Contract — F6-conf-01
Question. Does the weekday carry predictive content for the one-day change in the 10-year Treasury CMT yield (DGS10), 2000–2019?
Estimands.
- Joint test of equal weekday conditional means of Δy_t (bp), classical one-way equal-variance ANOVA F(4, n−5) over all changes.
- Monday contrast: mean(Δy | Monday) − mean(Δy | Tue–Fri), all changes, 95% CI (chosen: Welch two-sample interval).
- Out-of-sample: RMSE over the evaluation period (changes dated ≥ 2010-01-01) of (a) zero-change forecast, (b) estimation-period (≤ 2009-12-31) weekday means. Assumptions: changes are approximately independent with finite variance; ANOVA additionally assumes equal variance & approximate normality (robust in large n); weekday means are stationary across periods for (3) to be meaningful.
Information rules. Drop empty DGS10 rows first; Δy_t = 100(y_t − y_{t−1}) between consecutive available observations, dated t, weekday of t. First available obs has no change. Weekday means for forecasting from estimation period only. Only data/DGS10.csv used.
Outputs. out/results.json (keys: n_missing_rows, n_train, n_test, anova_f_stat, anova_p_value, effect_estimate, effect_ci_low, effect_ci_high, baseline_metric, model_metric, conclusion, task_id, feasibility, feasibility_reason, data_issues, n_input_rows, n_analysed, random_seed); out/figure1.png; out/table1.csv (weekday,n,mean_bp,se_bp); out/report.md (6 sections); out/analysis.py; out/claims.json; out/exclusions.csv; out/state/*. Conclusion code: ANOVA at 5%; if not rejected → no_detectable_effect; if rejected, sign from Monday CI (positive/negative_effect; if CI covers zero while ANOVA rejects, I will report no_detectable_effect for the Monday-sign and document it).
Invalidity. Missing/unparseable file, non-weekday dates, duplicate dates, too few observations per weekday, or using evaluation-period info in forecasts.