Deployment Policy A/B Result
Eligible weekly launch cohorts: 181 (overlapping; inference uses launch-year blocks).
Paired Deltas vs Fresh Gradual
| policy_vs_fresh | horizon | metric | n_overlapping_cohorts | n_year_blocks | mean_delta | median_delta | win_rate | year_block_bootstrap_90_low | year_block_bootstrap_90_high |
|---|---|---|---|---|---|---|---|---|---|
| legacy_copy_full | 1m | return | 176 | 10 | 0.0050 | 0.0051 | 0.5682 | -0.0170 | 0.0086 |
| legacy_copy_full | 1m | mdd | 176 | 10 | -0.0338 | -0.0262 | 0.0398 | -0.0464 | -0.0305 |
| legacy_copy_full | 3m | return | 174 | 10 | 0.0073 | 0.0050 | 0.5517 | -0.0195 | 0.0134 |
| legacy_copy_full | 3m | mdd | 174 | 10 | -0.0261 | -0.0201 | 0.2299 | -0.0501 | -0.0235 |
| legacy_copy_full | 6m | return | 166 | 9 | 0.0084 | 0.0078 | 0.5301 | -0.0240 | 0.0236 |
| legacy_copy_full | 6m | mdd | 166 | 9 | -0.0174 | -0.0110 | 0.3012 | -0.0460 | -0.0095 |
| legacy_copy_full | 12m | return | 143 | 9 | 0.0131 | 0.0058 | 0.5245 | -0.0226 | 0.0280 |
| legacy_copy_full | 12m | mdd | 143 | 9 | -0.0113 | -0.0056 | 0.3636 | -0.0444 | -0.0069 |
| legacy_copy_staged4 | 1m | return | 176 | 10 | 0.0001 | -0.0014 | 0.4830 | -0.0026 | 0.0039 |
| legacy_copy_staged4 | 1m | mdd | 176 | 10 | -0.0093 | -0.0042 | 0.3295 | -0.0150 | -0.0083 |
| legacy_copy_staged4 | 3m | return | 174 | 10 | -0.0026 | -0.0038 | 0.4770 | -0.0092 | 0.0049 |
| legacy_copy_staged4 | 3m | mdd | 174 | 10 | -0.0117 | -0.0090 | 0.2931 | -0.0279 | -0.0106 |
| legacy_copy_staged4 | 6m | return | 166 | 9 | -0.0019 | -0.0032 | 0.4639 | -0.0142 | 0.0101 |
| legacy_copy_staged4 | 6m | mdd | 166 | 9 | -0.0061 | -0.0053 | 0.4518 | -0.0256 | -0.0010 |
| legacy_copy_staged4 | 12m | return | 143 | 9 | 0.0075 | 0.0015 | 0.5035 | -0.0143 | 0.0257 |
| legacy_copy_staged4 | 12m | mdd | 143 | 9 | -0.0045 | -0.0022 | 0.4406 | -0.0252 | -0.0011 |
Decision
No alternative passes the pre-registered robustness gate. Keep fresh_gradual as the default deployment policy; no P2 or production change.
Notes
- Results are descriptive across overlapping launch dates; year-block bootstrap reduces but does not eliminate market-episode dependence.
- Segment table is in segment_summary.csv; raw paired output is in cohort_policy_results.csv.