1# F6-conf-01 — Day-of-week effects in daily changes of the 10-year Treasury yield23## Question4Does the weekday carry predictive content for the one-day change Δy_t (bp) in the 10-year US Treasury constant-maturity yield (FRED DGS10), 2000–2019? Three estimands: (i) joint equality of the five weekday means; (ii) the Monday contrast (Monday mean minus the Tue–Fri mean) with a 95% interval; (iii) out-of-sample RMSE of weekday-specific means versus a zero-change forecast.56## Data and exclusions7`data/DGS10.csv` has 5217 rows (`n_input_rows`), with SHA-256 matching the provenance file. 215 rows have an empty yield field (US market holidays; `n_missing_rows`) and were dropped before differencing. The first available observation has no change, which leaves 5001 daily changes (`n_analysed`). These split into 2500 estimation-period changes dated on or before 2009-12-31 (`n_train`) and 2501 evaluation-period changes dated on or after 2010-01-01 (`n_test`). The steps are in `out/exclusions.csv`. Data issues are listed in `data_issues`: 214 changes span a holiday, and there are large but genuine crisis-period moves (−51 bp on 2009-03-18). All of these were kept as specified. Full audit: `out/state/data_audit.md`.89## Method10- Joint test: classical one-way equal-variance ANOVA F-test of equal weekday means over all 5001 changes, F(4, n−5).11- Monday contrast: difference in means, Monday versus Tuesday–Friday, all changes. The 95% interval is a Welch two-sample t interval (unequal variances, Welch–Satterthwaite df; `monday_contrast_method`). As a robustness check, an OLS regression on a Monday dummy with HC1 standard errors was also run (`robust_hc1_ci_low`, `robust_hc1_ci_high`).12- Out-of-sample: each evaluation-period change is forecast by the estimation-period mean change of its weekday (`model_metric`). The comparison is the zero-change forecast (`baseline_metric`). The metric is RMSE in bp.13- Conclusion rule: if the ANOVA does not reject at 5%, the code is `no_detectable_effect`. Otherwise the sign comes from the Monday interval.14- The analysis uses no randomness (`random_seed` = null). Everything is regenerated by `out/analysis.py`.1516## Results17- ANOVA: F = 0.263 (`anova_f_stat`), p = 0.902 (`anova_p_value`). Equal weekday means are not rejected.18- Weekday means (`out/table1.csv`, Figure 1) range from about −0.22 bp (Tuesday) to +0.02 bp (Wednesday). Every weekday's 95% interval includes zero.19- Monday contrast: +0.050 bp (`effect_estimate`), 95% Welch CI [−0.319, +0.420] bp (`effect_ci_low`, `effect_ci_high`). The HC1 regression interval is practically identical: [−0.319, +0.419] (`robust_hc1_ci_low`, `robust_hc1_ci_high`).20- Out-of-sample RMSE: weekday means 4.884 bp (`model_metric`) vs zero change 4.876 bp (`baseline_metric`), a ratio of 1.0017 (`rmse_ratio_model_over_baseline`). Weekday means are slightly *worse* than the random walk.21- Conclusion: `no_detectable_effect` (`conclusion`).2223## Uncertainty24The Monday contrast's standard error is 0.188 bp (`monday_contrast_se`, Welch df ≈ 1549, `monday_contrast_welch_df`). The interval excludes Monday effects larger than about 0.42 bp in absolute value. That is small next to the daily standard deviation of about 5.7 bp (from the data audit), so the null result is reasonably precise rather than simply underpowered. The RMSE comparison is a single point estimate on one split, with no test of predictive accuracy (e.g. Diebold–Mariano). Because the gap is under 0.2%, it should be read as "no improvement", not as evidence that weekday means are harmful.2526## Limitations and open issues27- **What would invalidate the result:**28 - Strong serial dependence or volatility clustering. Crisis periods such as 2008–09 show both, and the classical ANOVA's independence and equal-variance assumptions would then misstate the p-value. The Welch and HC1 intervals handle unequal variances but not autocorrelation or clustering.29 - A weekday effect that reverses sign across sub-periods could cancel out in the pooled 20-year means. Only the 2000–09 vs 2010–19 split was examined, and only through the forecasting exercise.30- **Not checked:**31 - Sub-period or regime-specific weekday effects.32 - Weekday effects in volatility rather than in the mean.33 - Effects tied to FOMC or data-release calendars, which cluster on particular weekdays.34 - Robustness to winsorising outliers.35 - HAC standard errors.36 - Formal forecast-comparison tests.37 - Changes spanning holidays were attributed to the weekday of t without adjusting for elapsed days, as specified. Monday changes always cover a weekend.38- **Main threats:**39 - Mismeasurement from using the constant-maturity yield, an interpolated H.15 series, instead of traded prices.40 - Small-sample influence of a few large crisis-day moves on the weekday means.41 - Multiple testing, if one went looking for any weekday contrast. Here only the pre-specified Monday contrast was tested.4243None of these concerns is likely to turn a p-value of 0.90 into a rejection. They do, however, limit how far the result generalises beyond 2000–2019 and beyond the mean.44