Diego Garcia

Back to logs

> var / tempo

The Model Senses Draws, Then Buries Them

KairoX Tempo assigns P(draw) above 0.25 to 60% of the draws it misses: the probability is there, and the decision boundary discards it.

Jun 20267 min read

This log uses EXP-006 (V5, ECE=0.0193), the best-calibrated KairoX Tempo artifact, on 1,140 OOF predictions from three La Liga validation seasons (23-24, 24-25, 25-26). The question is whether the draws the model gets right differ structurally from the ones it misses, or whether the draw class is simply unpredictable.

Sample size caveat

The model predicted draw as the most likely outcome only 31 times across 1,140 matches. Seven of those were correct. That leaves a true positive sample of n=7, too small for statistical claims on its own, so the feature comparison below is directional at best. The underpricing finding rests on a much larger base: 290 missed draws and 668 high-P(draw) matches.

How Rarely the Model Commits

The base draw rate in the dataset is 26.1%. The model predicts draw as the most likely outcome in 2.7% of matches. When it does, it is correct 22.6% of the time. That precision barely clears usefulness, and at 2.4% recall, betting draws with this artifact is not worth running.

CategoryCount% of OOF
Actual draws29726.1%
True positive draws (predicted D, actual D)70.6%
False negative draws (actual D, predicted H or A)29025.4%
False positive draws (predicted D, actual H or A)242.1%

Draw prediction breakdown across 1,140 OOF matches. Precision: 22.6%. Recall: 2.4%.

Draws are hard to predict, but that does not account for a 2.4% recall. The model declines to predict draws even where it puts reasonable probability on them.

Probability Without a Prediction

Among the 290 missed draws, the mean P(draw) assigned by the model is 0.266. For the 7 correctly predicted draws, it is 0.388. That gap is significant (Mann-Whitney p less than 0.001, Cohen's d=2.417). More telling is that 174 of the 290 missed draws, 60%, had P(draw) above 0.25: the model put meaningful probability on them and picked H or A anyway.

Model's draw probability: correctly predicted vs missed draws

EXP-006 (V5, v5_dominance_30), La Liga OOF, walk-forward CV.

0102030405060700.000.050.100.150.200.250.300.350.400.450.500.550.600.65Model's predicted P(draw)Count1/3 (uniform threshold)
TP draws (n=7)FN draws, missed (n=290)
Figure 1: P(draw) distribution for correctly predicted draws (green) vs missed draws (red). The missed draws have substantial mass above 0.25.

Looking at all 668 matches where the model assigned P(draw) above 0.25, the actual draw rate is 27.1%, close to the full dataset base rate of 26.1%, so the elevated draw probability is not predictive of draws in aggregate. Within that subset the model predicted a draw only 31 times; for the other 637 matches it assigned above-threshold draw probability and still predicted H or A.

Prediction outcomes when the model senses a draw (P(D) > 0.25)

EXP-006 (V5), n=668 matches.

Actual HActual DActual APred HPred DPred A200(29.9%)103(15.4%)97(14.5%)16(2.4%)7(1.0%)8(1.2%)59(8.8%)71(10.6%)107(16.0%)Actual OutcomePredicted Outcome
correct predictionn = 668
Figure 2: Prediction outcomes for the 668 matches where P(draw) exceeded 0.25. Of 181 actual draws in this subset, the model correctly identified 7. The remaining 174 were routed to H (103) or A (71).

What Distinguishes the Draws the Model Gets Right

Even with n=7, two features separate correctly predicted from missed draws on both significance and effect size.

FeatureTP medianFN medianp-valueCohen's d
Home xG_op at Home (roll5)0.4780.9260.00261.082
Away xG_op Away (roll5)0.5060.7920.03610.770
Passes Acc delta (roll10)-103.75-9.500.062 (ns)0.639
Points Gap-56.0-8.50.249 (ns)0.339
xG Open Play delta (roll10)-0.091-0.0770.706 (ns)0.106

Feature comparison: correctly predicted vs missed draws. Only venue-specific xG features clear p<0.05 and Cohen's d>0.4 simultaneously.

The two significant features are both venue-specific xG measures. Correctly predicted draws happen in matches where the home team's open-play xG at home (roll5) is low (median 0.478) and the away team's open-play xG away from home (roll5) is also low (median 0.506). Missed draws happen in matches where both figures are higher: 0.926 and 0.792 respectively.

One reading: the draws the model catches are tight matches where neither side creates much at its usual venue, and that shared low output is a parity the model can detect. The draws it misses come in higher-xG games where both teams create chances and the match ends level anyway, which looks more like finishing variance than structural balance.

That second type of draw is plausibly undetectable from pre-match features. No feature set will reliably predict that a match with xG 1.8 vs 1.6 ends 0-0 rather than 2-1. The question the data cannot yet answer is whether the first type, the structurally constrained draw, can be recovered with better feature representation or a different decision boundary.

An Open Question on the Decision Boundary

This points to a representation gap more than a signal gap. The model has draw probability in the right places, but cross-entropy, optimised across all three classes, does not penalise missed draws heavily enough to move the decision boundary. A draw accounts for 26% of outcomes but produces only modest loss when incorrectly classified as H or A, because H and A together form the majority. The model learns to avoid the draw prediction unless the evidence is overwhelming.

It is unclear whether this is fixable at the loss function level, at the threshold level, or not at all for the high-xG draw subset. A draw-aware loss that penalises FN draws more heavily might shift the boundary, but it could also introduce draw over-prediction that destroys the H and A ROI the model currently has. That tradeoff has not been tested.