How Accurate Are Sports Predictions? · Part 1 of 7
A 70% Prediction Isn’t Wrong When It Loses
One upset cannot tell you whether a forecast was any good. Thousands of games can. Here is how BBMI’s probabilities have held up in college basketball, the NFL and MLB.
BBMI Research · November 20, 2026
On Oct. 18, BBMI gave the Los Angeles Rams an 89% chance to beat Arizona at home. The Cardinals won, 18-16.
Was the forecast wrong?
89%BBMI’s forecast for the Rams on Oct. 18
11%Arizona’s chance: about one game in nine
18–16Final score. Arizona won.
Most people would say yes. The model picked a team and the team lost. But an 89% forecast still gives the underdog an 11% chance, about one game in nine. Across hundreds of forecasts at 89%, some of those underdogs should win. If none did, that would raise a different question: was the model too cautious?
A single game cannot tell us whether a probability was well calibrated. Only a large group of forecasts can. The question to ask is this: of all the times the model said about 70%, how often did that team win? If the answer is close to 70%, the model is calibrated. If those teams won 85% of the time, or 55%, the probabilities are off, no matter how many winners the model picked.
What the record shows
BBMI publishes this test on its own forecasts. The chart groups games by the probability BBMI gave the home team, then compares that forecast with how often the home team won.
In 14 of 15 probability bands, the forecast was within 5 points of the result
College basketball, 2025-26 (3,361 games)NFL, 2023 to 2026 (967 games)MLB, 2026 (1,695 games)
Home team won
20%30%40%50%60%70%80%90%20%30%40%50%60%70%80%90%
MLB, under 35% (80 games)forecast 30.0%, won 51.2%
On the dashed line, the forecast matched the result
BBMI forecast: home team win probability
Each dot is one probability band: all the games where BBMI’s forecast for the home team fell in that range. Source: BBMI Model Performance pages, Nov. 20, 2026. 6,023 games.
College basketball, 2025-26 (3,361 games). Every band is within 5 points. Home teams given 65% or more averaged an 82.3% forecast and won 78.7%. Home teams given under 35% averaged 21.3% and won 24.7%. Both ends missed toward 50%. The pattern suggests BBMI was modestly overconfident in its most lopsided forecasts.
NFL (967 games). Every band is within 2 points. Home teams given 65% or more averaged 75.3% and won 76.3%, which means 69 of those 291 favorites lost. The Rams were one of them. That band’s average forecast was 75.3%, so it does not test 89% forecasts on their own. Most of this sample comes from walk-forward tests of the 2023 to 2025 seasons, in which each week is forecast using only earlier data. The 2026 games are live forecasts.
MLB, 2026 (1,695 games). Three of the five bands, covering 1,447 games, are within about 1 point. The other two are further off. Home teams given 65% or more averaged a 70.2% forecast and won 65.5%. The largest gap is in 80 games where BBMI gave the home team under 35%: the average forecast was 30.0%, and those teams won 51.2%.
The numbers behind the chart
Forecast band
Games
Average forecast
Home team won
Gap
College basketball, 2025-26 (3,361 games)
Under 35%
671
21.3%
24.7%
+3.4
35 to 45%
291
40.2%
45.0%
+4.8
45 to 55%
354
50.0%
50.0%
0.0
55 to 65%
384
59.5%
56.0%
−3.5
65% and over
1,661
82.3%
78.7%
−3.6
NFL, 2023 to 2026 (967 games)
Under 35%
172
26.5%
25.0%
−1.5
35 to 45%
159
40.1%
38.4%
−1.8
45 to 55%
157
49.7%
48.4%
−1.2
55 to 65%
188
60.0%
59.6%
−0.4
65% and over
291
75.3%
76.3%
+1.0
MLB, 2026 (1,695 games)
Under 35%
80
30.0%
51.2%
+21.3
35 to 45%
277
40.8%
40.4%
−0.4
45 to 55%
634
50.0%
51.1%
+1.1
55 to 65%
536
59.3%
58.2%
−1.1
65% and over
168
70.2%
65.5%
−4.7
Forecast band is BBMI’s win probability for the home team. Gap is the actual win rate minus the average forecast, in percentage points, as published. Zero means the forecast matched the result.
One number summarizes each record. The Brier score is the average squared gap between the forecast and the result, counting a win as 1 and a loss as 0. A perfect forecaster scores 0. Saying 50% for every game scores 0.25. BBMI scores 0.192 in college basketball, 0.215 in the NFL and 0.247 in MLB. Those three numbers do not rank the models. About half of the college basketball games had a home team at 65% or better, and about one in ten MLB games did, so confident forecasts come easier in one sport than the other. Judging a model takes a benchmark from its own sport, which is the subject of Part 2.
Brier scores: lower is better
College basketball, 2025-26
0.192
NFL, 2023 to 2026
0.215
MLB, 2026
0.247
00.25
The Brier score is the average squared gap between the forecast and the result, counting a win as 1 and a loss as 0. A perfect score is 0. The dashed line marks 0.25, the score for saying 50% on every game. Scores are not comparable across sports.
Weather forecasters have been graded this way for decades
When the National Weather Service gives a 40% chance of rain, it means a 40% chance that at least 0.01 inch falls at any given point in the forecast area. One dry afternoon does not make that forecast wrong, and one wet one does not make it right.
So meteorologists count. A 1977 study of National Weather Service forecasts covered 17,514 precipitation forecasts issued in Chicago from July 1972 through June 1976. The forecasters said 30% on 1,574 occasions. It rained on 449 of them, or 28.5%.
Chicago weather forecasters, 1972 to 1976
30%Forecast chance of precipitation
28.5%How often it rained after that forecast
1,574 forecasts of 30%, with precipitation on 449. From 17,514 National Weather Service forecasts for Chicago, July 1972 through June 1976. Source: Murphy and Winkler (1977).
Actuaries know the method by another name: an actual-to-expected study. Group the cases by expected rate, then compare each group with what happened.
The takeaway
A good probability forecast is not one that always picks the winner. It is one where 70% means 70%. That can only be judged over many forecasts, and the judging has to include the places where the model misses.
Next in the series: can a sports model beat the market’s forecast?
Data: BBMI figures are from the Model Performance pages for college basketball, the NFL and MLB, read Nov. 20, 2026. The Rams forecast is from the BBMI Rams team page.