Loading…
Options strategy tags (US + NSE) (M-004). Short answer: the hub rates it FIT FOR LIMITED USE.
Data as of .
FIT FOR LIMITED USE on US names: the sell_premium branch and the vol fields may be relied on to rank premium richness. long_premium is half-tested, avoid_illiquid untested, and the NSE fork is NOT VALIDATED.
Development summary: Only the sell_premium VRP leg has a measured forward edge (+1.28 vol pts over always-selling, audit as_of 2026-09-24). Its RV-percentile leg was retired at +0.01; long_premium is half-testable and avoid_illiquid is untestable at all.
| Item | Detail |
|---|---|
| Question it answers | Is this name's implied volatility rich, cheap, event-distorted or untradeable right now — the one-word vol stance the options tab and the structure selector read before any trade is shaped? |
| Method | Rule cascade on vol regime |
| How it works | Precedence rule over measured vol fields: avoid_illiquid > pre_earnings_iv > sell_premium > long_premium > neutral; thresholds set and retired on a bootstrap audit. |
| Inputs | VRP (implied − realised vol), Realised-vol percentile, Straddle edge, ATM spread & open interest, Days to earnings |
| Output | One tag per name per session (5 values) beside the vol fields it reads |
| Horizon | ~21 sessions (the near-expiry vol window the audit scores) |
| Coverage | US options panel ~4,000 rows, 614 with a live chain; NSE F&O 208 rows |
| Markets | US, IN |
| Model type | Model · Rule |
| Stage | Production |
| Version | v2, first recorded 14 Aug 2026 |
Training data: 2025-08→2026-07. Test data: not stated. Out-of-sample design: none.
| Measure | Result | Compared with | Note |
|---|---|---|---|
| sell_premium VRP>8 leg | +1.28 vol pts | +1.65 always-sell baseline | 175,722 symbol-days / 667 names (audit as_of 2026-09-24, inputs through 2026-10-23); 21-day block bootstrap clustered by date; CI [+0.54, +1.98] |
| RV %ile>60 leg (retired) | -0.00 | same baseline | CI [-0.53, +0.78] (as_of 2026-09-24); no standalone edge in any run — one line (TH_SP_RVPCT=60.0) reverts it |
| shipped two-leg tag, pre-retirement | +2.45 | same baseline | CI [+0.29, +4.64] on 8,290 rows (as_of 2026-09-24) — clears zero only narrowly, with no multiplicity correction across the audit's 8 tests |
| long_premium vol leg (VRP<2) | +0.55 | same baseline | CI [+0.10, +0.92] (as_of 2026-09-24); PARTIAL — the straddle_edge>1.05 and trend legs are not reconstructible |
| avoid_illiquid | not testable | no historical chains | per-strike OI/bid/ask are banked only since 2026-08-31 and bid/ask is zero since 10-13, so the tag stays an untested heuristic; D-003 (2026-10-26): its OI leg alone sets it on 53 of 485 rows |
| Metric | Sample / window | Compared with | Value | Range (CI) |
|---|---|---|---|---|
| Coverage (headline) | 423 of 485 rows carry IV (2026-10-23) | n/a | 87.20% | n/a |
| Edge vs control | audit 2025-08→2026-09 | always-sell +1.65 | +1.28 vol pts | [0.54, 1.98] vol pts |
| Observations | symbol-days · 667 names | n/a | 175,722 | n/a |
Is the design sound? The precedence order is right for the use: liquidity is a veto, an event distorts every vol read, and richness/cheapness is the decision. Scoring the tag against vol_pnl rather than stock direction is the correct outcome for a premium decision. The weak link is conceptual honesty at the edges — two branches name conditions the engine cannot observe on most rows.
Data quality. IV coverage is the binding constraint: 614 of 4,021 US rows have iv30, 126 have oi_total. The IV history behind iv_pctile is survivorship-filtered (621 names in today's universe, none backfilled for delistings) — exactly the bias that flatters short-vol rules. NSE has no OI, no contract volume and no LTP by design.
Backtest. One outcome window (21 sessions), one regime, ~11 months in the published run. The wider 24-month lens panel agrees in sign and keeps the VRP branch on the test half (+1.50 [+0.83, +2.40]), which is the strongest support the tag family has. The audit's own combined-leg result (+2.31 [+0.07, +4.80]) would not survive a BH correction across its 8 tests.
Against simple benchmarks. The right benchmark — always sell premium — is built into the estimator, and the tag has to beat it rather than merely be positive. The lens work adds a second benchmark: a HAR-forecast hurdle ranks vol richness better than the raw VRP (IC 0.1259 vs 0.1026), so the VRP threshold is not the best available ranker even where it is the measured one.
Independent replication. Threshold values, precedence order and branch conditions were read against the live feed: the current US snapshot returns 46 sell_premium, 4 long_premium, 80 avoid_illiquid, 9 pre_earnings_iv among IV-bearing rows, and NSE returns 25 long_premium, 6 avoid_illiquid, 0 sell_premium — consistent with the forked thresholds. The bootstrap itself was not recomputed.
Sensitivity. Threshold sensitivity is documented in the code rather than tested: reverting TH_SP_RVPCT to 60.0 restores the old tag in one line, and the two CIs overlap so heavily ([+0.07,+4.80] vs [+0.45,+1.98]) that neither variant is demonstrably better. No sensitivity run exists for the 5-21 day earnings window, the 12% spread gate or the 5,000 OI floor.
Monitoring. None beyond the pipeline's own coverage prints. Minimum viable: tag mix and IV coverage per run, alert when IV coverage falls below the last 20-session median, and a scheduled re-audit trigger.
Source: the Model Hub's public model register, as of 6 Nov 2026. The M-004 page in the live Model Hub carries the full development and validation record, monitoring and open decisions. All 30 models and their verdicts.