Can You Beat Vegas Long-Term? The Scoreboard Through September 7
BBMI has logged its numbers before tipoff, first pitch or kickoff, then left every result on the board. This series asks the only question that matters after the screenshots and hot streaks are stripped away: can a clearly defined set of picks beat the price over time? Here is the current score across five sports.
Every serious bettor eventually reaches the same wall: the closing line is hard to beat. As we wrote in June, a study of roughly 400,000 soccer matches found that observed win rates across Pinnacle's closing-odds bands tracked the implied probabilities with an R-squared of 0.997. That does not mean the favorite won 99.7% of the time, or that a line contains 99.7% of all available information. It means the closing prices were exceptionally well calibrated across a huge sample. That is still a formidable opponent.
Beating that opponent for a Saturday is common. Doing it for a season is not. Doing it for years, at prices a bettor could actually take, is the claim this series will test in public. A losing result stays in the article. A rule change gets dated. Future updates will use the same scoreboard.
The hypothesis
The test has two separate claims. First: can a model built from public statistics match the market's forecast accuracy? Second: can a rule defined before the games identify picks that beat the available price? A model can lose the first comparison and still win the second on a narrow slice of games. It can also call plenty of winners and lose money by paying too much for them. The tables below keep those questions apart.
The point is not to out-forecast the market on every game. It is to find a repeatable spot where BBMI and the market disagree, publish the rule, and see whether that slice survives the vig. Most games should produce no pick. If the boundary is drawn after the winners are known, the record is marketing, not evidence.
That disagreement is what this article calls edge: the gap between BBMI's number and the market's. It is measured in points on a football or basketball line, in runs on a baseball line, and in percentage points of win probability on a moneyline. Every pick rule below is a minimum edge, sometimes with a cap above it.
That test has to cover the whole board, not just the best headline. NCAA Basketball publishes spreads and totals. College baseball has spreads, totals and an underdog moneyline product. MLB's production board currently centers on first-five spreads and full-game run lines, alongside live first-five and full-game moneyline products with published records. College football currently publishes spreads, with a totals model still in development, while the NFL test covers both spreads and totals. MLB full-game totals and WNBA spreads and totals are earlier-stage calibration work: those models are still being adjusted and paper-tracked, so they are disclosed but not counted in the production-product standing.
A quick word on the price, since it decides everything below. At a standard −110 line you risk $110 to win $100, so you need to win 52.4% of your bets just to break even. That is the vig. Push the price to −150 and break-even climbs to 60.0%. At −184, the median price on BBMI's MLB run-line picks, it is 64.8%. A bettor who wins 55% of the time at −110 is roughly what a .300 hitter is in a .250 league: rare, valuable, and over any single month indistinguishable from a hot streak. The whole job is proving the .300 is real.
Every live pick is logged before the game and never removed. The article also separates live records from walk-forward replays, where a past season is processed chronologically and the model may use only information available before each game. A replay is more useful than an ordinary backtest, but it is not a bettable record. If a product threshold was chosen after reviewing that replay, the article says so.
Check the receipts. The headline is only useful if the underlying record is easy to inspect. BBMI keeps the game-by-game logs public, including the losses.
NCAA Basketball Live · 2025–26
Basketball has BBMI's longest live record and its strongest result. It also supplied the season's best reminder that even a strong ledger can feel awful for three straight weeks. The forecast-comparison sample covers 3,348 games from December 29 through April 6. The public picks log includes 2,012 spread picks and 1,279 totals. The denominator differs because the accuracy table covers every forecast in its comparison file, while the picks log contains only games that cleared a published betting threshold.
| Metric | BBMI | Vegas | Games | Read |
|---|---|---|---|---|
| Winner accuracy | 70.4% | 68.0% | 2,974 | BBMI ahead by 2.4 points on the same games |
| Brier score | 0.196 | 0.199 | 2,138 | Both beat a constant 50% forecast (0.250); BBMI lower by 0.003 |
| Margin MAE (points) | 8.80 | 8.83 | 2,975 | A dead heat |
| Total MAE (points) | 13.14 | 13.37 | 2,627 | BBMI closer by about a quarter point per game |
The basketball point estimates favor BBMI on winners, Brier score and totals. The winner gap is the clearest. The margin errors are effectively tied, and these summary rows do not by themselves prove that the smaller differences are statistically meaningful.
Claim two is where the money is, so here is the whole ledger, not just the good part. Spread picks fire at 2 or more points of disagreement with the market, with the premium tier beginning at 6. Totals fire at 4 or more points.
The tiers describe confidence, not just the size of the disagreement. A premium pick is one where BBMI believes the model's edge is strongest. A free, or marginal, pick is one where an edge appears to exist but the model is less confident that edge holds up long-term. The same labels carry the same meaning in every sport below.
| Product · edge | Tier | Record | ROI | Win % | 95% CI |
|---|---|---|---|---|---|
| Spread · 2 to 4 | Free | 561–465 | +4.4% | 54.7% | 51.6–57.7 |
| Spread · 4 to 6 | Free | 330–261 | +6.6% | 55.8% | 51.8–59.8 |
| Spread · 6 to 9 | Premium | 194–115 | +19.8% | 62.8% | 57.3–68.0 |
| Spread · 9 and up | Premium | 63–23 | +39.9% | 73.3% | 63.1–81.5 |
| Spread · all premium (6+) | 257–138 | +24.2% | 65.1% | 60.2–69.6 | |
| Spread · all picks (2+) | 1,148–864 | +9.0% | 57.1% | 54.9–59.2 | |
| Totals · all picks (4+) | Premium | 795–484 | +18.7% | 62.2% | 59.5–64.8 |
Then came March. The premium tier went 133–61 in January and 68–36 in February, but only 55–41 in March. From the opening round on March 17 through April 5, it went 20–20. Illinois failed to cover against UConn in a 71–62 semifinal loss, and Arizona missed against Michigan in a 91–73 defeat. Both were BBMI sides, and both lost. Forty picks at a true 65% win rate finish 20–20 or worse about 4% of the time, so the slump was unusual without being impossible. It was also the stretch a subscriber would remember. The national championship was in the log too. Michigan won 69–63, but the BBMI pick was Michigan −7.5, so it graded as a loss in the all-spreads record. It was not a premium pick. BBMI's tournament handling is being reworked for 2026–27. The 395-pick premium record matters more than those 40 picks, but the 40 still belong in the story.
Spreads are only half the basketball story. The totals product went 795–484 (62.2%) at 4 or more points of edge, returning 18.7% at −110. The same rule went 285–218 (56.7%) in the 2024–25 walk-forward. Both cleared break-even, and the live season was considerably stronger. That does not make every over or under automatic. It does make totals one of BBMI's two largest profitable live records, alongside the full basketball spread board.
College Baseball Live · 2026
BBMI priced 5,048 games from February 14 through June 21. The market posted a comparable line on fewer than half of the games BBMI covered.
| Metric | BBMI | Vegas | Games | Read |
|---|---|---|---|---|
| Winner accuracy, all games BBMI priced | 63.9% | n/a | 5,042 | Includes 2,717 games with no market line; BBMI called 66.4% of those |
| Winner accuracy, games both priced | 61.0% | 62.8% | 2,325 | Vegas ahead by 1.8 points on the same games |
| Brier score, games both priced | 0.230 | 0.221 | 2,328 | Vegas sharper; the site's 0.223 covers all 5,048 games |
| Margin MAE (runs) | 4.61 | 4.45 | 2,190 | Vegas closer by a sixth of a run |
| Total MAE (runs) | 4.61 | 4.46 | 2,201 | Vegas closer by a sixth of a run |
On the games both sides priced, the market won this comparison. It picked winners a little more often, posted the lower Brier score and missed margins and totals by less. BBMI was competitive, but it was not ahead.
Which makes the pick record more interesting, not less, because it shows the two claims are separable. You do not have to out-predict the market on every game to make money on the games where you disagree with it the most.
| Edge (runs) | Tier | Record | ROI at −110 | Win % | 95% CI |
|---|---|---|---|---|---|
| All graded spreads | Benchmark | 964–830 | +2.6% | 53.7% | 51.4–56.0 |
| Under 2 | Not a pick under the current rule | 341–363 | −7.5% | 48.4% | 44.8–52.1 |
| 2 to 4 | Pick all season | 457–361 | +6.7% | 55.9% | 52.4–59.2 |
| 4 to 6 | Pick only from May 30 | 165–104 | +17.2% | 61.3% | 55.4–67.0 |
| Over 6 | Excluded | 8–9 | −10.1% | 47.1% | 26.2–69.0 |
| 2 to 6 | Current rule, full season | 622–465 | +9.3% | 57.2% | 54.3–60.1 |
The dates matter here, so they go on the record. The spread thresholds moved three times during the season: the cap on edge was cut from 5 to 4 runs on April 9, the free floor and premium line were lowered on April 14, and the cap was raised from 4 to 6 on May 30 after a review of how the 4-to-6 games had done. The 2-to-6 rule this article reports is the one in force at season's end, applied back over the whole season. That makes the 622–465 record live results under a rule finalized in-season, not a pre-registered test, and the article treats it that way in the standing table below.
Under the 2-to-6 definition the product stayed in a fairly tight range for four months: 57.1% in February, 55.4% in March, 59.7% in April and 57.6% in May. Omaha was rougher. Six picks all lost, including Oklahoma +1.5 in Game 2 of the championship series. North Carolina won that game 6–2, then Oklahoma answered with a 13–2 title-clinching win the next night. Six losses can spoil a weekend. They do not outweigh nearly 1,100 decisions.
College baseball was not a one-market season. The totals board went 108–65 at 4 or more runs of edge, and the underdog moneyline board finished 176–241. One won nearly two-thirds of its picks at −110. The other lost most of its picks and still returned 25.0%, because the average winner paid +250. The full moneyline stream, favorites included, returned 9.5% over 838 picks; the favorites half of it lost 5.8%.
| Product | Rule | Record | ROI | Win % | 95% CI |
|---|---|---|---|---|---|
| Spread | 2 to 6 runs of edge | 622–465 | +9.3% | 57.2% | 54.3–60.1 |
| Totals · Premium | 4+ runs of edge | 108–65 | +19.2% | 62.4% | 55.0–69.3 |
| Totals · full stream | 1.5+ runs of edge | 577–480 | +4.2% | 54.6% | 51.6–57.6 |
| Underdog moneyline | 1+ percentage point of edge | 176–241 | +25.0% | 42.2% | 37.6–47.0 |
MLB Live · since May
MLB is BBMI's newest model and its hardest market. It is also where the difference between picking winners and beating a price is impossible to miss. The per-plate-appearance simulation went live on May 23 and had priced 1,391 completed games through September 7. The first-five-inning spread tracker began May 15.
| Window | Games | BBMI winners | Vegas winners | BBMI Brier | Vegas Brier | BBMI total MAE | Vegas total MAE |
|---|---|---|---|---|---|---|---|
| Full sim era | 1,391 | 55.1% | 56.4% | 0.249 | 0.244 | 3.66 | 3.51 |
| May 23–31 | 123 | 49.6% | 59.3% | 0.264 | 0.239 | 3.65 | 3.42 |
| June | 385 | 53.8% | 54.8% | 0.254 | 0.249 | 3.57 | 3.42 |
| July | 371 | 54.4% | 54.7% | 0.246 | 0.247 | 3.91 | 3.75 |
| August | 416 | 57.5% | 58.9% | 0.243 | 0.237 | 3.54 | 3.39 |
| Sep 1–7 | 96 | 60.4% | 54.2% | 0.250 | 0.244 | 3.59 | 3.56 |
The market remains ahead over the full MLB sample. The winner gap fell from 9.7 percentage points in the first nine days to 1.4 in August, and BBMI moved ahead for the first week of September. Ninety-six games are nowhere near enough to call that a turn. Since August 1, both sides picked winners at 58.0% over 512 games, while Vegas retained the better Brier score, 0.239 to 0.244, and the lower total error, 3.42 to 3.55 runs. One small exception belongs on the record: BBMI's July Brier was 0.246 against 0.247 for Vegas.
Claim two in MLB is a story about prices, one BBMI covered in July under the title Sixty percent of the time, it loses every time. It is still true.
| Product | Status | Record | ROI | Win % | 95% CI | Median price | Break-even |
|---|---|---|---|---|---|---|---|
| Full-game run line, +1.5, all picks | Live | 205–121 | −1.4% | 62.9% | 57.5–67.9 | −184 | 64.8% |
| First-5 spread, full tracked stream | Live | 125–98–1 | +7.0% | 56.1% | 49.5–62.4 | −110 | 52.4% |
| First-5 moneyline, home side | Live | 103–67–28 | +1.2% | 60.6% | 53.1–67.6 | −152 | 60.3% |
| Full-game moneyline | Live | 76–49 | −0.9% | 60.8% | 52.0–68.9 | −158 | 61.2% |
Read the run-line row twice. Winning 62.9% of 326 decisions looks strong until the price enters the box score. The typical pick cost −184, a price that needs 64.8% to break even. Because prices varied from game to game, the portfolio's exact verdict is its return: −1.4%. June made 10.2%; July lost 5.4%; August lost 3.1%. That is a season spent close to the rail. The two moneyline trackers tell a similar story, with win rates around 61% and returns within about a point of zero.
The first-five spread has been the better lane. It stops after five innings and usually costs −110, so 56.1% earns a profit instead of becoming an expensive way to be right. In baseball, the price can matter as much as the pick.
NCAA Football Walk-forward · 2025 Live · 2026, Week 1
College football has BBMI's strongest replay and almost no live history. Those facts cannot be blended. The model was replayed through the 2025 season using only information available before each kickoff. The current 6-point product rule, however, was finalized after that historical work was reviewed. Treat the replay as design evidence, not independent proof of the product. Live picks began August 29. Through September 7, the new ledger held 49 graded games and eight premium picks.
| Metric | Season | BBMI | Vegas | Games |
|---|---|---|---|---|
| Winner accuracy | 2025 walk-forward | 69.0% | 73.4% | 794 |
| Margin MAE (points) | 2025 walk-forward | 13.74 | 11.84 | 794 |
| Total MAE (points) | 2025 walk-forward | 12.90 | 12.68 | 725 |
| Winner accuracy | 2026 live | 81.6% | 85.7% | 49 |
| Margin MAE (points) | 2026 live | 14.15 | 13.77 | 49 |
Claim one, college football: the market wins, and not narrowly. Vegas called 4.4 more winners per hundred games in 2025 and missed the margin by nearly two fewer points. That is BBMI's widest accuracy gap across the five sports. It also sits next to BBMI's strongest replay record, which is the clearest example here of how forecast accuracy and betting results can diverge.
| Edge (points) | Season | Tier | Record | ROI at −110 | Win % | 95% CI |
|---|---|---|---|---|---|---|
| Under 5 | 2025 walk-forward | Not a pick | 113–116 | −5.8% | 49.3% | 42.9–55.8 |
| 5 to 6 | 2025 walk-forward | Not a pick | 54–60 | −9.5% | 47.4% | 38.4–56.5 |
| 6 and up, spread ≤ 23 | 2025 walk-forward | Premium | 270–157 | +20.8% | 63.2% | 58.6–67.7 |
| 6 and up, spread ≤ 14 | 2025 walk-forward | Premium ★ | 222–124 | +22.5% | 64.2% | 59.0–69.0 |
| 6 and up, spread ≤ 23 | 2026 live, Week 1 | Premium | 7–1 | +67.1% | 87.5% | 52.9–97.8 |
| All graded games | 2026 live, Week 1 | 27–22 | +5.2% | 55.1% | 41.3–68.1 |
How can a model miss the margin by almost two points more than the market and still post a strong replay against the spread? The two scores cover different sets of games. MAE averages every matchup. The pick record isolates the games where the numbers differ by at least six points and the market spread is no wider than 23. That slice covered 63.2% in the 2025 replay. Because the slice was studied while the product rule was being set, only the 2026 live log can tell us whether it travels.
The opening week gave the article some football. Tulsa was catching 14.5 against Oklahoma State and won outright, 24–10. Notre Dame laid 20.5 at Lambeau Field and pulled away from Wisconsin, 41–13. The lone miss was Troy −16.5: the Trojans won 24–21, nowhere close to covering. Add five other covers and the first live card finished 7–1. It could hardly have started better, and it still means very little. Seven wins are a headline. A full season is a test.
NFL Walk-forward · 2023–25 Live · starts 9/9
The NFL is the market everyone wants to beat, and BBMI's live test starts September 9. Every graded number below comes from a chronological replay of 816 games from 2023 through 2025. The current product bands were set after that history was available, so their records are candidates to test, not independent validation. The 2026 spread card was locked Tuesday morning, before the opener.
| Metric | BBMI | Vegas | Games | Read |
|---|---|---|---|---|
| Winner accuracy | 66.6% | 71.0% | 815 | Vegas ahead by 4.4 points; ahead in each of the three seasons |
| Margin MAE (points) | 10.00 | 9.30 | 816 | Vegas closer by 0.7 points per game |
| Total MAE (points) | 10.28 | 9.45 | 815 | Vegas closer by 0.8 points per game |
Claim one, NFL: the market is ahead, by about the same margin as college football. Claim two is where the NFL differs from every other sport here, because the model's clearest replay edge is on totals rather than spreads, and its spread record is the thinnest of any BBMI product.
| Market | Edge (points) | Record | ROI at −110 | Win % | 95% CI |
|---|---|---|---|---|---|
| Totals | Under 4 · not a pick | 279–297–6 | −7.5% | 48.4% | 44.4–52.5 |
| Totals | 4 to 6 · Premium | 67–55–1 | +4.9% | 54.9% | 46.1–63.5 |
| Totals | 6 and up · Premium ★ | 66–44 | +14.6% | 60.0% | 50.7–68.7 |
| Totals | All fired (4 and up) | 133–99–1 | +9.5% | 57.3% | 50.9–63.5 |
| Spreads | Under 1 · not a pick | 91–86–3 | −1.8% | 51.4% | 44.1–58.7 |
| Spreads | 1 to 4 · Premium | 225–195–15 | +2.3% | 53.6% | 48.8–58.3 |
| Spreads | 4 to 5 · Premium ★ | 42–35–1 | +4.2% | 54.5% | 43.5–65.2 |
| Spreads | All fired (1 to 5) | 267–230–16 | +2.6% | 53.7% | 49.3–58.1 |
| Spreads | Over 5 · not a pick | 56–66–1 | −12.3% | 45.9% | 37.3–54.7 |
The first NFL card has 16 spread picks, most within a point and a half of the market. Two disagreements are worth circling for Sunday: the Chargers are −9.5 against Arizona while BBMI makes the line −13.2, and Minnesota is −1.5 at home against Green Bay while BBMI makes the Packers a slight favorite. Seven games qualify for Premium under the current tiers; none reaches Premium ★. The totals product will wait for 2026 scoring data. For now, the NFL entry is a replay and a promise: the replay identified totals as the stronger candidate, and the live log will keep every result whether that idea survives or not.
The standing
Selection only matters if the selected games beat a fair benchmark in the same market. Exhibit 11 makes that comparison.
| Market | Window | Broad benchmark | Win % · ROI | Cohort | Win % · ROI | Share |
|---|---|---|---|---|---|---|
| NCAA Basketball spreads | Live 2025–26 | 1,449–1,158 | 55.6% · +6.2% | 257–138 | 65.1% · +24.2% | 15% |
| College Baseball spreads | Live 2026 | 964–830 | 53.7% · +2.6% | 622–465 | 57.2% · +9.3% | 61% |
| College Baseball totals | Live 2026 | 577–480 | 54.6% · +4.2% | 108–65 | 62.4% · +19.2% | 16% |
| NCAA Football spreads | 2025 replay | 437–333 | 56.8% · +8.4% | 270–157 | 63.2% · +20.8% | 55% |
| NFL totals | 2023–25 replay | 412–396 | 51.0% · −2.6% | 133–99 | 57.3% · +9.5% | 29% |
| NFL spreads | 2023–25 replay | 414–382 | 52.0% · −0.7% | 267–230 | 53.7% · +2.6% | 62% |
| Sport · cohort | Window | Forecast comparison | Record | ROI | Read |
|---|---|---|---|---|---|
| NCAA Basketball · premium spreads | Live 2025–26 | BBMI ahead, 70.4 vs 68.0 | 257–138 | +24.2% | Positive live ROI on 395 picks |
| NCAA Basketball · totals, edge ≥ 4 | Live 2025–26 | BBMI ahead, total MAE 13.14 vs 13.37 | 795–484 | +18.7% | Positive live ROI on 1,279 picks |
| College Baseball · spreads 2 to 6 | Live 2026 | Market ahead, 62.8 vs 61.0 | 622–465 | +9.3% | Positive; rule finalized May 30, so partly post-selection |
| College Baseball · totals, edge ≥ 4 | Live 2026 | Market ahead, total MAE 4.46 vs 4.61 | 108–65 | +19.2% | Positive live ROI on 173 decisions |
| College Baseball · underdog moneyline | Live 2026 | Market ahead, Brier 0.221 vs 0.230 | 176–241 | +25.0% | Highest live ROI on the board, at an average price of +250 |
| MLB · run line +1.5 | Live since May 23 | Market ahead, 56.4 vs 55.1 | 205–121 | −1.4% | Negative after paying a median −184 |
| MLB · first-five spread | Live since May 15 | Same model as above | 125–98–1 | +7.0% | Positive live ROI on 224 picks at −110 |
| MLB · first-five moneyline, home side | Live since July 30 | Market ahead, Brier 0.244 vs 0.249 | 103–67–28 | +1.2% | Positive by a point after paying a median −152 |
| MLB · full-game moneyline | Live since August 4 | Same model as above | 76–49 | −0.9% | Negative after paying a median −158 |
| NCAA Football · premium spreads | 2025 replay | Market ahead, 73.4 vs 69.0 | 270–157 | +20.8% | Positive replay; current rule reviewed after the fact |
| NFL · totals, edge ≥ 4 | 2023–25 replay | Market ahead, 71.0 vs 66.6 | 133–99–1 | +9.5% | Positive replay; nothing live |
| NFL · spreads, edge 1 to 5 | 2023–25 replay | Same model as above | 267–230–16 | +2.6% | Inconclusive replay; nothing live |
So, can you beat Vegas long-term? Not enough time has passed to say yes. BBMI trails the market's winner-accuracy point estimate in four of the five sports. The useful news is narrower. College basketball supplied two large positive live records: spreads returned 9.0% over 2,012 decisions, and totals returned 18.7% over 1,279. The premium spread cohort was stronger still, and its interval sits well above zero under a rule that did not move. College baseball also finished positive across spreads, totals and underdog moneylines. Its 2-to-6 spread rule was widened on May 30 after results were reviewed, so that record is partly post-selection; the 2-to-4 band that was a pick all season sits right at break-even on its lower bound. The MLB first-five spread has started well, but 224 picks are still a short baseball summer. The full-game run line is the cautionary tale: 205 wins in 326 decisions, and still a 1.4% loss after paying the price.
The replays are leads, not trophies. College football's 2025 cohort produced the strongest historical result, while the NFL totals record was positive but inconclusive. Now both have to face fresh games. That is the part worth watching: not whether Tulsa covered once as a two-touchdown underdog, but whether the same published rule is still standing in November.
What would change the answer
College football needs a full live season before its replay earns real confirmation. The MLB run-line rule should remain frozen while the site judges both its win rate and the prices actually paid. NFL totals and spreads need live records, period. College basketball needs another season from both spreads and totals. College baseball needs its spread, total and underdog moneyline rules carried forward without another look-back adjustment. One good season is evidence. Two begin to look like a record.
Follow the test, not just the headline. The live dashboard shows every graded pick and every miss. Premium unlocks the complete pregame board, but the historical record stays public so readers can judge the claims first.
Methodology and limitations
Scope. The ROI standing covers BBMI's current production products in NCAA Basketball, College Baseball, MLB, NCAA Football and the NFL. MLB full-game totals and WNBA spreads and totals remain calibration experiments while those models are still being adjusted and paper-tracked. NCAA Football totals and WIAA spreads are still in development. None of those calibration or development tracks is counted as a current product in Exhibit 12.
Source and cutoff. Figures come from the public pick logs behind the Model Performance pages, using those pages' grading rules. This snapshot was frozen after games on September 7, 2026. Picks are logged before the game and are not removed. Win percentages exclude pushes; records show them where they occurred.
Cohorts and rule timing. A cohort is one market plus a published minimum edge and, in some products, a cap or upper exclusion. NCAA Basketball, College Baseball and MLB rows use live results. College Baseball's spread thresholds changed on April 9 (cap 5 to 4 runs), April 14 (free floor 1.0 to 0.5, premium line 3 to 2) and May 30 (cap 4 to 6, premium line 2 to 4); the 2-to-6 cohort reported here is the season-end rule applied to the full season. The current NCAA Football outer spread cap was changed on August 30 and the NFL tiers were set on September 4, then applied to earlier graded history. Those cohort results are post-selection evidence.
Live, replay and calibration. Live means a pick was published before the game. Walk-forward means a past season was processed chronologically with information available before each game, then graded later; no reader could have bet the replay as presented. NCAA Basketball 2025–26, College Baseball 2026, MLB first-five spreads since May 15, the production MLB products since May 23 and NCAA Football since August 29 are live. NCAA Football 2025 and NFL 2023–25 are replays. The NCAA Basketball 2024–25 totals result is also a replay. Calibration products are published paper tests while the model and product rules are still being adjusted; they are not treated as established products or evidence of a durable edge.
Accuracy comparisons. Winner accuracy, Brier and MAE are computed on the games where both BBMI and the market posted the relevant number, so the two columns describe the same games. The exception is noted in the College Baseball section: the site's published 63.9% covers all 5,042 games the model priced, more than half of which had no market line. Vegas win probabilities are derived from the moneyline where captured; the NCAA Football and NFL files carry no market probability, so no market Brier is shown for them.
Prices and intervals. ROI uses flat one-unit stakes. NCAA Basketball, MLB and the College Baseball moneyline use stored pregame prices. College Baseball spreads and totals, NCAA Football and NFL assume −110 because the logs do not store a per-pick price for those markets. Win-rate intervals are Wilson 95% intervals. ROI intervals are percentile intervals from 4,000 resamples of individual pick profit. That bootstrap treats picks as independent, so games from the same slate or changing conditions may make uncertainty wider than shown. For a fixed −110 cohort, a Wilson win-rate interval can be converted to an ROI range, but it is not mathematically identical to a percentile bootstrap interval.
Selection and sample size. Looking across several sports, markets and thresholds creates multiple-comparison risk. A positive interval is useful descriptive evidence, not proof of a durable edge, especially when the rule was chosen after examining the same history. Records under 100 decisions are shown for transparency, not treated as settled results. That includes NCAA Football 2026 (8 decisions).
Inputs the models may miss. The models do not independently ingest every late injury, lineup or weather update beyond what is present in the captured data feed. The market may react to information the model does not contain. That is one plausible reason for poor results at extreme disagreements, but the performance record alone cannot establish a cause.