The method is open. The calibration is sealed.
We are not offering alpha. We are offering an honest, falsifiable diagnostic for volatility model risk — and the rigor that produced it. We work with a small number of desks, platforms, and research groups.
What we put first: what doesn’t work
Honesty is the moat, so the limits lead. Across equities, rates, commodities, FX and crypto, our coherence order parameter forecasts the size of the next move — and provably nothing about its direction (demonstrated against a powered momentum control). It does not beat traded 30-day implied volatility in any asset class tested; the options market already prices the signal.
There is no standalone trading edge at standard tenors today. What is confirmed, and useful, is below.
What is confirmed — and why an institution should care
The coherence hub forecasts forward realized volatility after proper controls, with an effect graded by market depth, and it flags the regimes where the industry-standard HAR volatility model tends to under-forecast. That is a model-risk signal, not a trading signal — a transparent, seven-indicator challenger you can run alongside your existing vol models.
Magnitude is lawful
Hub → forward realized vol is positive and survives a horizon-matched control, spline nonlinearity control, block-permutation nulls and out-of-sample split-half. Pooled p<0.003.
Graded by market depth
Strongest in deep, structured markets (equities, long Treasuries, gold, metals); fades toward zero in hyper-efficient intraday FX. The gradient itself is the result.
Flags HAR blind spots
Identifies, out-of-sample, where a backward-looking HAR forecast is most likely to under-forecast — a challenger model in the exact sense your validation framework already contemplates.
How we know what we know
The credibility is in the protocol, not any single number. We tried to break every finding before believing it, and we report where it broke.
| Discipline | What it does |
|---|---|
| Within-unit confirmation | A signal must hold inside each asset, not just pooled. |
| Block-permutation nulls | Inference at the autocorrelation length — i.i.d. permutation over-rejects on autocorrelated series. |
| Powered directional control | A null is only meaningful while a positive control fires on the same series. |
| Provenance-locked corpus | One real-market dataset; synthetic-data artifacts from earlier exploration quarantined and excluded. |
| Texas-sharpshooter guard | Stop re-slicing an era once it passes; only unseen forward data de-risks further. |
Every empirical claim referenced traces to results computed on a single, provenance-locked corpus of real market data. No standalone trading return is claimed at standard tenors.
Magnitude lawfulness and phase sovereignty across asset classes. The full method, controls, and every reported null — written falsification-first.
Read the method →Three ways to engage
- Fund the short-dated (0DTE) and rates/credit trigger tests on the questions that matter to you
- Early optionality on the upside tier if an edge survives the same rigor
- A definitive, co-publishable answer either way
- Pre-registered hypotheses, all nulls reported in full
- We run the diagnostic on an asset list you choose (e.g. SPY / GLD / your benchmark set)
- A short report showing where coherence flags HAR under-forecasting
- No data from you required; ~30 minutes of your time
- Not a trading signal — a validation overlay
- Co-authorship on the formal HAR head-to-head (QLIKE / Diebold–Mariano)
- Complexity-finance grant work supporting the data and the science
- Independent third-party validation of the program
All figures in any shared briefing are order-of-magnitude planning ranges, not commitments or forecasts. Nothing on this site constitutes investment advice or an offer of securities.