←BACK TO BLOG
>

The Favorite-Longshot Bias Doesn't Care How Smart You Are

Prediction MarketsQuantMarket Microstructure
JUL 2026·6 MIN READ

#THE PATTERN THAT WON'T DIE

Richard Griffith documented it in 1949 and nobody has managed to make it go away since: bettors systematically overpay for longshots and underpay for favorites. It shows up in pari-mutuel horse racing data from essentially every country that runs a track, with the exceptions rare enough to be footnotes. It shows up in tennis betting markets too, which rules out "horses are weird" as an explanation. Seventy-five years of replication is about as close to a physical constant as social science gets.

The magnitudes are the part people underrate. A bet on a horse near 1/1 odds returns roughly 85 cents on the dollar, a 5-15% takeout-adjusted loss, which is already bad but at least in the ballpark of the track's cut. A 30/1 shot returns about 63 cents on the dollar. Push out to 100/1 or beyond and you're looking at a rate of return around -61%. The house edge doesn't stay flat as you move down the probability curve; it compounds, hard, exactly where the payout looks most exciting.

#THE STANDARD ALIBI

The textbook explanation has always had a demographic flavor to it: racetrack bettors are recreational, they're chasing a story, they don't have the numeracy to correctly discount a 100-to-1 shot, and they'd rather lose small and often on a dream than win small and often on a favorite. Snowberg and Wolfers formalized the two competing versions of this in NBER working paper w15923, using a huge exacta/quinella/trifecta dataset to ask whether the bias comes from rational risk-love (people genuinely enjoying lottery-like payoffs) or from Prospect Theory-style misperception of small probabilities. Their evidence leans toward misperception: people just don't weight tail probabilities correctly.

Either way, the implicit background assumption was that this is a retail phenomenon. Give the market enough sharp, well-capitalized participants and the mispricing should get arbitraged down, the same way it did in options markets once market makers got good at pricing tails. The bias was treated as a bug in the bettor pool, not a bug in pricing itself.

#THEN A CRYPTO-NATIVE ORDER BOOK SHOWED UP

Polymarket is about as far from a betting window at Santa Anita as a market can get. It's a blockchain-settled, on-chain central limit order book running on Polygon, open globally, and its flow includes algorithmic desks, professional traders, and crypto-native funds alongside everyone else, precisely the population that was supposed to arbitrage the favorite-longshot bias out of existence. Philipp Dubach's 2026 paper, 'The Anatomy of a Decentralized Prediction Market,' takes this market apart at the microstructure level: roughly 30 billion order-book events over 52 days, joined to on-chain trade records, across a pre-registered panel of 600 markets. It's the kind of dataset that doesn't leave much room to argue the sample was cherry-picked.

The first stylized fact Dubach reports is what he calls a longshot spread premium: quoted bid-ask spreads on the lowest-probability decile of outcomes are dramatically wider than on high-probability outcomes. Structurally, it's the same shape as the racetrack curve: the market gets worse exactly where the implied probability gets small.

The Longshot Tax at the Racetrack

illustrative
0.85~1/1 (favorite)0.7610/10.6330/10.39100/1+

Return per dollar bet slides from $0.85 near even money to just $0.39 on 100/1+ longshots, a real, monotonic 'tax' documented across decades of racetrack studies (the 10/1 bucket is an illustrative interpolation showing the same downward trend).

#THE NUMBER THAT KILLS THE EXCUSE

Here's the part that should bother anyone still leaning on the unsophisticated-bettor story: Dubach measures the full quoted spread on Polymarket's lowest-probability decile at roughly 1,300 to 1,800 basis points, an order of magnitude wider than classical longshot premia in horse racing. This is not a smaller version of the racetrack effect surviving contact with smart money. It's a bigger version of it, on a venue that skews toward exactly the participants who were supposed to correct it.

Compare the two:

  • ▸Horse racing, favorite (~1/1): ~$0.85 return per dollar bet
  • ▸Horse racing, 30/1 longshot: ~$0.63 return per dollar bet
  • ▸Horse racing, 100/1+ longshot: roughly -61% rate of return
  • ▸Polymarket, lowest-probability decile: ~1,300-1,800 bps quoted spread, an order of magnitude wider than the racetrack premium
  • ▸Polymarket, self-counterparty (wash) trading: median ~1% of volume, 22% upper tail, present, but not remotely large enough to manufacture this on its own

#WHAT'S ACTUALLY GOING ON

Dubach's own read leans partly on a liquidity-provision constraint rather than pure misperception or risk-love: quoting a tight two-sided market on a 2-cent YES is genuinely expensive. Your inventory risk is asymmetric and your hedges are bad: the position can go to zero or to a dollar, and there's no clean way to lay off gamma near a probability boundary the way you can mid-curve. That's a structural cost of market-making at the tails, not a behavioral quirk of the person on the other side of the trade.

Which is what makes this finding sharper than a simple replication. Snowberg and Wolfers were right that misperception matters: humans misweight small probabilities, on a racetrack or anywhere else. But 'unsophisticated bettors' was doing double duty as an explanation for both the mispricing and the reason nobody arbitraged it away. Strip out the unsophisticated bettors, hand the order book to algo desks and professional flow, and the effect doesn't shrink; it gets worse. That's only possible if the binding constraint was never bettor IQ. It's the cost of providing liquidity in the tail, and that cost doesn't care who's on either side of the trade.

#THE POINT

I don't think this is noise waiting for someone smarter to show up and correct it. Correcting it would require someone to make markets in extreme tail probabilities for free, and nobody does that, not at the track, not on a crypto-native order book with professional flow and sub-50ms data feeds. The favorite-longshot bias looks less like a market failure and more like a structural risk premium: the price of insurance against a rare event will always be quoted wide, because bearing that risk is genuinely costly, independent of venue, technology, or who's holding the bag.

For anyone trading prediction markets, the actionable version of this is boring but real: be structurally suspicious of anything priced below roughly 5-10% implied probability, and treat consistent overpayment there as a standing source of edge on the sell side rather than a place to buy lottery tickets, the same lesson the racetrack has been teaching for 75 years, just now denominated in USDC instead of two-dollar win tickets.

Spread Widens as Probability Shrinks

illustrative

Polymarket's quoted bid-ask spread runs roughly 1,300-1,800 basis points for the least-likely outcomes (the reported figure from Dubach 2026) and narrows sharply toward near-certain outcomes, echoing the racetrack's favorite-longshot curve; only the longshot-decile point is a real number, the rest of the curve is illustrative of the paper's described shape.

#Further Reading