AXOFUND
AXOFUND
Coverage
Enforced relationships are individually small. A prospectus rebalance mandate, a defended band with a formal intervention obligation, a share-class legal identity — each is real, and each has a capacity ceiling. A book built on them only works if you can find many, across many venues. Breadth is therefore an arithmetic requirement.
This page is the home of record for every estate figure on this site. It carries two breadth counts, because there are two true numbers and they had never been given different names. It carries what is actually held, not what the vendor markets. And it carries a register of the defects we have found in our own store, which is the part of a coverage page that is usually missing.
We have spent more of this year repairing the store than searching it, and the register below is why.
Elsewhere on this site the estate has been described as twenty-nine exchange feeds, and here as twenty-one exchanges. Both are true. They measure different things, and the discrepancy was a naming failure, not an arithmetic one.
| Count | Definition |
|---|---|
| 29 | Catalogued breadth. Exchange feeds we hold vendor entitlements to, whether or not a bar from them has ever entered a study. This is the number that describes what we could ingest tomorrow. |
| 21 + 2 | Ingested breadth. Twenty-one exchanges plus two derivative venues that land in the store on a nightly schedule and are read by live research. This is the number that describes what a result can currently be computed on. |
A third count would be the honest one, and it is not yet available: feeds that have passed the acceptance protocol further down this page. No feed currently in the store has been through it, because the protocol was written after the stores were built, which is precisely how the register below came to have seven entries in it. We state that instead of quietly reporting the larger of the two numbers.
Every number that appears on more than one page of this site is defined once, on its home page, with its definition and its census date, and cited from elsewhere with a link. Estate counts live here. Funnel counts live on the cover. Corpus counts live on the reading list. A number restated on a second page drifts; the twenty-nine-against-twenty-one discrepancy this section resolves was produced exactly that way, and the rule exists so it cannot recur silently.
| Asset class | Composition held | Instruments |
|---|---|---|
| Equities — Europe | Cash equities | 30,455 |
| Equities — North America | Cash equities | 16,344 |
| Equities — Asia-Pacific | Cash equities | 12,213 |
| Foreign exchange | Spot crosses, majors through exotics | 1,983 |
| Crypto | Perpetual futures and spot | 846 |
| Listed futures | Individual contracts, plus validated continuous roll series | 682 |
| Listed options | Full chains with greeks, implied vol and open interest | 18 chains |
| Index reference | One reference family back to 1990 | 4,134 |
| Total | 21 exchanges + two derivative venues | ~66,700 |
The stores are refreshed nightly. The raw vendor mirror was retired in 2026 and the stored bars are now the record; re-downloading from a vendor is a deliberate four-case procedure. That decision is defensible and it is also one reason the register below matters: when the bars are the record, a defect in the bars is a defect in the evidence.
Coverage and access are different things, and conflating them is how research becomes untradeable. Every market above is priced through a venue-specific cost model (commissions, exchange fees, transaction taxes, borrow, and the currency leg) before any result is believed.
| Constraint class | How it is handled |
|---|---|
| Venues with unrestricted two-way access | Tradeable. Transaction taxes decide which mechanics survive: a purchase-side stamp duty of tens of basis points rules out anything with high turnover, and the honest answer is to design around it, not to ignore it in the cost line. |
| Venues with administered daily price limits | Tradeable long-only, with next-day settlement. A limit-locked session is modelled as no fill, not as a fill at the limit. A backtest that fills at the limit is trading against a queue that was not there. |
| Crypto | Sandbox execution against live market data, on the production code path. Real-money access is pending a regulatory restriction outside our control. The sandbox is not a rehearsal environment; it is the production path with the venue's own fills withheld. |
| Any book with an unhedged currency leg | Blocked pending its hedged arm. An apparent cross-listing edge measured on this estate turned out to be an uncontrolled currency basis and died when the hedge was applied. A hedged arm is now mandatory before a forward slot, not after one. |
Every defect below was found in our own store, by us, and each one had already cost research time before it was named. We publish them because a coverage page that lists breadth and not defects is exactly the selection-bias exhibit this page warns about elsewhere.
Five of the seven entries are not data errors in the ordinary sense. They are components that reported success while doing nothing: a store that silently dropped a series instead of retaining it with a terminal date, an empty table that reads as "no corporate actions" when the truth is "not loaded", a membership table with zero rows that reads as "no index changes" when the truth is "never loaded", a streamer that refreshed a recent window and called that a refresh, an archiver that died inside its own error-logging path. Silent-zero is the dominant failure mode of an estate like this one, and it is why the engineering is fail-loud. The general treatment is in what the machines got wrong; the store-specific treatment is in the working paper on data defects.
| Defect | How it surfaced, and what it invalidated | Status |
|---|---|---|
| Survivorship absence2026-07-31 | One national equity store held no delistings at all across more than four thousand tickers over four years. That is not survivorship bias, it is survivorship absence: every series that stopped existing was silently dropped rather than retained with a terminal date and reason. It surfaced through a merger-arbitrage screen in which none of 26 price-locked candidate episodes ever stopped printing, which is arithmetically impossible in a real market. It independently blocks merger arbitrage in court-sanctioned acquisition structures, index deletions, tax-loss reversal, and every cross-sectional book on that market, because a universe that cannot lose members overstates every backtest run on it. | Open Moratorium on new books |
| Distributions unadjusted2026-07-31 | The corporate-action table is empty, so bars are unadjusted. This does not add noise, it manufactures results: a screen on an income-security class returned 111 trades that cleared every conventional test statistic, and was entirely the distribution sawtooth; 65% of entries fell within five days of an ex-distribution drop against a 17% base rate. On a distribution-adjusted series the same rule fires three times in four years, and loses when it does. | Open Feed specified for purchase |
| Splits unadjusted2026-07-31 | Eleven unadjusted split events in a single venue's bars, and four more in a second store of the same market. A pair search returned a top-ranked winner that was entirely four unadjusted splits; on split-adjusted data the same instrument's null bar sits at approximately zero. The defect cuts both ways, which is the expensive direction: a single unadjusted forward split understated one instrument's multi-year compound growth by a multiple, not a margin, which is enough to invert a ranking and turn a pass into a kill. The instrument and both growth figures are withheld; the direction of the error is not. | Contained Back-adjust at read is mandatory; feed still unadjusted |
| Permanent hole in four stores2026-08-02 | A vendor streamer that refreshes only a recent window turned a three-week upstream outage into a permanent 15-session hole across four stores, 13,079 active symbols, with nothing watching for it. It surfaced when a study's out-of-sample window silently lost sessions and its dispersion statistic moved on their removal. Any study on those stores crossing that window is provisional until re-run. | Repaired Scheduled doctor + backfill; re-runs outstanding |
| Archiver died in its own error path2026-08-02 | An event archiver died on its first reconnect because of a malformed argument inside its own error-logging call (the fail-loud loop killed itself), and then lay dead for three and a half days under a supervisor that only ran at boot. The lost days cannot be reconstructed after the fact: it is a live-capture archive, not a queryable history. | Closed Fixed and guarded by a liveness loop |
| No security master2026-07-31 | The store holds tickers and bars only; there is no issuer identity, no international identifier, no security master anywhere. It surfaced while building a cross-listing registry, where identity had to be inferred statistically from price behaviour because there was no field to read it from. It blocks verified enumeration of cross-listings, and it blocks the classification that decides tradeability: the distinction between fungible interlisted lines, depositary receipts, and dual-listed structures. | Open Specified for purchase |
| No index membership history2026-07-31 | Zero membership rows, so an index-rebalance study, one of the cleanest mandated-flow events available anywhere, had to reconstruct its events from volume spikes as a proxy. The resulting test landed at t = 1.90 on n = 15, with a control that beat the treatment at three of four horizons. That is the difference between an event study and a proxy study, and the proxy study is not evidence. | Open Specified for purchase |
The pattern across the register is worth stating plainly, because it changed our order of operations: the defects were found by instruments we built to watch for them, not by anyone noticing an odd number. Two full market screens were spent refuting artefacts the store itself had produced. That converts research capacity into refutation capacity, and the correct response is a moratorium on new books in the affected market, not a smarter screen.
Feed acceptance protocol
The reusable part of the above is the rule, not the incident list. A vendor feed adopted without calibration does not remove the silent-zero risk; it relocates it, into a system where nobody is looking any more because the data is now "good".
A coverage list that only shows productive markets is a selection-bias exhibit. Ours is mostly negative, and the negatives are the part we are most confident in.
US equities is the clearest of them: a full campaign across the liquid universe (cointegration and tracking families, cross-sectional momentum and mean reversion, breakout, share-class pairs) with matched control ensembles per candidate, produced no book that passed both the survivability and specificity gates at real per-share costs. The planted control in that campaign was correctly refused, which is the reading that matters: the instrument returned a zero where a zero was the truth. Those gates were retired on 2026-08-08; the correction of record is on the method page. The verdict is kept as that closed era's published finding. It has not been restated under the current protocol.
The broader sweep that followed screened eight markets for mechanisms with a named obligated party and a dated or conditional trigger. Seven of the eight returned nothing that survives cost; measured residuals sat below the round-trip cost at every venue screened in those seven. One returned a single candidate that survives its cost model, and the house labels it a risk premium rather than a dislocation, which is the weaker of the two claims and the one the evidence supports. We are not going to enumerate which markets those were: the set of markets a firm screens is live information even when every individual result in it is dead.
The negatives are what moved the search. A tradeable structural relationship needs an enforcer: an agent contractually or legally obliged to push prices back together, whose obligation does not depend on the relationship having held before. The classes we recognise are a prospectus rebalance mandate, a defended band with a formal intervention obligation, a share-class legal identity, contract convergence to settlement, and an index tracking mandate. Statistical co-movement has no enforcer; its persistence is a bet that a past correlation regime continues, and that is exactly the bet a composition null prices.
The cross-market gradient is ordinal and it is stark. Across three asset classes (lettered A, B and C, with names withheld for the reason given two paragraphs above), statistical cointegration persists well above its own random baseline in A, above that baseline by a smaller margin in B, and statistically indistinguishably from chance in C. Statistical relationships are therefore abundant in one asset class, present in a second, and arbitraged away in a third. Enforced relationships survive in all three, because arbitraging one away requires the enforcer to stop enforcing. The full treatment, with the figure, is the 2026-06 working paper Enforced structure ports. Statistical structure does not. The ordering and its direction are published; the classes are lettered rather than named, and the hit-rates, their random baselines and every ratio between them are calibration of our own screen and are withheld.
Two honest limits sit under that. Enforcement can end (pegs break, prospectuses change, share classes collapse), so enforced books carry a discontinuity risk that in-sample statistics are structurally blind to, and every enforced book we run registers a same-day kill trigger on its own mechanism. And our search has a detection floor, which this year we measured for the first time: at the dislocation scales real contractual obligations actually offer, a pair search finds them somewhere between rarely and never. The direction and the affected classes are published; the floor's value is a calibration of our own instrument and is withheld. The operating consequence is a rule rather than a disclaimer: a kill verdict issued without a stated minimum detectable edge is not a finding, and is now refused as one.