Loading…
Option Lens Suite V/E/S/T/F/X (M-006). Short answer: the hub rates it FIT FOR LIMITED USE.
Data as of .
FIT FOR LIMITED USE for the V and T hurdles as RANKING inputs, on the strength of a pre-registered, chronologically split, date-clustered test. NOT VALIDATED for E, S and F, which have no data to test, and for the thresholds as levels.
Development summary: The HAR forecast is a better-calibrated, better-ranking delivered-vol benchmark than the raw VRP the tab ships, and AR(1) reversion predicts the next IV change cleanly. 4 of 6 cells hold on the test half; 3 lenses have no data to test.
| Item | Detail |
|---|---|
| Question it answers | For one optionable name: is the option charging more volatility, event move, tail risk or term premium than the underlying is forecast to deliver — expressed as one comparable hurdle per lens? |
| Method | HAR-RV + AR(1) lenses |
| How it works | Six lenses, each implied minus delivered = hurdle, ranked as h_pct; delivered side of V is a pooled HAR-RV OLS forecast, of T a per-name AR(1) on own IV. |
| Inputs | 30-day implied vol (iv30), HAR-RV delivered-vol forecast, IV mean reversion, AR(1), IV percentile (252 sessions), Option wings / skew |
| Output | 6 hurdles + h_pct + a one-word Read per name; heat only on measured cells |
| Horizon | 21 sessions (the window the V hurdle forecasts and the audit scores) |
| Coverage | 614 US names carrying an iv30; E 139, S/F 122, T 611, X 605 |
| Markets | US |
| Model type | Model · ML/Stat |
| Stage | Production (4 of 6 registered cells hold on the test half; E/S/F untested) |
| Version | v1.2, first recorded 2 Sep 2026 |
Training data: 2021-08→2024-08. Test data: 2024-08→2026-08. Out-of-sample design: holdout.
| Measure | Result | Compared with | Note |
|---|---|---|---|
| HAR hurdle rank IC vs vol_pnl | 0.1259 (t 9.5) | 0.1026 (t 7.58) raw VRP | 492 dates, Spearman within date, Newey-West lag 21; HAR wins in BOTH halves (train 0.1306 / test 0.1207) |
| rev_pred vs actual 21-session IV change | IC -0.4406 (t -12.33) | no benchmark — nothing else predicts IV change | 365 dates, IC negative on 99% of them; MZ slope 0.69 R2 0.2461 — the AR(1) over-promises by ~31% |
| Forecaster race, held-out year | HAR QLIKE 8.3429, MZ slope 1.167 | naive rv20 QLIKE 8.4933, slope 0.759 | EWMA(0.94) beats HAR on RMSE (16.34 vs 16.91, DM t 9.62 vs 2.42) but is worse calibrated; the hurdle cares about calibration |
| Cells keeping test-half edge | 4 of 6 | baseline vol_pnl +0.38 on the panel | V:vrp_gt8_sell, V:hurdle_gt8_sell, V:ivpct_ge90_sell, T:rev_ge5_sell; the two buy cells sign-flip between halves |
| Panel | 297,173 symbol-days / 614 names | 141,296 in the house audit (T-006) | 2024-08-14 to 2026-08-03; the house panel drops every row without a strict-252 IV percentile and loses its first year |
| Metric | Sample / window | Compared with | Value | Range (CI) |
|---|---|---|---|---|
| Rank IC (headline) | OOS panel 2024-08→26-08 | raw VRP rank 0.1026 | 0.1259 | n/a |
| Excess return at horizon | test half to 2026-08 · hurdle>8 sell | raw VRP cell +1.50 | +1.65 vol pts | [0.61, 2.43] vol pts |
| Observations | OOS panel 2024-08→26-08 · dates | n/a | 492 | n/a |
Is the design sound? The central idea is right and is the best-argued piece of options work on the dashboard: the shipped VRP measures implied against PAST realised vol, and the honest comparison is against a FORECAST of the same 21 sessions the seller is short. The evidence supports it — naive rv20's MZ slope of 0.759 with a +10.49 intercept means the raw VRP is a biased hurdle by construction, and HAR's slope of 1.167 fixes the calibration. Refusing to print a theoretical price against a market price, and making lens X explicitly a ruler, are both correct. The weak point is that the suite's six-lens symmetry implies six comparable readings when only two carry evidence.
Data quality. Coverage is the binding constraint and the engine says so: V 100%, T 99.5%, X 98.5%, but E 22.6% and S/F 19.9%, because the per-contract book has persisted three sessions. Nulls are correctly 'not measured on this name', never zeros. The IV history (T-007) is survivorship-filtered to today's 621 names, which flatters short-vol results. Percentile and AR(1) cells only exist from ~2025-02 under min_periods 126, so their train half is about six months; the strict-252 versions cannot be split at all.
Backtest. The grid was pre-registered before scoring, split chronologically, bootstrapped in 21-day blocks clustered by date and corrected with Benjamini-Hochberg — the most careful design in the options stack. It still tests stability inside a single regime: 24 months, both halves sharing one macro backdrop, and the only vol explosion (2025-03, baseline -18.83 vol pts) sitting entirely in the train half. The two cells that fail are precisely the two whose train and test halves disagree, which is the tier rule working as intended.
Against simple benchmarks. Properly benchmarked on the forecast side: three forecasters raced on a held-out year with MZ, QLIKE and Diebold-Mariano. The honest reading is that EWMA(0.94) beats HAR on squared error (RMSE 16.34 vs 16.91, DM t 9.62 vs 2.42) and HAR wins only QLIKE and calibration — which is the right criterion for a hurdle, but the spec's description of EWMA as a mere 'fallback' understates it. On the trade side the benchmark is always-sell (+0.38 on this panel), which every cell must beat rather than merely be positive.
Independent replication. Not recomputed. Constants, lens definitions, gate values, coverage counts (614/139/122/611/122/605), the five heat cells and the tier map were read against the live feed and the validation JSON and agree. The validation file's own house-reproduction table is the useful internal check: same estimator against gives the same sign and overlapping CIs on all four shared cells (VRP>8 +1.91 [+0.69,+3.33] here vs +1.27 [+0.54,+2.07] published), with the difference explained by the house panel dropping every row without a strict-252 percentile.
Sensitivity. Thin. No sensitivity run exists for the lens thresholds (8 vol pts, +/-5 rev_pred, the 10th/90th percentiles), for the phi clip, for the pooling choice (one HAR for 614 names) or for the 21-session horizon. The validation itself names the one incremental test that matters and refuses to run it post hoc: whether hurdle>8 AND vrp<=8 finds rich vol the raw VRP misses, or merely re-selects the same rows.
Monitoring. Better than most: the feed already carries the per-run forecaster race and per-lens coverage. Missing is drift monitoring on the re-fitted coefficients against the frozen set, an alert when V or T coverage falls, and a scheduled re-run of the pre-registered grid rather than an ad-hoc one. Addendum 2026-09-24: the per-run holdout table, coef_source and now provide coefficient drift monitoring; an alert on coef_source flips is still missing. Addendum 2026-10-26 (v1.2): with the refit off the nightly path, monitoring is the frozen set's own out-of-sample table (meta.har.oos_monitor) against EWMA(0.94) and naive RV20.
Source: the Model Hub's public model register, as of 6 Nov 2026. The M-006 page in the live Model Hub carries the full development and validation record, monitoring and open decisions. All 30 models and their verdicts.