Research Integrity
Why we pre-register
If you test enough strategies, some will look brilliant purely by luck. This is the central hazard of quantitative research, and most published "edges" do not survive an honest correction for it (Harvey, Liu & Zhu, … and the Cross-Section of Expected Returns, Review of Financial Studies, 2016). Our defense is procedural: for every hypothesis we record, in advance, the test design, the data window, the success metric, and the minimum evidence required — and we freeze them before looking at results. We cannot move the goalposts after the fact, because the goalposts are written down and timestamped.
The bar we hold to
A signal is only allowed to influence a model's capital — or to appear anywhere in our marketing as a claim — after it clears a significance bar set well above the conventional level, chosen specifically because it penalizes for the number of ideas we have tried. The bar is applied to out-of-sample evidence: how a signal behaves on data it was not built on. Until a signal clears the bar, it stays observational — measured, logged, and reported internally, but never driving live capital and never sold as a result.
What we have confirmed — and rejected
A representative, plain-language summary of hypotheses we have tested. Method names are generalized; the verdicts are not.
| Hypothesis (plain language) | Verdict |
|---|---|
| A crowding / short-horizon mean-reversion signal carries an edge that remains after removing exposure to the standard style factors (value, size, momentum, low-volatility, quality). | Confirmed |
| A trend-following signal is a standalone stock-selection edge in its own right. | Rejected — reduces to well-known factor exposure. |
| A low-volatility tilt adds selection edge beyond the standard factors. | Rejected — explained by existing factors. |
| A structural-stress / correlation-regime indicator can time broad market risk better than standard volatility measures. | Rejected — no edge beyond a standard volatility index. |
| That same indicator can reduce drawdowns by de-risking at detected transitions. | Rejected — a simple trend filter managed downside better. |
| Path-shape pattern matching of historical sell-offs can forecast future ones. | Rejected — did not generalize out of sample. |
| A reduced-turnover rebalancing rule cuts trading frequency versus a fixed calendar. | Rejected on the headline claim — it traded more, not less; a separate quality-per-trade benefit held up. |
| An estimation-noise reduction step improves portfolio construction. | Under live evaluation |
| Signal weighting conditioned on the market regime beats a fixed blend. | Under live evaluation |
| Our signals add value under independent, blinded third-party validation. | Under evaluation — below median so far; reported honestly, claim withheld until the sample is adequate. |
Confirmed cleared the pre-registered out-of-sample bar. Rejected failed it and does not drive live capital or claims. Under evaluation is being measured live and is not yet claimed either way. We expect this table to grow more rejections than confirmations over time — by design.
What we claim, and what we don't
We claim: a modest, factor-neutral edge from crowding and mean-reversion dynamics, expressed through disciplined portfolio construction and risk management. We do not claim: an ability to time the market, a proprietary forecast of prices, or that any idea still under evaluation works. When something is unproven, we say so on this page rather than implying it is settled.
Models, not advice
Everything we publish describes the output of our quantitative models, tracked forward in real time against live market data. It is research content, not personalized investment advice, and the performance shown is forward-tracked model output, not the return of any client account. Please read the full Disclaimer, and see the tracked performance record for the numbers themselves.
← Back to home