We say this sentence in every letter we send. A sentence like that is either an argument or a boast. This page is the argument. It defines both words, states the tests, shows the evidence, and lists what we do not claim, so the adjective is yours to check rather than ours to assert.
A natural problem is a law of markets: non-stationarity, capacity, crowding, alpha decay, the efficient-markets limit. No mechanism repeals them. A structural problem is created by an institutional arrangement and can be removed by a different one. Solvable means structural. The test: does the problem survive a change of mechanism? If yes, it is natural and outside this claim.
Among structural problems, size is measured on three axes a reader can check: dependency, how much of the field's machinery takes this problem's output as input; duration, how long the field's own literature has documented it as open; price, the capital screened on the unverified number. The biggest is the one that dominates on all three.
We tried to find a structural problem that sits upstream of the trial count. We could not. Every other one either already had a partial fix, or inherits from this one.
| Structural problem | Prior fix existed? | Upstream of the trial count? | What it inherits |
|---|---|---|---|
| The unverifiable trial count. Every overfitting correction divides by N, how many strategies were tried, and N was declared by the party being judged. | NO · 46 years | IT IS THE ROOT | Deflated Sharpe, the Harvey–Liu haircuts, PBO, SPA, minimum track length: all take N or the trial set as input. |
| Survivorship bias in data | Partial · point-in-time datasets | No | A clean dataset with a fabricated N still fails the correction. |
| Look-ahead bias in backtests | Partial · tooling, purged CV | No | Detectable only if the trials are committed; otherwise it is one more undeclared trial. |
| Cost and slippage misstatement | Partial · cost models | No | A cost model is one parameter of the search; it enters N. |
| Reported-return manipulation | Yes · audit, since 1933 | No | Audit verifies the statement, not the search that produced the strategy. |
| Process opacity in reporting | Yes · GIPS, 1999 | No | GIPS verifies process, and is silent on backtests and trial counts by design. |
| Disclosure destroys the edge | No | Same problem, other face | This is the deadlock that kept the trial count unverifiable: verify me, but I cannot show you. |
| Allocation by pedigree | No | No · a symptom | The substitute the industry adopted because the number could not be checked. |
Every multiple-testing correction in the literature takes the trial count or the trial set as input. Bailey and López de Prado's deflated Sharpe ratio, Harvey and Liu's haircuts, the probability of backtest overfitting, Hansen's SPA test, the minimum track-record length. None of them can be computed honestly without N. No other structural problem is an input to all of them. The mechanism, in full →
Grossman and Stiglitz stated the disclosure deadlock in 1980. Bailey and López de Prado formalised the correction in 2014 and, in the Notices of the American Mathematical Society, called failure to report the trial count a form of fraud. Harvey's 2017 AFA presidential address said most claimed findings are likely false. Bollen and Pool found the fingerprint of massaged returns in the Journal of Finance. Forty six years, five journals of record, never closed. Every citation, sourced →
Roughly $8 trillion of systematic capital is screened on backtested numbers whose trial count cannot be checked. If those numbers overstate by one percent, that is $80 billion a year. The assumption is stated so a sceptic can turn it down; at a quarter of a percent it is still $20 billion. The industry already pays for weaker substitutes: operational due diligence at $25K to $75K a fund, GIPS verification at $10K to $50K a year, none of which can check N. The price, in full →
It stayed open because the two halves were jointly unsatisfiable by any disclosure-based mechanism: to verify the search you had to see it, and a seen strategy stops being an edge. Every prior fix picked a side. Manual audit, GIPS, platform custody, NDA disclosure, self-reported databases, pedigree: what each verifies and what each structurally cannot →
It is closed because the primitive that dissolves the pair, a proof that recomputes the statistics over a committed trial set without revealing it, became production-grade about two years ago. The trial ledger is committed as a Merkle tree before evaluation. The circuit reads its length. N stops being a declaration and becomes a count. VTR-1, the frozen specification →
Closed means the impossibility is gone, not that the industry has switched. The leaf count binds only under register-before-run: the circuit proves the shown result is the maximum of the committed set. It cannot prove that no trials were run off-ledger. That boundary is stated in VTR-1 §3.5, on the credential, and here. A claim that hid it would be the thing this standard exists to abolish.
Name a structural problem in quantitative finance whose output does not enter the trial count, and which the trial count does not enter. If it exists, the word "biggest" is wrong and we will change it on this page, dated. We could not find one. Send it to [email protected].
Name a mechanism, between 1980 and 2024, that verified the trial count without disclosure and without custody. Audit, GIPS, platforms, NDAs and databases each fail one of the two. If you find one, "solvable but unsolved for 46 years" is wrong and we will say so.
Download the verifier and a credential. Recompute the data root from your own CSV. Watch the trial count come from the ledger, never from a field. About 81 milliseconds. Verify it yourself →
The sentence, restated with its argument attached: the unverifiable trial count was the one deadlock in the field's own literature that was structural rather than natural, upstream of every overfitting correction, documented for forty six years, and priced in the tens of billions a year. It is closed. The proof is public.