. They prefer CHEAP.

Polymarket's question design makes "cheap" and "YES" synonymous in most markets.

Combined with the resolution data: the "Will X happen?" framing simultaneously (a) makes YES the cheap token (most questions are about unlikely events) and (b) makes YES resolve less often (unlikely things usually don't happen).

The apparent "YES bias" may be the compound effect of long-shot preference (well-documented since Griffith 1949) channelled through editorial question framing.

To truly separate YES-specific preference from long-shot preference, you may need a platform that randomises which outcome gets the YES label- or perhaps the same question framed both ways across platforms. Neither condition currently exists.

Some Tangential findings

These weren't part of our original hypotheses but interesting threads that could be pulled further

Small trades are the only profitable bucket

Trades under $20 earn +$0.024 per trade. Every larger bucket is negative. YES buyers profit at every size, but edge shrinks with size. Note: this is measured on our 88-market sample which over-represents YES resolutions. The profitability advantage would likely shrink with a population-representative sample. Needs replication.

Wallet-level patterns

We really like this chart. Small-wallet makers average ~55% YES, large-wallet makers ~42% YES. Consistent with the price composition story - different sized wallets access different price ranges.

Limitations

• 88 trade-level markets, 7,292 resolution-level markets. Trade analysis is limited to 88 high-volume events. Resolution analysis covers the full population.
  
  

• Volume selection bias. Our trade sample over-represents YES-resolving events (51% vs 41% population).
  
  

• Maker-centric classification. We checked taker-side as a robustness test - same results - but taker behaviour deserves deeper analysis.
  
  

• First 200 trades per token. Our data skews toward early-life trading.
  
  

• Resolved markets only. Open markets, where bias might be most active, are excluded.
  
  

• No Mention Markets. Our single-market filter excludes celebrity/social media events - identified by the Deleep et al research as the most bias-prone category.

What we take from this

1. "YES bias" may be long-shot bias in disguise. Polymarket's question framing systematically assigns the longshot to the YES token. When we control for token price, traders buy whatever's cheap - YES when YES is cheap, NO when NO is cheap. The preference follows price, not the label.
   
   

2. Question framing is load-bearing. The "Will X happen?" structure simultaneously creates cheap YES tokens (unlikely events) and NO-heavy resolution rates (unlikely things don't happen). The editorial layer that decides how questions are framed shapes both how bias appears in the data and the actual base rates. Sports questions, which are symmetrically framed, resolve near 50/50. Everything else skews NO.
   
   

3. Market microstructure constrains what's observable. The order book's complementary pricing (YES + NO = $1) limits trade-count ratios to near 50/50. The trade-size gradient is a price composition effect. Different-sized capital accesses different parts of the price spectrum, and the data reflects that structure, not directional conviction.
   
   

4. Small doesn't mean dumb. Small traders capture positive edge in our sample, consistent with Becker's and Deleep et al.'s larger studies.
   
   

5. The unit of analysis matters. Population-level resolution rates vary dramatically by category (Sports 47% vs World 7%). Any aggregate "YES bias" claim that doesn't account for category composition is incomplete.

We'll continue running small experiments on questions we find interesting. If you have ideas or want to challenge any of this, the notebook is below.

————

*Analysis based on 7,292 resolved single-market events (resolution data) and 29,463 on-chain trades from 88 events (trade-level analysis). Full methodology and reproducible notebooks: [link to sprint]*

————

Appendix: Related work

Two substantial studies have examined prediction market bias at much larger scale:

Becker (2026) analysed 72.1 million trades on Kalshi ($18.26B volume) and found a systematic wealth transfer from takers to makers (-1.12% vs +1.12% excess return). His "Optimism Tax" concept (takers disproportionately buy YES longshots while makers take the other side) is consistent with our finding that traders prefer cheap tokens. His category breakdown (Finance: 0.17pp gap near-efficient, Entertainment: 4.79pp, Media: 7.28pp) parallels our resolution rate gradient. Crucially, he shows makers don't need to predict better - they profit structurally by being the counterparty to biased flow.

Deleep et al. (2026) at UC Berkeley analysed 5,456 markets across Polymarket and Kalshi (~7M observations) and found pervasive YES overpricing. They controlled for contract lifecycle timing and identified Mention Markets as the most biased category. Their finding that small traders capture positive edge (by fading whale bias) is consistent with both our results and Becker's.

Where we add something different: Neither study separates single-market from multi-market events, addresses the question framing confound, or tests whether the YES/NO preference is symmetric when controlling for token price. Our observation that traders buy whatever's cheap (not whatever's labeled YES) raises a question about how much of the documented "YES bias" is actually longshot bias channelled through editorial framing. This would require randomised YES/NO labelling or cross-platform experiments to resolve definitively.