The Milwaukee Brewers finished the World Series in five games on 28 October, and with that Momus's first baseball season is closed. The headline is 20 wins, 10 losses on real money. The headline is also the least informative number in this article, so it gets one paragraph and then the rest gets taken apart.
The record
Between the September go-live and the last week of October, Momus took 30 MLB positions on Polymarket and wrote up 96 passes: 20–10, a 66.7% hit rate and +20.7% return on the stakes. Stakes ran from a third of a percent to 2% of bankroll, 1.2% on average. In bankroll terms: 37 points staked across the season, just under 8 points returned as profit. Every position is on the track record.
Now the cold water. At the prices he paid, a bettor with no edge at all wins about 17 of those 30. Twenty or more comes up roughly 18% of the time by luck alone. Thirty bets do not prove a model. They are consistent with one, which is a smaller claim and the only one the sample supports.
The bar was set in public before the first bet: positive closing-line value and a 60%+ win rate after twenty settled bets, or the engine goes back to the lab. At twenty settled the record stood at 13–7 (65%) with the closing line a point in his favour on average. Cleared — narrowly enough that nobody should frame it.
Where the profit came from
- Regular season: 22 bets, 13–9, +5.5%. Roughly break-even with a tailwind.
- Postseason: 8 bets, 7–1, +67.4%. Most of the season's profit, from a sample that fits on one hand and three fingers.
- By month placed: September 16–9 (+15.0%), October 4–1 (+57.5%).
- Home picks went 10–4, road picks 10–6.
One team accounts for a lot of it. Momus backed the Brewers seven times and they won six, including Games 1 and 3 of the World Series. He rated the eventual champions higher than the market did for six weeks, and the market kept paying him for it. He also backed the Mets three times and lost all three. A season this short is a handful of opinions about a handful of teams.
Thin edges lost. Real ones paid.
The cleanest split in the record is the size of the edge at entry:
- Under 2 points of edge: 11 bets, 5–6, −30.4%.
- 2 points and up: 19 bets, 15–4, +33.6%.
A one- or two-point gap between a model and a liquid market is inside the noise of both. The live floor sat at one point all season. Eleven bets is not enough to move it on, so the floor goes into the off-season as the first thing to test against the full paper record — but nobody reading this should expect it to survive.
Closing-line value
Win–loss says what happened. Closing-line value says whether the bet was good when it was made. Measured from the price quoted at the moment of decision to the last Polymarket price before first pitch, the 30 positions beat it by 0.95 points on average: 16 moved his way, 12 did not move, 2 moved against him.
The split behind that average is the most encouraging number of the season. Bets where the line moved toward Momus went 13–3. Bets where it stayed flat or went the other way went 7–7. When the market agreed with him after the fact, he won; when it shrugged, he flipped coins. That is what an edge is supposed to look like, and it is the test that matters most at small samples. One footnote on the accounting: results are booked at the order's limit price, the worst price he was willing to pay, not at the quote.
The model against the market, on every game
Bets are the tip. Underneath, the pitcher-aware model priced every game it could get a clean pre-game read on: 706 of them from late July into the postseason.
- Picking the winner: the model's side won 60.5% of the time, the market favourite 58.5%.
- Where they disagreed on who would win (134 games), the model was right in 74 — 55%.
- Probability quality (Brier score): model 0.2372, market 0.2372. A dead heat. The 60/40 blend of market and model that Momus actually bets from scored 0.2362, better than either parent.
- Postseason: 30 games, model 16 right, market 18. The model's advantage is a regular-season one; in October the market had the better of it.
Read that honestly: the model is not smarter than the market. It is about as good, and it is wrong in different places, which is the only reason blending the two is worth anything. A bet happens when the blend and the Polymarket price are far enough apart — not when the model is feeling confident.
What the paper season changed
The engine ran on paper before it ran on money, and kept a paper book alongside the real one all season: 160 candidates the rules flagged, 94–66, +15.8% at flat stakes. Three findings from it shaped the live engine.
- The biggest edges were the worst bets. Candidates with a gap of 12 points or more went 11–13 and lost money. Gaps of 4 to 12 points went 64–37. When the model disagrees with a sharp price by that much, the model is the one that is missing something. Those candidates are refused.
- Momus's own veto was costing money. Until 7 September he could wave off a candidate the numbers liked. Over the paper season the 27 he waved off went 16–11, +39.5% at flat stakes — better than the bets he approved. A strictness audit that day took the veto away for qualifying candidates and fixed a leak where candidates expired without ever being ruled on.
- When the engine looks matters. Early scans ran before lineups were posted: both starting pitchers were known on 41% of games. After the scan moved closer to first pitch in August, 94%. A pitcher-aware model without pitchers is a team model with extra steps.
The ceiling trial
The engine went live refusing any price above 2.00. The paper test had shown its largest claimed edges to be model error, so long prices were banned outright. By 23 September that ceiling had blocked 35 live candidates — and 17 of them had won, against a break-even rate of 43%. A rule written to avoid model error was throwing away the best prices on the board.
On 24 September the ceiling moved to 3.00 as a declared trial: above 2.30 the edge has to be at least 3 points, and every underdog is staked at half size. Day one produced two underdog bets: the Mets in Texas at 2.20 lost, the Brewers in Philadelphia at 2.25 won 5–1. A fair introduction.
- Underdogs actually bet: 6 positions, 4–2, +61.6% on half stakes.
- All underdog candidates the engine priced since the trial began: 27 settled, 13–14, against a break-even rate of 43.6%. +5.7% at flat stakes.
- Favourites (2.00 and under), whole season: 24 bets, 16–8, +15.9%.
The trial was set to be judged at 50 settled underdog candidates. The season ended at 27. So there is no verdict: the numbers lean the right way and prove nothing. The 3.00 ceiling, the half stakes and the counter all carry over to Opening Day, and the trial reports when it reaches 50.
The lesson October taught: starters to be announced
After a bet on 3 October, Momus did not place another until the 16th. The Division Series and the first half of the League Championship Series went by without him. Not for lack of opinions. For lack of a retry.
The engine makes one decision per market and that decision is final. In the regular season that is discipline. In the postseason, managers name starters late, and a game scanned while a starter was still unannounced was written off as a pass for good — even when the pitcher was named two hours later and the gap to the market was five to nine points. Seven of the last fourteen postseason passes before the fix died that way.
The fix shipped on 13 October: a game with an unannounced starter is no longer decided, it is deferred, and the last run before first pitch rules on it with whatever is known. The bets came back three days later: two in the ALCS, then the Brewers in Games 1 and 3 of the World Series. The cost of the two weeks before is unknowable, which is the irritating thing about bets that were never placed.
Best calls
- Brewers at Phillies, 24 September. Priced at 2.25 (44.5%) against a fair 49.7%. Won 5–1, on the first day underdogs were allowed.
- Yankees at Twins, 15 September. Bought two days early at 56.5% against a fair 62.1%. The line moved five points his way before first pitch. Won 8–1.
- Yankees over Red Sox, Wild Card Game 2. Bought at 52.5%, closed at 55.5%. Won 9–2.
- Brewers at Rays, World Series Game 3. A road underdog at 48.5% against a fair 52.1%. Won 10–1.
Worst calls
- 16–17 September: four straight losses. Dodgers in Cincinnati at 65.5%, the joint-shortest price he paid all season, lost 6–2. Red Sox in Texas, Brewers in Pittsburgh and a Mets side shut out by the Phillies followed. Three of the four carried under two points of edge. See above.
- Guardians at Royals, 27 September. A full 2% stake, the maximum, on a 5.5-point edge. The line never moved toward him and Cleveland lost 3–2. The same day the Mets lost in Washington to complete their 0–3.
What carries into 2027
- The blend stays. Model and market tied on accuracy and the mix beat both. Nothing about the season argues for trusting the model more.
- The edge floor gets tested. Sub-two-point bets went 5–6. The off-season study decides whether the floor moves; the number will be published with the evidence.
- The ceiling trial resumes at 27 of 50, same rules.
- Late starters are deferred, not passed — from the first game, not from mid-October.
- Closing-line value stays the scoreboard. +0.95 points across 30 bets is a start. A few hundred is an answer.
Twenty and ten, one championship team read correctly, one rule that cost two weeks of October, and a model that is exactly as good as the market it bets against. The engine rests until Opening Day. The track record stays where it is.