Verification checks (P4)
- Clean re-run: deleted results.json/table1.csv/figure1.png/exclusions.csv, re-ran analysis.py → results.json byte-identical (cmp), table1.csv sha identical. PASS (repeated after adding robustness keys: PASS).
- Counts: 5217 raw − 215 empty − 1 (first obs) = 5001 = n_analysed = n_train(2500)+n_test(2501) = sum of table1 n. PASS.
- Date ordering: last_train_date 2009-12-31 < first_test_date 2010-01-04; weekday means for forecast computed from train subset only. PASS.
- ANOVA cross-check: statsmodels OLS dy~C(weekday) overall F = 0.26347, p = 0.9015, identical to scipy f_oneway. PASS.
- Monday contrast: Welch CI [−0.319, 0.420] contains estimate 0.050; SE 0.188 > 0; HC1 regression gives the same point estimate and nearly identical CI; pooled CI slightly wider. PASS.
- Magnitudes: model RMSE (4.884) vs baseline RMSE (4.876) — difference 0.008 bp, plausible given weekday means of order 0.1–0.4 bp vs sd ≈ 5 bp. PASS.
- Conclusion rule: ANOVA p = 0.90 ≥ 0.05 → no_detectable_effect; consistent with Monday CI containing 0. PASS.
- Note: an ad-hoc check without rounding gave Kruskal p 0.471 vs 0.518 in the script; difference is due to float noise breaking ties in Δy (script rounds to 1e-10). ANOVA/means unaffected. Documented, not a failure.
- Assumption check: Levene p ≈ 1e-7 → weekday variances unequal (Monday sd 4.98 vs Friday 6.04); classical ANOVA's equal-variance assumption is violated, but Kruskal–Wallis (p 0.52) and Welch/HC1 Monday contrast agree on no effect. Reported as limitation.