Tier 1 on canonical IV30 vs the proxy, same days (#48)¶

Days: 1184 (2022-01-03 .. 2026-09-22). Scored (label + forward outcome): proxy 1155, canonical 1155. VRP sign differs between the two on 483/1182 labelled days (40.9%).

Labels: production's pure functions, composed in backtest/replay_variants.py; the proxy variant matches production main() on every field of every day (0 mismatches over 1184 days, checked before this was written).

Outcome on each variant's own measure: vrp_realised = iv_t - rv_fwd_30.

The two contrasts #39 turns on¶

95% moving-block bootstrap intervals, block = 30 trading days, 2000 resamples, seed 48. An interval that straddles 0 means the sign of the difference is not established on this sample.

variant	contrast	n (a / b)	mean diff	95% CI	hit-rate diff	95% CI
proxy	transition - settled	743 / 412	-0.0045	[-0.0226, +0.0102]	-4.3 pp	[-14.8 pp, +4.4 pp]
proxy	calm-contango: vrp- minus vrp+	282 / 440	-0.0125	[-0.0245, +0.0007]	+1.2 pp	[-8.1 pp, +14.1 pp]
canonical	transition - settled	551 / 604	-0.0005	[-0.0269, +0.0187]	+0.2 pp	[-14.9 pp, +13.9 pp]
canonical	calm-contango: vrp- minus vrp+	637 / 85	-0.0139	[-0.0253, -0.0021]	-16.8 pp	[-29.4 pp, -2.9 pp]
canonical-svi	transition - settled	565 / 589	-0.0015	[-0.0270, +0.0172]	+1.0 pp	[-13.6 pp, +14.5 pp]
canonical-svi	calm-contango: vrp- minus vrp+	633 / 88	-0.0135	[-0.0258, -0.0018]	-13.7 pp	[-29.0 pp, +2.7 pp]

Churn: how often the gate is locked out¶

variant	year	labelled days	in transition	share	confirmed-label changes	VRP-sign changes
proxy	2022	251	189	75.3%	36	48
proxy	2023	250	148	59.2%	34	32
proxy	2024	252	174	69.0%	37	45
proxy	2025	248	175	70.6%	38	30
proxy	2026	181	62	34.3%	11	9
proxy	all	1182	748	63.3%	156	164
canonical	2022	251	183	72.9%	38	52
canonical	2023	250	98	39.2%	22	14
canonical	2024	252	81	32.1%	14	18
canonical	2025	248	96	38.7%	20	26
canonical	2026	181	95	52.5%	18	34
canonical	all	1182	553	46.8%	112	144

Tables — proxy¶

Transition vs settled (proxy)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`settled`	412	0.0362	0.0401	-0.0141	0.0231	0.0566	0.0742	0.876	-0.113
`transition`	743	0.0317	0.0397	-0.0264	0.0114	0.0676	0.094	0.833	-0.2622

By raw label (proxy)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`calm-contango-vrp+`	440	0.0369	0.0428	-0.0116	0.0247	0.061	0.0765	0.882	-0.2458
`calm-contango-vrp-`	282	0.0244	0.0309	-0.0009	0.0157	0.0438	0.0556	0.894	-0.2622
`calm-flat-vrp+`	38	0.005	0.0407	-0.2244	-0.0216	0.0717	0.0766	0.658	-0.2377
`calm-flat-vrp-` ⚠ thin	2	0.0121	0.0121	-0.0227	-0.0227	0.0469	0.0469	0.5	-0.0227
`normal-backwardation-vrp+` ⚠ thin	4	-0.0159	-0.0188	-0.1678	-0.1568	0.1419	0.1419	0.5	-0.1678
`normal-contango-vrp+`	123	0.0278	0.0345	-0.0412	-0.0161	0.0779	0.0934	0.683	-0.0991
`normal-contango-vrp-`	63	0.0198	0.0359	-0.0568	0.0049	0.0493	0.0525	0.794	-0.117
`normal-flat-vrp+`	123	0.0383	0.0796	-0.1949	0.0136	0.0995	0.1174	0.821	-0.2303
`normal-flat-vrp-` ⚠ thin	19	0.0561	0.0573	0.0024	0.0142	0.099	0.1138	0.947	-0.0015
`stressed-backwardation-vrp+` ⚠ thin	7	0.1077	0.1475	-0.1137	0.0779	0.1589	0.2343	0.857	-0.1137
`stressed-backwardation-vrp-` ⚠ thin	3	0.1538	0.1569	0.1335	0.1335	0.171	0.171	1.0	0.1335
`stressed-contango-vrp+` ⚠ thin	1	-0.007	-0.007	-0.007	-0.007	-0.007	-0.007	0.0	-0.007
`stressed-contango-vrp-` ⚠ thin	1	0.0339	0.0339	0.0339	0.0339	0.0339	0.0339	1.0	0.0339
`stressed-flat-vrp+`	45	0.0695	0.0649	0.0137	0.0345	0.0987	0.1177	1.0	0.0069
`stressed-flat-vrp-` ⚠ thin	4	0.0887	0.1015	0.0112	0.0976	0.1404	0.1404	1.0	0.0112

calm-contango-vrp+ by ex-ante VRP bucket (proxy)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`[0, 1)`	172	0.0283	0.0364	-0.0384	0.0166	0.0593	0.0726	0.843	-0.2458
`[1, 2)`	117	0.0399	0.0423	-0.0019	0.0269	0.0586	0.0798	0.889	-0.0457
`[2, 3)`	90	0.0391	0.0425	-0.0064	0.0247	0.057	0.0762	0.889	-0.0863
`[3, 4)`	36	0.0553	0.0593	0.0285	0.0352	0.0775	0.0882	0.972	-0.0005
`[4, 6)` ⚠ thin	25	0.0481	0.0565	0.0206	0.0379	0.0681	0.0705	0.96	-0.113

Tables — canonical¶

Transition vs settled (canonical)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`settled`	604	0.0073	0.011	-0.0453	-0.004	0.0294	0.0447	0.704	-0.1414
`transition`	551	0.0068	0.0249	-0.0646	-0.0087	0.048	0.0663	0.706	-0.2814

By raw label (canonical)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`calm-contango-vrp+`	85	0.02	0.0222	-0.0036	0.0069	0.0353	0.0418	0.882	-0.15
`calm-contango-vrp-`	637	0.0061	0.0124	-0.0473	-0.0026	0.0294	0.0445	0.714	-0.2814
`calm-flat-vrp+`	30	-0.0036	0.0314	-0.1379	-0.0227	0.0452	0.055	0.733	-0.2743
`calm-flat-vrp-` ⚠ thin	10	-0.0644	-0.0377	-0.2561	-0.0523	-0.0154	0.0441	0.2	-0.259
`normal-backwardation-vrp+` ⚠ thin	4	-0.0726	-0.0731	-0.1978	-0.1966	0.0535	0.0535	0.5	-0.1978
`normal-contango-vrp+`	61	0.0313	0.0435	-0.0423	0.009	0.0608	0.0754	0.787	-0.0988
`normal-contango-vrp-`	125	-0.0145	-0.0051	-0.1007	-0.051	0.0315	0.0451	0.48	-0.1414
`normal-flat-vrp+`	96	0.009	0.0444	-0.2225	0.0048	0.0708	0.0824	0.771	-0.2579
`normal-flat-vrp-`	46	0.0101	0.0338	-0.0308	-0.0181	0.0608	0.0712	0.63	-0.2519
`stressed-backwardation-vrp+` ⚠ thin	7	0.0415	0.0556	-0.1544	0.0029	0.1147	0.1445	0.857	-0.1544
`stressed-backwardation-vrp-` ⚠ thin	3	0.1114	0.1196	0.0927	0.0927	0.1218	0.1218	1.0	0.0927
`stressed-contango-vrp-` ⚠ thin	2	-0.0189	-0.0189	-0.0476	-0.0476	0.0098	0.0098	0.5	-0.0476
`stressed-flat-vrp+`	36	0.0332	0.0332	-0.0178	0.0171	0.0611	0.0725	0.778	-0.0372
`stressed-flat-vrp-` ⚠ thin	13	0.0314	0.051	-0.0292	-0.0236	0.0664	0.0807	0.692	-0.0384

calm-contango-vrp+ by ex-ante VRP bucket (canonical)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`[0, 1)`	49	0.0165	0.017	-0.0067	0.0027	0.0341	0.0456	0.857	-0.15
`[1, 2)`	32	0.0251	0.0267	0.0032	0.0143	0.0385	0.0416	0.906	-0.0153
`[2, 3)` ⚠ thin	4	0.0224	0.0256	0.0033	0.0194	0.035	0.035	1.0	0.0033

Robustness — canonical from the SVI surface instead of chain inversion¶

Transition vs settled (canonical-svi)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`settled`	589	0.0078	0.0126	-0.0486	-0.0046	0.031	0.0453	0.699	-0.1422
`transition`	565	0.0063	0.0233	-0.0609	-0.0071	0.0447	0.0649	0.71	-0.2824

calm-contango by VRP sign (canonical-svi)¶

group	n	mean	median	p10	p25	p75	p90	hit rate	worst
`calm-contango-vrp+`	88	0.0201	0.0231	-0.0046	0.005	0.0355	0.0532	0.852	-0.1492
`calm-contango-vrp-`	633	0.0066	0.0138	-0.0463	-0.0021	0.0302	0.0446	0.716	-0.2824

Limitations — these bound every number above¶

Short window. 2022-01-03 .. 2026-09-22 is ~4.7 years, one rate-hiking cycle and one bear market (2022). The 2009-2026 proxy tables in the 09-27 receipt are not comparable row-for-row; the proxy rows HERE, on the same days, are the like-for-like baseline.
Overlapping outcomes. Each day's outcome shares up to 29 of its 30 forward days with its neighbours. Read n as an upper bound; the block-bootstrap intervals are the honest uncertainty for the two contrasts.
Outcome is entry richness, not P&L. vrp_realised = iv_t - rv_fwd_30 says whether implied exceeded what was realised. It is not the P&L of any traded structure; that needs Tier 2 (historical chains through the EV engine).
One underlying. SPY only; no cross-sectional evidence.
Canonical series provenance. Chain-inverted SPY ATM IV30 from EOD chains (backtest/iv30_from_chain.py, r = 10Y TNX, a known tenor mismatch whose ATM effect is small once the forward is parity-implied). The SVI series is shown as a robustness check, not as a second independent source (same chains).
Hysteresis starts empty on 2022-01-03 in both variants, so the first few days bootstrap a confirmation. Identical across variants, so it cannot favour either.