AXOFUND
AXOFUND
Working paper · 2026-08 · data quality
Two full market screens were spent refuting artefacts our own store produced. The finding is not that data quality matters — everyone says that. It is that a defective store converts research capacity into refutation capacity at a fixed exchange rate, and that the correct response is a moratorium on new work in that market rather than a smarter screen.
The usual account of a data defect is that it manufactures false positives. It does, and Section 2 gives the mechanism and two demonstrations. But the expensive direction is the other one. The same defect can suppress a real result, and then the rejection looks like rigour: a kill computed on an unvalidated store is indistinguishable, from the inside, from a kill that was earned. Section 3 is our own instance of it, and it very nearly cost us the one idea a multi-market campaign found.
Exchanges, vendors, instruments, tickers, return magnitudes and every dollar figure are withheld throughout; the defect classes, the counts and the direction of every error are not. What is withheld is declared where it is withheld.
Section 1
One national equity store in our estate held zero delistings across several thousand tickers over four years of daily bars. Zero is not a low number. Zero is a structural statement about the store: a feed that silently drops a series when the security stops existing, rather than retaining it with a terminal date and a reason, will report zero delistings forever, on any market, in any period.
Survivorship bias is a sampling problem you can reason about and sometimes bound. This is different. There is no surviving record of the dead to reason from, so the defect is invisible to every check run on the data itself. A universe that cannot lose members overstates every backtest run on it, and it does so silently.
The proof that this was not a theoretical concern arrived from a screen we were running for another purpose. We were looking for price-locked merger candidates — securities trading in a tight band against an announced consideration, which by construction stop printing when the deal completes. Not one candidate episode in the screen ever stopped printing. In a real market that is arithmetically impossible. The screen had not measured the mechanism; it had measured the store. The exchange, the ticker count and the candidate count are withheld, because together they locate a market we are still working in (see Redaction policy).
The absence independently blocks four research classes. Each of them needs the dead.
Merger arbitrage on announced deals trades an event in which a security ceases to exist, so a store with no terminations cannot represent the payoff at all. Index deletions are the forced-sale side of a rebalance, and usually the larger dislocation of the two; deletions are disproportionately names on their way out of the market. Tax-loss reversal draws its candidate set from the worst performers of a period, which is exactly the population most likely to be missing. Every cross-sectional book on that store runs its ranking, weighting and neutralisation over a universe that cannot lose members, so the ranking is computed against a survivor set and the backtest inherits the overstatement.
Our own leakage rules already require point-in-time universe reconstruction — what was listed on date D, determined from data observable before D. That requirement is unsatisfiable on a store that discards terminated series, and no amount of care in the research code repairs it. It is an acquisition problem, not a modelling problem.
Section 2
In the same estate, the corporate-action table was empty. Every bar was therefore unadjusted: raw closes, with dividends, distributions, coupons and splits left in the price path as discontinuities.
Treating that as noise is the error. Noise degrades a signal. This generates one, and it generates it in the specific shape that mean-reversion screens are built to find. On unadjusted closes, two instruments over the same underlying with different distribution calendars print a sawtooth into their spread: the spread steps down on one leg's ex-date, drifts back as the price recovers, and steps again on the next. A spread that reliably departs and returns on a published calendar is not a dislocation. It is the calendar. But a screen looking for departure and return cannot tell the difference, and the resulting crossings are numerous, regular and profitable-looking.
Demonstration one. A screen over listed securities carrying a scheduled coupon returned a positive mean across roughly a hundred trades, with the large majority of individual lines positive and test statistics clearing every conventional threshold. On its face it was the best result in its campaign. It was one hundred per cent coupon sawtooth. Entry timing gave it away before any economic argument did.
Demonstration two. In a separate campaign on a different market, a share-class pair with a genuine, documented conversion relationship was tested against a null of unrelated pairs drawn from the same universe. Three pairs in that null were pure ex‑dividend sawtooth on unadjusted bars — same underlying exposure, different distribution calendars, no economic relationship being traded at all. All three scored higher than the real mechanism. The artefact did not merely survive the screen. It outranked the thing the screen was built to find. The instruments and the scores are withheld; the ordering is the finding.
This is why we treat the pattern as a named artefact class rather than as a war story. It has a mechanism, it is reproducible, it is predictable from the calendar alone, and it will appear in any screen over same-underlying pairs run on any unadjusted store. It also explains a failure mode that would otherwise look like bad luck: the artefacts are not weak signals that slipped through. They are strong signals, because a calendar is more regular than a market. The rule that follows is to back-adjust at read, and to budget for an unadjusted store as an artefact generator.
The corollary is uncomfortable and we state it plainly: every positive result our screens produced on that store, before the defect was found, has to be re-run rather than re-read. A result computed on a generator of exactly the shape the screen rewards carries no information about the market.
Section 3
Everything above is the direction people expect. Defects manufacture winners; discipline kills the winners; the house is inconvenienced but safe. The asymmetry that argument hides is that a defect suppressing a real result is far more costly than one manufacturing a false one, because the false positive reaches review and the false negative never does.
Our instance: a single unadjusted forward split in our own store understated one instrument's compound growth rate by a multiple, not a margin. Not a rounding error, not a marginal shift in a ranking — the headline property of the series was wrong by more than the distance between the campaign's best candidate and its worst. Note what the quantity is: the compound growth of the instrument's own published price history under two adjustment treatments, a property of the data, not of any position we held. Both growth figures and the ratio between them are withheld (see Redaction policy); the direction of the error and its consequence are not.
Uncorrected, that figure would have inverted the instrument ranking of the only live output the campaign produced. The idea would have been ranked last instead of first and dropped, and it would have been dropped for a reason that reads as rigour: the compound growth does not justify the risk taken to earn it. Nobody in that review would have been careless. The kill would have been argued from evidence, recorded with its artefacts, and wrong.
We found it because the split was visible in a corporate-action audit we were running for Section 2 — that is, by luck of adjacency, not by a control we had designed. That is the honest account and it is the reason this section exists rather than being a footnote to Section 2.
Most research discipline is built asymmetrically: passes are scrutinised, kills are filed. The scrutiny is aimed at the direction where the house loses money, which is a reasonable instinct and an incomplete one, because the ideas a defective store kills leave no trace to audit later. A house that only audits its winners will be exactly as wrong as its data, and will never find out.
The instrument, the split ratio, the dates and every risk statistic attached to them are withheld. This idea is still live for us, so we publish the defect class and never the position.
Section 4
Two full market screens were consumed refuting artefacts described above. The natural response (build a smarter screen, add an artefact filter, tighten the null) is the wrong one, and it is wrong for a reason worth stating precisely: a defective store produces artefacts faster than any research programme can refute them. The exchange rate is unfavourable and it does not improve with cleverness. Each filter you add is calibrated against the artefacts you have already found.
So the rule is a moratorium, not a filter. No new books are authorised on a market whose store carries a known unrepaired defect of either class above. Existing verdicts on that market are marked provisional and each is required to name which defect could have produced it and whether that defect was controlled. Where the answer is "not checked", the verdict does not stand — it waits.
The second half is the reusable part: what it takes to trust a replacement feed. Buying data does not remove the risk, it relocates it. A vendor feed adopted without calibration is the same silent-zero failure we already pay for elsewhere (a component that reports success while doing nothing), just with an invoice attached.
When a feed lands, it is calibrated against two known answers, not zero: we re-run two studies whose answers we already established by other means, one where the defect suppressed a real result and one where it manufactured a false one, because a feed can be right about one direction and wrong about the other. Both target answers are written down before the feed is queried, so agreement cannot be reached by adjusting what counts as agreement. If the new feed does not reproduce both known answers, the failure attributes to the feed, not to us: it goes back to the vendor, and nothing downstream is authorised in the interim.
Then, and only then, the blocked work reopens. Merger arbitrage, index deletions and the cross-sectional books are unblocked by the acceptance test passing, never by the data arriving. Where identity had to be established statistically because the store carried no security master, it is re-verified against real identifiers, and the count of statistically-inferred relationships that turned out to be coincidence is published.
Which two studies serve as the calibration cases is withheld: together they identify the funded idea from Section 3.
Section 5
The obvious commercial response to a research programme finding little is to buy more markets. We had two additional national markets on a purchase list and we declined both, on the evidence in this paper.
The argument is a comparison of two constraints. A screening campaign across eight markets returned survivors in one. If the binding constraint were breadth, the marginal market would be worth its price. But Sections 1 to 3 establish that the store we already held manufactures artefacts that survive naive screening and suppresses results that should pass, which means the campaign's yield tells us about our data rather than about those markets. Buying a ninth market before repairing the first eight adds surface area that produces artefacts, and it adds it faster than we can refute them.
Ranking the acquisition list by measured defect rather than by coverage inverted it entirely. Point-in-time bars including terminated securities, and a corporate-action history, both for markets we already hold, went to the top. Two new markets went off the list. Neither decision was available from a coverage map; both fell out of the defect measurements, which is the argument for measuring defects as research output rather than as maintenance.
The markets, the vendors and every dollar figure are withheld.
Redaction policy
Our published position is that a refuted result is safe to name and a live one is not, and that the set of markets we screen is itself live information even when every individual result in it is dead. A list of where we looked reconstructs the programme. So this paper names mechanisms and withholds venues.
Exchanges and markets are withheld everywhere: the defect classes are general; the map of where we screen is not. Instruments and tickers are withheld, because Section 3's instrument is live for us. Return magnitudes, growth rates and t‑statistics are withheld as a set, and so are the ratios between them; published individually they are harmless, published together with the mechanism and the window they are attributable. Vendors and dollar figures are withheld, and neither bears on the finding.
What is published in full is every count, every base rate, every n, the direction of every error, and the four blocked research classes. These are the evidence, and withholding them would make this an assertion rather than a paper.
Every withheld claim on this page is wrapped in a marker, so the omission can be audited rather than taken on trust.
Cited by
Cross-references are hand-maintained; a link check runs before publish.