> var / blaze
The Bilateral ROI Hypothesis: Right Structure, Wrong Sample Size
The Bundesliga BTTS model finds positive NO edge in every validation fold, and the bootstrap confidence interval on its NO ROI is 77 percentage points wide. The edge may be real; three seasons of bets cannot show it.
A BTTS binary classifier has two prediction directions: YES (both teams score) and NO (at least one team fails to score). Most deployed models find edge on one side or neither. Positive ROI on both sides at once, called bilateral ROI here, needs the model to be well calibrated in both directions and the market to misprice both outcomes, which is a stronger claim than usual and needs stronger evidence.
The KairoX Blaze Bundesliga model (BU v3, V3 Poisson feature set, 200 Optuna trials) records +12.08% headline ROI on its most recent validation season. That figure is from one fold. The question is whether the bilateral signal holds across the full walk-forward history, and whether there is any statistical basis for deploying on it.
What the Three-Fold History Shows
Across three walk-forward validation seasons (23-24, 24-25, 25-26), bilateral ROI first appears at the 62.5% confidence threshold with 22 NO bets. At the 60% threshold, which is the one that matters for deployment, YES ROI is +3.32% and NO ROI is -3.23% across all folds combined. The folds disagree sharply:
| Fold | Season | N YES | ROI YES | N NO | ROI NO |
|---|---|---|---|---|---|
| Fold 0 | 23-24 | 85 | -0.53% | 23 | -36.13% |
| Fold 1 | 24-25 | 113 | -4.24% | 10 | +16.30% |
| Fold 2 | 25-26 | 85 | +17.21% | 19 | +26.32% |
Fold-level ROI at the 60% confidence threshold. Bilateral ROI holds only in fold 2. Fold 0 produced strongly negative NO ROI on 23 bets.
Fold 0 placed 23 high-confidence NO bets in 23-24 and lost 13 of them, for -36.1% ROI, which on its own pulls the three-fold NO aggregate negative. Fold 1 and fold 2 are both positive on NO, with fold 2 showing bilateral ROI at +17.2% YES and +26.3% NO simultaneously.
Fold 2 could easily be noise. What the model selected says more than what it returned.
Edge Per Bet Is the More Honest Signal
ROI on a small number of bets is dominated by variance. The mean edge per value bet is less noisy: the average difference between the model's predicted probability and the market's implied probability, measured only on bets where the model has positive expected value. It does not depend on which matches happened to end YES or NO, so it measures whether the model keeps finding matches where the market underprices NO.
| Fold | Season | N NO value bets | Mean NO edge | N YES value bets | Mean YES edge |
|---|---|---|---|---|---|
| Fold 0 | 23-24 | 71 | 0.033 ± 0.021 | 121 | 0.044 ± 0.033 |
| Fold 1 | 24-25 | 59 | 0.038 ± 0.029 | 150 | 0.040 ± 0.031 |
| Fold 2 | 25-26 | 109 | 0.045 ± 0.031 | 110 | 0.046 ± 0.037 |
Mean edge per value bet across all confidence levels. The NO edge is positive and increasing across all three folds, including fold 0 where realized NO ROI was -36%.
The NO edge is positive in every fold (3.3%, 3.8%, 4.5%) and trending upward. In fold 0, where realized NO ROI was -36.1%, the model was still identifying NO bets with a mean 3.3% edge over the market. On 23 bets, losing 13 is consistent with variance and says little about the selections.
Mean edge per value bet by fold, YES vs NO
Bundesliga BU v3.
The Bootstrap Result
With 52 NO bets at the 60% confidence threshold across three folds, the bootstrap 95% confidence interval for NO ROI is (-42.64%, +35.15%) around an observed -3.23%, a span of 77.8 percentage points. The YES direction, with 283 bets at the same threshold, has a much narrower interval: (-5.90%, +12.79%), also straddling zero.
With 52 bets, the observed -3.23% is consistent with a true ROI anywhere from strongly negative to clearly positive, so the interval neither supports the bilateral hypothesis nor rules it out. The deployment conclusion is the same as the statistical one: do not bet NO on this model until the sample can separate signal from noise.
Bootstrap distribution of NO bet ROI, Bundesliga ≥60% threshold
1000 resamples, n=52 bets. FAIL: lower bound ≤ 0%.
The Structural Conditions Are Present
Three observations suggest the bilateral hypothesis is still worth testing.
First, the model's NO zone is precise. It assigns P(NO) above 0.45 to only 16.6% of Bundesliga matches. In those matches, the actual NO rate is 50.7% against average implied odds of 42.9%, a 7.8 percentage point structural edge. The model makes targeted NO calls in a narrow set of matches and is right more often than the market expects. La Liga by contrast assigns P(NO) above 0.45 to 82.4% of matches with an actual NO rate of only 46.2% in that zone. That is why Bundesliga is the more likely league to support bilateral ROI: there the model discriminates between matches.
Second, the Bundesliga model is well-calibrated in both directions (ECE=0.018, mean absolute deviation 0.031 for both YES and NO). Bilateral ROI needs calibration on both sides, since a model has to price NO accurately before it can find NO edge. La Liga's model (mean absolute deviation 0.075) is too poorly calibrated to sustain bilateral edge even if the market mispriced both sides.
Third, the bilateral signal persists independently of the feature set. Running the same test on BU v4 (V4 venue tendency features, 37 features, a different architectural approach to Blaze) finds bilateral ROI at the 60% threshold with positive mean NO edge across all three folds (0.029, 0.040, 0.042). The v4 bootstrap also fails, but two structurally different models finding NO edge in the same market makes a feature-specific artifact much less likely.
The Data Requirement
A power analysis for the NO direction (one-sided proportions z-test, H0: win rate equals implied NO probability of 34.8%, alpha=0.05, 80% target power) puts the required sample at roughly 220 high-confidence NO bets for a medium-edge scenario (+8pp above implied). At approximately 17 qualifying NO bets per Bundesliga season at the 60% threshold, that means roughly ten more seasons of data, which back-testing cannot supply. Only live deployment at a small flat stake can build the sample.
Vig makes this harder. The raw bilateral signal survives vig adjustment (mean overround 1.052, approximately 5.2% headwind per bet) but only at the 62.5% threshold, not 60%. After vig the model needs a larger edge per bet than the raw odds suggest, and at 62.5% there are only 22 NO bets across three folds.
Whether the bilateral hypothesis holds live, or is an artifact of three particular Bundesliga seasons, is still open. The structural evidence makes it worth testing; the sample so far is too small to have tested it.