betaider.com

Track record

Basketball: what the model predicted, against what actually happened

What you are looking at
Two records, answering different questions. Historical replays every prediction the walk-forward backtest made — the model only ever saw matches played before the one it was predicting, so these are genuine forecasts, and there are enough of them to mean something. Live is picks this app actually published, settled after the fact. It starts empty and grows.

Historical uses the raw model probability rather than the calibrated one, because calibration is itself derived from this history — scoring against it would be marking its own homework.

Historical — out of sample

Selections settled
6,023,971
Model said
49.4%
Actually won
49.4%
Calibration gap
-0.03pp

A gap near zero means the model's probabilities mean what they say: things it calls 70% happen about 70% of the time. That is the single most important number on this site.

Calibration by confidence

Model saidSelectionsPredicted ActualGapPredicted vs actual
0.05–0.10 236,309 7.3% 8.3% +0.93
0.10–0.15 372,063 12.7% 13.8% +1.17
0.15–0.20 298,238 17.3% 18.2% +0.85
0.20–0.25 337,427 22.5% 24.0% +1.45
0.25–0.30 282,993 27.6% 28.5% +0.94
0.30–0.35 354,501 32.5% 33.2% +0.74
0.35–0.40 340,498 37.6% 38.3% +0.73
0.40–0.45 296,186 42.6% 42.7% +0.01
0.45–0.50 583,220 47.8% 47.8% -0.04
0.50–0.55 580,420 52.2% 52.1% -0.03
0.55–0.60 290,951 57.4% 57.3% -0.09
0.60–0.65 331,146 62.4% 61.7% -0.76
0.65–0.70 329,353 67.6% 66.7% -0.85
0.70–0.75 260,679 72.4% 71.4% -1.01
0.75–0.80 326,686 77.5% 76.1% -1.43
0.80–0.85 204,147 82.6% 81.3% -1.31
0.85–0.90 294,110 87.4% 85.8% -1.55
0.90–0.95 138,587 92.4% 90.9% -1.47
0.95–1.00 166,457 98.1% 97.3% -0.76
predicted actual

By market

Market familySelectionsPredicted ActualGap
Handicap 1,180,987 56.3% 56.2% -0.10
Team total 978,472 50.0% 50.0% -0.01
Total points 666,898 50.0% 50.0% -0.00
H1 total 651,948 50.0% 50.0% +0.00
H2 total 651,948 50.0% 50.0% +0.00
Q3 total 237,852 50.0% 50.0% +0.00
Q2 total 237,852 50.0% 50.0% +0.00
Q4 total 237,852 50.0% 50.0% +0.00
Q1 total 237,852 50.0% 50.0% +0.00
Winning margin 156,105 13.1% 12.8% -0.26
Odd/Even 133,470 50.0% 50.0% +0.00
HT/FT 79,355 26.7% 26.6% -0.07
Highest scoring half 62,552 50.0% 50.0% +0.00
H1 result 62,175 48.3% 48.3% +0.00
H2 result 62,056 48.5% 48.4% -0.13
Highest scoring quarter 52,856 25.0% 25.0% +0.00
Half time 44,158 48.4% 48.4% +0.00
Moneyline 42,371 52.4% 52.4% -0.02
Overtime 38,100 57.8% 57.9% +0.06
Win both halves 37,455 28.7% 28.9% +0.20
Century marks 37,129 52.9% 52.1% -0.89
Q1 result 33,962 38.1% 38.2% +0.06
Q2 result 33,873 38.2% 38.3% +0.03
Q3 result 33,430 38.7% 38.7% +0.04
Q4 result 33,263 38.8% 38.7% -0.12

A family with a large negative gap is one the model oversells. The radar's calibration already corrects for this before ranking, but it is worth knowing which markets to trust least.

By league

CompetitionSelectionsPredicted ActualGap
NCAAW results-only 2,486,188 49.5% 49.5% -0.02
NCAAB results-only 2,193,748 49.8% 49.8% -0.04
NBA results-only 1,223,392 48.6% 48.6% -0.05
WNBA results-only 120,643 48.0% 48.0% -0.02

Live record

Model
ai-blend-live
Settled
5,128
Model said
35.7%
Actually won
36.3%
Model
dc-xg-live
Settled
5,264
Model said
35.5%
Actually won
36.2%
Model
deep-live
Settled
1,052
Model said
36.5%
Actually won
36.2%

Profit at fair odds

Flat 1 unit per selection
+93299
Per selection
+0.0155

This is not a profit claim. Every bet is settled at the model's own fair odds, where a perfectly calibrated model scores exactly zero by construction. It is a second calibration measure, and it is more sensitive than the headline gap because a longshot priced at 7% pays about 13× — so small errors on unlikely outcomes cost far more than the same error on a favourite. A negative figure alongside a near-zero overall gap points to the familiar favourite–longshot pattern: the model slightly oversells long odds and slightly undersells short ones. Real bookmaker prices are shorter than fair odds, so actual betting would do worse than this.

betaider.com

Sport

Basketball sections

Preferences

Appearance
Aider Reading the Basketball model · online
Open the Radar
Bet slip
Open full page
to move to open esc to close See all results

On this page

Track record

What this page is for

The scoreboard. Not what the model thinks it can do, but what happened when it was asked in advance and the matches were then played.

How to read it

Everything here is out of sample: each prediction was made before the match it describes. Break the record down by market and league before drawing a conclusion from the headline — a model can be genuinely good at one league and useless at another, and an average over both hides it.

Things to watch

Sample size is the trap. A market with thirty settled picks tells you almost nothing; a run of good results over a few weeks is well within what luck produces. Football is low-scoring enough that the better side loses often, and a fair chunk of every result is simply not predictable.

Where to go next

Profit at fair odds, at the foot of the page, is the harshest test on the site: it asks whether the edge survives being paid for.

Press esc to close Read the full guide