Archive / Prediction markets / E080
Backing favourites, draws and sure things in sports markets fails
Buying esports favourites looked like +4.5c per share in a first sample but was -1.0c on 755 untouched events; soccer draws, favourites and near-certain outcomes were priced within the noise.
Bettors are said to overpay for underdogs and underpay for favourites, to ignore soccer draws, and to leave near-certain outcomes cheap. We tested all three on pre-game sports and esports markets of an on-chain prediction market. Buying esports favourites one hour before the match returned +4.5c per share in the first sample, but that figure assumed a purchase price that could not have been known in advance. Priced at the last quote before the decision it was +1.3c, and on untouched events it was -1.0c with the original fill (755 events) and -0.5c priced at the last quote (1,315 events). Soccer draws (-1.4c) and favourites (+0.2c) were flat.
Pre-game prices in sports markets are biased: underdogs are overpriced and favourites underpriced, soccer draws are neglected, and outcomes priced at 90c or more are worth more than their price.
Why people believe it
People argue that casual bettors back teams they like and long shots with large payouts, that nobody wants to back a draw, and that sure things offer too little profit per dollar locked to attract capital. If so, favourites, draws and near-certain outcomes would win more often than their prices say.
Why it would persist The argument, as stated before testing
Each edge is small per share and the markets are many and short-lived, so the argument runs that no one profitable specialises in them and the bias is never competed away.
For each pre-game market we read the price at fixed times before the scheduled start, built only from the trades that had already happened. Then we bought one share under a fixed rule and held it to the result. Rules were written down before any price was looked at: buy the favourite when its price is 60c to 90c, buy the soccer draw when its price is 10c to 45c, and buy any outcome priced 90c to 98.5c. The first esports sample looked good, so we found a flaw in how we had priced the entry, re-priced all of it, and repeated the rule on esports events from earlier months that the first sample did not contain.
At least 150 independent events; net mean above 0 in each of selection, validation and holdout; pooled event-clustered 95% interval lower bound above 0; pooled net mean at least 1c per share; still above 0 with the +0.5c stress. Esports replication, fixed after the first sample was seen and before replication data were read: net mean above 0 with event-clustered t at least 1.65, above 0 in each third, and above 0 with the stress.
Any of the pass conditions fails. Fewer than 150 events counts as underpowered.
Taker fee 0.05 x price x (1 - price) per share on sports, esports and weather; stress case +0.5c per share.
The price-versus-outcome calibration of each bucket is its own control. The first-fill and all-eligible versions are compared on the same events. An out-of-period esports sample with the rule frozen is the replication. Standard errors are clustered by event (2,000-draw bootstrap).
The esports favourite edge failed its out-of-period replication on all three pre-registered replication conditions (the original-fill version: t -0.6 against 1.65 required, one of three thirds positive, -1.5c under the +0.5c stress), and the soccer draw, soccer favourite and near-certain rules failed the pre-registered interval and 1c requirements.
- First sample, esports favourites (mid 0.60 to 0.90 at 1h before the start), priced at the first fill after the decision, which could not have been known in advance: +4.5c per share net (547 events, 95% interval +0.9 to +7.9c, t 2.49), positive in selection (+5.0c), validation (+2.4c) and holdout (+5.6c), and +4.0c with the +0.5c stress. This was an apparent pass of the pre-registered criteria, but it rested on that flawed fill.
- The fill was the flaw. Favourites that had a taker buyer in the 10 minutes after the decision won 77.9% of the time at a mid price of 0.718; favourites without one won 72.0% at a mid of 0.732. The later buyer is informed, so pricing the entry at that fill uses the future.
- Re-priced at the last pre-decision ask for every eligible favourite, the same rule was +1.3c (979 events, interval -1.4 to +3.9c, t 0.97).
- Out-of-period replication on 3,180 events with the rule frozen: -1.0c with the original fill (755 events, interval -4.3 to +2.2c, t -0.6; selection +0.6c, validation -3.1c, holdout -3.4c) and -0.5c all-eligible (1,315 events, interval -2.8 to +1.7c, t -0.4). Every replication condition failed. The headline statistics in this record are the all-eligible set (-0.5c, 1,315 events); the median of +20.3c sits above a negative mean because most bets win a small amount and a minority lose almost the whole price.
- At 6h before the start, all-eligible favourites were +4.5c in the first sample (614 events, t 2.75) and -1.1c in the replication (855 events, t -0.7). Buying the underdog at 6h lost 7.5c in the first sample (t -4.5) and 1.2c in the replication (t -0.8).
- The favourite's mid moved -0.15c from 6h to 1h before the start in the replication (t -1.1, 851 events). Buying at the 6h ask and selling at the 1h bid lost 2.8c per share, mostly spread and two fees.
- Soccer, all 3,562 eligible events (the full set was used): buying the draw at 0.10 to 0.45 returned -0.6c with the first-fill rule (275 events, t -0.2) and -1.4c all-eligible (938 events, interval -4.0 to +1.4c, t -1.0). The 0.20 to 0.35 draw bucket was priced within +0.3c of its outcome rate (standard error 1.6c).
- Soccer favourites at 0.60 to 0.90: -1.7c with the first fill (355 events, t -0.7) and +0.2c all-eligible (577 events, interval -3.2 to +3.5c, t 0.1). The favourite's mid drifted +0.38c from 6h to 1h (t 2.8, 355 events), far below the cost of trading it: a round trip lost 2.5c.
- Near-certain outcomes (mid 0.90 to 0.985, bought at the pre-decision ask; soccer and esports pool the 6h and 1h decision times, weather pools three fixed times of day): esports +0.9c (87 tokens), esports replication +1.8c (153), soccer -0.4c (330), two-way sports -4.2c (53) and weather -3.1c (121 tokens in 59 events). Each interval includes zero. Loss rates were 4.6% to 9.4%, and none of these tokens went to a multi-round dispute.
The esports favourite edge disappears twice
Net result per share of buying the esports favourite 1h before the start, with event-clustered 95% intervals. The first-fill version in the first sample (+4.5c) shrinks to +1.3c when every eligible favourite is priced at the last pre-decision ask, and both versions are at or below zero on the untouched replication sample.
Chart data (cents per share)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| First sample, first post-decision fill (n 547) | 4.5 | 0.9 | 7.9 |
| First sample, all eligible at pre-decision ask (n 979) | 1.3 | -1.4 | 3.9 |
| Replication, first post-decision fill (n 755) | -1 | -4.3 | 2.2 |
| Replication, all eligible at pre-decision ask (n 1,315) | -0.5 | -2.8 | 1.7 |
A fill that knows a buyer is coming
Mid price at decisionShare that won
Esports favourites in the first sample, split by whether a taker bought them in the 10 minutes after the decision. Their mid price at the decision was lower (71.8 against 73.2), yet they won 77.9% against 72.0%: the later buyer is informed, so entering at that fill adds profit that no real order placed at the decision time could earn.
Chart data (%)
| Mid price at decision | Share that won | |
|---|---|---|
| Taker buyer within 10 min after decision | 71.8 | 77.9 |
| No taker buyer | 73.2 | 72 |
Near-certain outcomes show no edge, and cannot show a small one
Net result per share of buying any outcome priced 90c to 98.5c at the last pre-decision ask and holding to resolution (soccer and esports pool the 6h and 1h decision times, weather three fixed times of day), with event-clustered 95% intervals. Every interval includes zero, and the two-way sports and weather samples are too small to rule out losses of 8c or more.
Chart data (cents per share)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Esports (87 tokens) | 0.9 | -5.6 | 5.6 |
| Esports replication (153 tokens) | 1.8 | -2.9 | 5.6 |
| Soccer (330 tokens) | -0.4 | -3.7 | 2.4 |
| Two-way sports (53 tokens) | -4.2 | -15.3 | 3.9 |
| Weather (121 tokens) | -3.1 | -8.6 | 1.1 |
Why it failed Interpretation · not established
The first-sample esports gain fits a market that is close to calibrated (a mild first-sample bias of -4.2c in the 0.2 to 0.35 bucket, standard error 1.9c, did not replicate) plus two artefacts: entering at the first later buyer selects trades that the market moved toward, and a sample of 547 events gives a standard error of about 1.8c, so a t of 2.5 is reachable by chance among the several rules and times tried. Prices one hour before the start sat within about 4c of outcome frequencies in every bucket of the first sample, with a median spread of 1c, so the spread and fee take what a small bias would leave.
Lesson
In a historical test of an order book, the price you assume for entry must be knowable at the decision time. Pricing at the first trade after the decision conditions on later informed flow and turns a calibrated market into a profitable one. The cheapest protections were to price every eligible item at the last pre-decision quote and to freeze the rule and rerun it on events the discovery sample never contained.
Scope of this result
Covers pre-game sports and esports moneylines and weather markets on one on-chain prediction market, decision times 1h and 6h before the scheduled start (weather: three fixed times of day), end dates 2026-04 to 2026-10, sampled every second or third day. It covers taker entries held to resolution and does not cover live in-play prices or other venues. The two-way sports (527 of 975 events had a trade near the decision time) and the near-certain slices are underpowered.
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 5Finding 6Finding 7Finding 8Finding 93. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade