Ablation Note v0.1

Synathra.Sport Ablation Note v0.1

Event-only baseline versus full hybrid_score_v0_1 across five football validation groups. The note answers whether the hybrid event-chain layer adds measurable ranking value beyond calibrated event risk alone.

5
validation groups
822 283
event rows
4.79x
mean full hybrid top-10 lift
+2.22x
mean lift gain vs event-only

Main result: the full frozen hybrid_score_v0_1 improves the event-only baseline in all five validation groups. Mean top-10 lift increases from 2.57x to 4.79x. Mean ROC-AUC increases from 0.637 to 0.742.

1. Purpose

The public v0.5 report showed that a frozen hybrid score transfers across men's international tournaments, club football, and women's football. The remaining reviewer question was whether the full score adds value beyond a simple event-risk ranking. This ablation note compares an event-only baseline with the full hybrid score on the same target and the same five validation groups.

2. Compared scores

Event-only baseline

The baseline ranks events by event_risk_calibrated only. In the v0.5 exported event-level tables, the operational column used is synathra_calibrated_score, because the public scoring script falls back to this column when a literal event_risk_calibrated column is absent.

Full hybrid score

The full model ranks events by the frozen public score hybrid_score_v0_1. It combines event risk with chain membership, local chain-delta context, and subtype reliability.

hybrid_score_v0_1 = 0.50 * event_risk_calibrated + 0.05 * inside_hardened_chain + 0.10 * chain_delta_context + 0.35 * subtype_reliability

The validation target is future_goal_h10: a positive case means that a goal occurs within the next 10 event steps after the evaluated event. Top-10 lift compares the positive rate in the top decile by score with the base rate of the corresponding validation group.

3. Group-by-group results

Validation group Domain type Rows Positives Base rate Event-only lift Full hybrid lift Lift gain Event-only AUC Full AUC AUC gain
FIFA World Cup 2018 same_family_base 71 838 590 0.82% 2.51x 4.66x +2.15x 0.637 0.742 +0.105
FIFA World Cup 2022 same_family_cross_season 234 637 1 808 0.77% 2.59x 5.01x +2.41x 0.632 0.748 +0.116
UEFA Euro 2024 cross_competition_mens_international 187 924 1 140 0.61% 2.84x 4.77x +1.93x 0.670 0.756 +0.086
Ligue 1 2021/2022 club_transfer 101 766 810 0.80% 2.58x 4.69x +2.11x 0.626 0.734 +0.108
Women's World Cup 2023 gender_transfer 226 118 1 664 0.74% 2.33x 4.84x +2.52x 0.618 0.729 +0.111

4. Interpretation

The full hybrid_score_v0_1 improves the event-only baseline on top-10 lift in every validation group. The mean lift increase is +2.22x, or approximately 87.4% relative to the event-only baseline. This supports the product claim that Synathra.Sport is not merely a calibrated event-risk ranking: the chain-context and subtype-reliability terms add measurable value for identifying the highest-risk event windows.

ROC-AUC also improves in every group. The mean AUC gain is +0.105. This global ranking improvement is smaller than the top-decile lift gain, which is expected because the product use case is concentrated in the highest-ranked event windows used for reports, chain cards, and analyst review.

5. Product conclusion

The ablation result strengthens the public v0.5 validation package. The hybrid layer is empirically useful: it preserves transfer across domains while improving ranking quality over event-risk alone. For media, academy, and small-club use cases, the practical implication is that Synathra.Sport can prioritize a smaller set of interpretable high-signal sequences rather than only ranking isolated events.

6. Caveat

This note uses synathra_calibrated_score as the event-only baseline because the v0.5 event-level exports do not contain a literal event_risk_calibrated column. This follows the fallback order used by the public compute_hybrid_score_v0_1.py script. If a future pipeline exports a literal event_risk_calibrated column, the ablation should be re-run with that exact column name for strict nomenclature alignment.

7. Data files

Group metrics CSVRounded public comparison table Full precision CSVFull numeric precision for audit/recalculation Input table inventorySource event-level tables used in the computation Summary JSONMachine-readable summary metrics

The computation draft is also available as Markdown: ABLATION_NOTE_v0_1_COMPUTATION_DRAFT.md.

Back to report Technical Method Note Method notes