Model accuracy · transparency

Sixty percent of the time, it loses every time

Half a season inside the BBMISports MLB spreads model: its two products, the First 5 Innings spread and the Full Game run line. The Full Game picks win almost 61% of the time. At the prices you actually get, they lose money. That contradiction, and what we did about it, is the most honest thing we can show you about betting a real model.

BBMI Research Desk · July 11, 2026 · 8 min read

The paradox 60.6% and −2.3% A Full Game run-line record most bettors would kill for, quietly underwater once the juice is real

The fix +13.0% The disciplined slice, First 5 Innings above the floor and Full Game high-conviction only, returns +13% at true prices

The tax 2-8, then 18-2 Even the winning slice runs through ten-game skids and twenty-game tears in the same half-season

Everything below is the BBMISports MLB spreads model, and only that model. It posts two things and nothing else: a First 5 Innings spread (who covers over the first five innings) and a Full Game spread, which in baseball is the run line, the underdog getting +1.5 runs or the favorite laying them. We just finished calibrating both products, and these are the results to date: roughly half a season of picks, priced at what the bets actually cost. Calibrating a brand new model works the way you’d hope it does: it makes “paper picks” along the way, graded and priced like real ones with no money behind them, and we analyze the wins and the losses and adjust the model to reflect its weaknesses. That process is where everything below comes from.

Part one: the trap

Here is a number that should bother you, because it bothered us: so far this season the model’s Full Game run-line picks have gone 100 and 65. That is a 60.6% win rate. It is the kind of record people put in advertisements. And at the prices those bets were actually available, it is a losing record, down 2.3% on every dollar risked.

The culprit is the juice,¹ and it is not a rounding error. A Full Game run-line pick is usually the underdog getting +1.5 runs, and the market knows that side wins often, so it charges for it. The median price on our run-line picks was −180: risk $180 to win $100. At −180 you have to win 64.3% of the time just to break even. We won 60.6%. Four points short of the toll, on the right side of the result and the wrong side of the ledger.

Fig. 1 · The win rate vs the toll

What the model won vs what the price required · red tick = break-even at the median price actually paid

Full Game run linemedian price −180 60.6% 3.7 points short → −2.3% on every dollar

First 5 Innings spreadpriced near −110, above the floor 60.6% 8.2 points clear → +15.7% at real prices

Same win rate. Opposite outcomes. Both products won 60.6% of their (kept) picks, but the run line’s price demanded 64.3% and the F5 spread’s only 52.4%. The price is the point.

This is the single most important thing to understand about a betting model, and the one most likely to be hidden from you: a win rate is not a return. Put both products together (the First 5 Innings spreads and the Full Game run lines, everything the model published) and the record is 172 and 127, a tidy 57.5%. Priced at the fiction of −110 that record looks like a 9–10% edge. Priced at what the games actually cost, it is −0.1%. Break-even. Half a season of being right more often than wrong, netting nothing, because being right was already priced in.

It gets sharper when you sort the picks by how loudly the model was talking. The lowest-confidence Full Game picks, the quiet ones, the marginal edges, went 44.6% and lost nearly 29 cents on the dollar. A small pile of picks where the model backed a favorite to lay the −1.5 went 1 and 8. Those were not the model being unlucky. Those were bets that never should have been graded as picks. Left in the pile, they turn a real skill into a break-even wash.

Part two: the fix, and the tax you pay for it

The fix is not a better model. It is refusing to make the bad bets the model was already telling us not to make. The model does not emit one undifferentiated opinion; it emits graded ones, and the grades are honest. The high-conviction Full Game picks, the ones it flags most strongly, went 68.3%, and at an average price of −163 that clears the juice with room to spare: +9.6%. The First 5 Innings spreads, once we raised the edge floor and threw out the thin sub-threshold band that had been bleeding at 47%, sit at 60.6% and, priced near −110, return +15.7%.

Fig. 2 · A win rate is not a return

Every slice of the season, at the prices actually paid · struck-through rows are the cuts

SliceRecordWin rateReturn

Everything publishedboth products, all grades172–12757.5%−0.1%

Full Game run line, all picksmedian −180100–6560.6%−2.3%

Lowest-confidence Full Gamethe quiet, marginal edges25–3144.6%−28.9%

Favorites laying −1.5never should have been picks1–811.1%−78.3%

High-conviction Full Gameavg −16369–3268.3%+9.6%

First 5 Innings, above the floornear −11040–2660.6%+15.7%

The disciplined bookhigh-conviction FG + above-floor F5 only115–6065.7%+13.0%

The discipline is the product. The winning book throws away more than a third of the bets that won, and it is the only version of this book that makes money.

Keep only those, the high-conviction Full Game run lines and the above-floor First 5 Innings spreads, and the combined record is 115 and 60, 65.7%, a +13.0% return at real prices. That is the actual edge. It is smaller than the raw win rate, it throws away more than a third of the bets that won, and it is the only version of this book that makes money. The discipline is the product. The rest is just being right in ways the market already charged you for.

And even that edge does not arrive gently. This is the second thing nobody tells you. A +13% half-season is not a smooth climb; it is a violent one. Inside this record there is a stretch that went 2 and 8 over ten picks, a six-game losing streak, and a twenty-game trough of 6 and 14, a run long enough and cold enough that every instinct says the model has broken. It has not. The same half-season holds a nine-game winning streak, a run of 18 and 2 over twenty picks, and stretches where it felt like free money. Month to month it barely wobbles: April 53%, May 54%, June 62%, July 56%. But week to week it lurches between those two extremes with no pattern you can trade on.

Fig. 3 · The same edge, two lenses

Inside one +13% half-season · the streaks that scream vs the months that shrug

The trough

2-8 · 6-14

A ten-pick skid, a six-game losing streak, and a twenty-game trough cold enough to feel like the model died. It didn’t.

The tear

18-2

The same half-season: a nine-game winning streak and a twenty-pick stretch that felt like free money. Same edge, same lens problem.

53%

April

54%

May

62%

June

56%

July

Month to month the edge barely wobbles. Week to week it lurches between the two panels above, with no pattern you can trade on.

That is the law of large numbers seen from the inside, and it is not comforting in the moment. Over ten picks the edge is invisible; the window swings from 20% to 90% on noise alone. Over a full sample it is the quiet, grinding truth. The 2-8 skid and the 18-2 tear are the same 65.7% edge, observed through a lens too small to show it. The only mistake is to read the cold stretch as news, to conclude the model died, and change what you’re doing right before the sample gets large enough to pay you.

The work is not finding a number that wins 70% of the time. It is being honest about which bets actually clear the vig.

So the two things nobody tells you, said plainly: winning more than half your bets can still lose you money, because the price is the point and the juice is real; and the edge that does survive the juice is thin, so the ride that delivers it is brutal. The work is not finding a number that wins 70% of the time. It is being honest about which bets actually clear the vig, disciplined enough to skip the ones that don’t, and still standing through the 2-8 stretch when the ones that do finally pay. That is the whole job of the BBMISports MLB spreads model, and it is the whole job of anyone betting it.

1The vig (short for vigorish; also called the juice) is the bookmaker’s built-in commission: the gap between the odds you’re offered and the true odds of the outcome. It’s why a coin-flip bet doesn’t pay even money: at a standard −110 line you risk $110 to win $100, so you have to win about 52.4% of the time just to break even. Push the price to −180 and that break-even climbs to 64.3%. The vig is how the house gets paid whether you win or lose. Clearing it, winning more often than the price requires, is the entire job.

Analytical and educational content about model performance. Results to date through roughly the midpoint of the 2026 season, immediately following a full recalibration of both products; the record keeps accumulating from here.