AXOFUND

Proprietary quantitative research

We look for someone who has to trade. Then we try to break the result.

AxoFund trades its own capital. No clients, no outside money, nothing here to sell. What follows is the method and the record.

Somebody has to be obliged. A fund whose rules make it trade at the close whether it wants to or not. An authority that has promised to hold a price inside a band and has to spend money when it strays. Two claims on the same asset that a contract says must settle at the same value. The obligation is written down, and it is written down before anyone looks at the price.

The usual alternative is to start at the other end: two things moved together for years, so trade them. We do not accept that as a reason. A candidate names the obligated party and the trigger that binds them, or it is not a candidate. A pattern that merely held is not a thesis.

Then we try to break it. A candidate is re-run against controls built to strip out the one thing it claims to be trading; if the controls do as well, the claim was decoration. Of 2,029 strategies written to date, 265 lasted long enough to face that test and 19 came through it (Fig. 1).

A second and harder test is built but not yet proven: data with no edge in it at all, which asks whether a book beats nothing rather than whether it beats its neighbours. The first time we pointed it at those nineteen, fourteen sat below it, three were marginal and two cleared. We publish that as readily as we publish the nineteen: three of the eleven items in the library are corrections or retractions of our own earlier work.

Most candidates do not survive. Increasingly, neither do our own verdicts.

Models author strategy code and propose hypotheses here at volume. Nothing a model says is permitted to adjudicate; the gate is arithmetic. What the machines got wrong →

FIG. 1 — THE SPECIFICITY GATE
authored 2,029 abandoned before test 1,764 gate-tested 265 refuted at the gate 246 survived 19
Of 2,029 strategies authored to date, 1,764 were abandoned before any gate test: killed at authoring, at the novelty check, or by a sweep that produced nothing worth testing. 265 accumulated enough independent windows to face the specificity gate, where each candidate is re-run against a control ensemble built to null exactly the mechanism it claims; 246 were refuted there and 19 survived, under Benjamini–Hochberg false-discovery control with consistency required across independent windows (the pre-registered false-discovery rate is not published). Two readings the bars do not give you. Each of the 265 was itself the surviving parameterisation of a several-thousand-trial optimisation, so the gate sees a population already selected for looking good. And we have since measured how often that selection manufactures a survivor out of nothing at all; that measurement, and its correct reading, are published in full as We measured our own instrument (2026-08) rather than summarised here. Earlier copy on this page called 2,010 strategies "refuted". That was wrong: it counted the abandoned with the tested, and abandonment is a weaker word than refutation. Corrected 2026-08 by “Refuted” overstated what we had actually tested, which is the correction of record; this page is the home of record for the counts themselves.

Correction of record — 2026-08

The gate itself failed its own test

The survivability score that ordered candidates through the funnel above was measured against the forward window it claimed to predict. Rank correlation: 0.98 across the full gated cohort. A score that tracks its own evaluation window that closely is restating the outcome, not predicting it. A second measurement compounded the verdict: across every forward cell we hold, the top 1% of cells carried 117% of total return, a shape a mean-based objective cannot see. We retired the gate mechanism on 2026-08-08.

Two consequences, stated plainly. The funnel stands as the record of what was tested. And the nineteen survivors are unproven rather than disproven: their claims returned to the queue with everyone else's. The corpus now stands at 3,080 books under a governed naming standard in which a control is named a control, and every book re-runs over 27 pre-registered cells: three entry windows, three direction arms, three objective functions, leverage pinned at one. No score adjudicates. A book earns or loses standing on realised forward evidence, judged per family against criteria frozen before the data arrived.

How we got here

Built, measured, corrected

2025–2026, in order. A custom backtest engine, retired when its fills diverged from live. A sequential portfolio fleet on an industrial-grade engine, kept. Five separate attempts at vectorised backtesting, each abandoned when measurement showed the speed was purchased with the information; the fifth post-mortem is published in the library. A four-gate evaluation chain, published, then retired by the correction above when we measured the instrument instead of trusting it. A sensitivity analysis over the engine's own trial histories that let us freeze the parameters the objective never responds to, cutting the search severalfold with no loss it could measure. Each step ran on the cluster we built for the purpose (the machines), and each correction cost more compute than the result it replaced. The mathematics, the trading, and the engineering are one argument here: what survives is what we could not break.

The argument

Two nulls, and a floor under every kill

We run two nulls, because they answer different questions: a matched control asks whether a book beats its neighbours, and only a calibrated, edge-free generator (series required to reproduce the market's volatility clustering, tails and correlation structure while containing nothing to find) asks whether it beats nothing at all. We had been reading the first answer as though it were the second; the three incidents that forced the distinction, and the doctrine reversal we published against ourselves when it landed, are Two nulls, because they answer different questions (2026-08). The generator is built and its out-of-sample moment-match table is not yet in; until it is, it is an instrument rather than a null, and we say so where it is used.

And a kill is not a finding without a detection floor. State the smallest edge the search could have seen, or the verdict is a statement about the search rather than about the market. Every verdict is now required to cite the floor for its class. Two mechanism classes have been measured; the rest are outstanding, and a verdict citing one of those is a prerequisite batch rather than a kill. The floor's value is proprietary.

The mechanism classes themselves (a rebalance mandate, a defended band, share-class identity, contract convergence, an index-tracking mandate) and the order the gates run in are set out on method.

The library

  1. 3 papers to read if you are reading once
    how models are governed and why none of them adjudicates · what the pipeline returned when we ran it over panels containing nothing to find · the defects we found in our own bar store, including one that hid a real result     start here →
  2. 11 items published to date, each under its title and the month it was published
    five working papers, a decommission, a data-quality finding, a charter, two corrections and a retraction · dated · status tracked · never silently edited     library →
  3. 244 papers shelved across 20 shelves
    verdicts assigned shelf by shelf: 48 recorded so far, and the page says which shelves are done     reading →
  4. 3 corrections and retractions issued against our own record
    a funnel figure that called abandonment refutation; two different breadths published under one name; and a working paper retracted in whole, its corpus optimised at a fraction of production budget     corrections →

Each figure has one home of record and is cited from it, not restated. Estate figures (how many venues, how many asset classes, which bar intervals) live on coverage. Corpus figures live on reading. Cross-references across the library are hand-maintained; a link check runs before publish.

FIG. 2 — THE SHELF, BY MONTH
1 3 11 2026-06 2026-07 2026-08
Cumulative items in the library, by month of publication: 1 in 2026-06, 3 by 2026-07, 11 by 2026-08. n = 3 months. This is a record, not a trend, and at this length it cannot be anything else, which is the reason to date every step rather than assert a body of work. An item is cited by its title and the month it was published; a date is never revised and a title is never reused for a different item. A wrong item is left in place and corrected by a new item that cites it, and the status vocabulary is closed: current, revised, superseded, corrected, retracted. A retracted item stays on this count and in the index; withdrawal changes its status and nothing else, so the line never falls.

Published against ourselves

Recently refuted

  1. Carry as alpha A G10 carry basket cleared survivability with a stable Sharpe across every test window and was refused at the specificity gate: p = 0.62 against a beta-adjusted null. It is the carry factor, correctly named, and we publish it as one. It was also that market's planted control (a factor strategy seeded among the candidates, expected to clear survivability and die at specificity), and it died as designed, which is the only reason the pass rate around it can be read at all. Enforced structure ports. Statistical structure does not. →
  2. Cross-sectional mean reversion, US equities Negative after real per-share costs in every test window. No parameterisation rescued it, because the fee floor is not a parameter. The fee floor is the finding, and the campaign that produced it returned no book that passed both gates. Enforced structure ports. Statistical structure does not. →
  3. Our own first equity flagship It passed both gates and was void 48 hours later, killed by our own fills ledger: the instrument records showed the backtest had filled on a venue the book's label did not claim. Three further asset-class routing defects were excavated underneath it. The verdict was withdrawn and the campaign's equity results with it; the pre-flight check that caught it is now standing. The discovery chain is the strongest exhibit that campaign produced. Enforced structure ports, revised 2026-07 →