Loading…
Strategy Board — 24 structures × 4 expiry rungs (M-008). Short answer: the hub rates it NOT VALIDATED.
Data as of .
NOT VALIDATED. Pricing, greeks and both probability engines were checked; the 2026-08-30 fix removed a 1.28x variance artifact that had made 100% of rows positive — but no structure has ever been tested against realised P&L.
Development summary: The board prices every structure honestly; its edge column now behaves like an unbiased measure — median +0.2pp, 53% positive. It has never tested a structure against realised P&L, so the ranking is an unvalidated research shortlist.
| Item | Detail |
|---|---|
| Question it answers | Which multi-leg option structure, at which expiry, is priced away from what the underlying's own realised distribution delivers — as opposed to the broker win rate, which carries no edge by construction? |
| Method | Risk-neutral vs empirical P |
| How it works | Price all 24 templates at 4 expiry rungs off the banked per-strike chain; score each twice — risk-neutral P_model vs dedrifted, moment-matched empirical P_emp — and rank on edge_pp = P_emp − P_model |
| Inputs | Banked per-strike option chain, Own daily return distribution, Risk-neutral vs empirical odds, 24 structures × 4 expiry rungs, Bid/ask mid leg pricing |
| Output | edge_pp per structure × expiry; best template per symbol on the outputs view |
| Horizon | the rung's expiry — near ~21d, mid ~55d, far ~120d, LEAPS ~300d |
| Coverage | US names with a banked chain (176 symbols / 1,408 rows on the 2026-08-31 run) |
| Markets | US |
| Model type | Model · Structural |
| Stage | Research |
| Version | v2, first recorded 28 Aug 2026 |
Training data: not stated. Test data: not stated. Out-of-sample design: none.
| Measure | Result | Compared with | Note |
|---|---|---|---|
| Edge median after the variance fix | +0.2pp | +9.7pp before | the median row now has no edge, which is the correct answer for fairly priced options |
| Rows with positive edge | 53% | 100% before | a measure positive on every row of 960 was an artifact, not a discovery |
| Empirical sd ÷ implied sd (the bug) | median 1.28, above 1.0 on 213 of 213 | 1.00 required | a noisy mean-reverting vol estimate in a denominator; excess kurtosis was 0.3, so no fat-tail story was involved |
| Dedrift on NVDA, 13-day ATM call | 30.9% vs P_model 31.6% | 44.5% vs 31.6% before dedrifting | the apparent +12.8pp was the stock's own uptrend, 3.27% per 13-day window |
| Greeks vs TradingView's builder (NVDA Sep-11 217.5C) | delta 50.67, gamma 1.70, vega 16.31, rho 3.59 | 50.40 / 1.73 / 16.37 / 3.61 | net cost, breakeven and max loss exact; theta differs only because this engine prices at r = 0 |
| Forward P&L test | none | required for any use beyond a shortlist | needs historical option CHAINS; nothing was stored before 2026-08-29 and no subscription sells them back |
Forward record: 4 of 20 published lists have reached their 21-session horizon. Average pick return minus SPY's over the same dates: -4.23 percentage points (standard error 0.99). 29% of the matured picks beat SPY over the horizon. Mean rank IC of the matured lists: -0.027.
Is the design sound? The central idea is right and rare: rank on the gap between an empirical probability and the risk-neutral one, because the risk-neutral number is by construction edge-free. The moment-match refinement is the sharpest part — it concedes that the LEVEL of the empirical distribution cannot carry information once it is calibrated to today's implied vol, leaving only skew and kurtosis, which is exactly what Black-Scholes gets wrong. The dedrift is equally load-bearing: without it every upside structure inherits the stock's trend.
Data quality. Chains are real, banked per-strike, and legs are dropped rather than extrapolated when a target strike is absent — a strong discipline. The underlying panel is survivorship-biased. The chain store started on 2026-08-29, so there is no history to test against and the earliest boards were partly modelled surfaces.
Backtest. None, and this is the finding that sets the rating. Entry is today's price, exit is modelled at expiry intrinsic, and no structure has been marked to a later real book.
Against simple benchmarks. P_model is itself the benchmark for P_emp, which is the right comparison. The board's edge distribution after the fix (median +0.2pp, 53% positive) benchmarks the measure against what an unbiased estimator should look like, and it passes. Median edge by family moved from a uniform +10pp on long vol to +6 to +7pp on the pinning family, consistent with the variance risk premium this platform measured independently (VRP > 8 worth +1.18 vol points over a +1.09 always-sell baseline).
Independent replication. Greeks reproduce an external commercial builder to within 0.5% on delta, gamma, vega and rho, with net cost, breakeven and max loss exact. The synthetic reproduces put-call parity exactly (delta 100, gamma/theta/vega zero). Both probability engines were re-derived by hand on the ABBV worked example in the guide.
Sensitivity. The sd_ratio column is a per-row sensitivity monitor for the exact bias that was found, which is good practice. No sensitivity has been run on the winsorisation percentiles, on MIN_N_EFF, on the delta targets in the templates, or on the r = 0 assumption.
Monitoring. Every row ships n_eff and sd_ratio; the board-level edge distribution is the standing bias check. There is no stored history of that distribution, so drift has to be re-measured rather than read off a series.
Source: the Model Hub's public model register, as of 6 Nov 2026. The M-008 page in the live Model Hub carries the full development and validation record, monitoring and open decisions. All 30 models and their verdicts.