6-month international-entry backtest (monthly)

Validation selected context_only by log loss. Parameters refit each January; inputs update at each forecast cutoff.

Test cutoffs: 2022-01-01 through 2024-12-01; 36 forecast dates. 1987 player snapshots; 93 players; 139 positive snapshots; 26 distinct observed entrants.

Labels are recorded entry and appearances. Neither a mature horizon nor more snapshots repairs archive gaps.

Model Log loss Brier Average precision ROC AUC Top-10 precision Mean probability
base_rate 0.258 0.066 0.067 0.481 9.8% 4.8%
boosting 0.236 0.061 0.204 0.748 21.7% 4.6%
boosting_context 0.239 0.061 0.206 0.741 22.2% 4.5%
context_only 0.258 0.065 0.102 0.538 9.8% 4.7%
logistic 0.229 0.058 0.297 0.735 20.9% 5.2%
logistic_context 0.231 0.058 0.289 0.731 20.9% 5.1%

Observed entry frequency: 7.0%. Lower log loss/Brier are better; higher ranking measures are better. Top ten are selected separately within each country/date; small pools can make recall mechanically perfect. Baseline ties use arbitrary UUID order.

Count model MAE Poisson deviance
mean_count 0.525 2.187
poisson 0.522 2.092
poisson_context 0.522 2.092
zero_count 0.353 inf

Counts measure opportunity and persistence. A zero forecast can win MAE but has infinite deviance on positive outcomes.

Country Snapshots Players Positive snapshots Unique entrants Max missing affiliation fixtures
England 1 1 1 1 0
India 1978 91 132 24 5
South Africa 8 1 6 1 1

Country is dated auction nationality, a proxy for affiliation. Outcomes and prior-cap exclusions span all national teams. The missing-fixture flag covers only the cutoff affiliation. Zero-coverage countries are excluded; partial coverage still limits labels.

Overlapping horizons, repeated players and country/calendar shocks make rows dependent. Improvements are descriptive backtest differences, not precise independent-row estimates. See equal-player metrics, year/country metrics, common-period scores and calibration CSVs. Overseas samples remain too small for broad validation.