Reject variance-lookback-short-symmetric hypothesis
Pre-registered 2026-04-17 to test whether lowering VARIANCE_LOOKBACK (10 → 5/6/7/8) would preserve CLV while enabling coverage for early- season bets that can't hit 10-match history. Tested via
Pre-registered 2026-04-17 to test whether lowering VARIANCE_LOOKBACK (10 → 5/6/7/8) would preserve CLV while enabling coverage for early- season bets that can't hit 10-match history. Tested via backtest-eval.ts --variance-filter --lookback=N across 5 values.
Results: lookback=5: n=23101 CLV=+10.1% ROI=-2.7% Hit=40.3% lookback=6: n=25141 CLV=+10.1% ROI=-2.8% Hit=40.4% lookback=7: n=26563 CLV=+10.0% ROI=-3.0% Hit=40.3% lookback=8: n=27802 CLV=+10.0% ROI=-2.5% Hit=40.4% lookback=10: n=29193 CLV=+10.0% ROI=-2.6% Hit=40.4%
Two unexpected findings:
- CLV is lookback-independent (+10.0-10.1% across all five values).
The signal quality is flat — not sensitive to window size.
- Lower lookback produces FEWER bets, not more. The 3.0-goal absolute
threshold is harder to hit in a shorter window — the coverage-unlock premise was wrong.
Verdict rejected: no CLV improvement, no coverage unlock, no meaningful ROI change. The real fix for early-season-bet coverage would be a two-parameter tuning (relax min-history floor AND scale threshold proportionally to window length). That's a distinct hypothesis — not registered in this commit.
Confirms the pattern from the rejected attack-defense-regression- asymmetry: the variance filter's edge is insensitive to lookback choices within 5-10 matches. The 10-match default stands.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>