The preprint says the safeguards intended to protect a systematic strategy from difficult conditions can consume the statistical room needed to choose and size a trade. In its end-to-end test, a search across 72 candidates was over budget at the measured cost of comparing block scores. The combined search haircut and dispersion penalty exhausted the selected rule's reported edge, so stage S5 returned no position.
That result has a narrow reading. The paper does not establish that its architecture produces profit or make a general claim about markets. It describes one committed walk on one index series under its declared rules for robust selection. It is a test of the procedure, not a market-wide verdict.
The objective is derived from the assumptions
The research question is whether the accumulated rules of systematic trading amount to a complete and forced architecture. The framework begins with a latent state, meaning an underlying condition that is not directly observed, and puts limits on how much that state can change, how often difficult conditions can recur, how long signals remain coherent, how large a signal may be and how stable the representation must remain. In the paper's terminology, these limits are an invariance defect, a recurrence bound, coherence times, a signal ceiling and an invariance ratio.
From those premises, the selection objective is derived rather than chosen as a preference. Conditional deployment loss is bounded by the average loss in the worst historical tail of block risks. That tail average is called conditional value at risk, or CVaR. Its tail fraction is set by the declared recurrence bound, one divided by Lambda. In ordinary terms, the method tells a researcher to judge a strategy by the blocks in which it fares worst, provided the assumptions hold at the declared scale.
The same framework treats model complexity, candidate searches and other data-dependent research choices as claims on one information budget. That budget is tied to the effective sample and the declared signal ceiling, and it includes a charge for how far the final choice moves from a pre-data prior. The paper measures the budget in nats. A larger search therefore leaves less room for other choices.
Tail selection adds a separate price. For independent block risks, the tail-price factor c(Lambda) reduces the effective sample available for searching by the square of that factor, so the permitted search size falls as the factor rises. A second result shows that mean-zero noise in block scores cannot make their tail average lower than the tail average of the underlying block risk. The authors present this as a mathematical comparison, not a measured market effect.
The same budget shapes position sizing. Under the paper's Gaussian and quadratic approximation, the allowed Kelly fraction, the share of a growth-optimal position, is bounded by one divided by one plus the fraction of capacity already spent. At full budget, that ceiling is half-Kelly. It is a sizing bound within the stated approximation, not a profitability result.
The diagnostics were tougher than the baseline
The real-series diagnostics gave the tail price a concrete scale. In the index experiment, the reported median estimate of c-hat at tail level 4 was 1.95, compared with 1.43 under a Gaussian reference. Across 20 assets, raw c-hat at the same level exceeded the Gaussian reference on 19, while the studentized values did not show the same excess. The reported effective-sample ratio had a median of 0.86, with a value of 3.3 in a stale-NAV high-yield fund case. Those estimates use noisy block scores and are not direct measurements of the latent block-risk law.
The diagnostics formed part of six committed real-market experiments using fixed-seed scripts and recorded outputs. The paper uses those tests to challenge declarations at their tested levels. A declaration that survives such a test is not treated as confirmation of the underlying axiom.
What the theorem does and does not settle
The broader theorem makes a stronger internal claim. Each listed stage is necessary under the framework: for every stage, the paper constructs an axioms-compatible law in which leaving that stage out performs strictly worse. The result supports stage necessity, not one unique implementation or a theorem that every good procedure must factorize in the same way.
The main caveat is that Lambda is not estimated from the past. It is treated as a premise about future paths. Historical data can provide contradiction diagnostics for a declaration, but cannot turn Lambda into an observed parameter. In the reported analyses, the authors say tests contradicted declarations rather than the framework: no axiom was refuted at the stated honest level, while kappa=1 and a declared coherence time of 60 were rejected. The reported volatility-coordinate rank was 138.
The preprint offers a conditional account of how a trading system should spend information, search and sizing capacity. It does not show that the canonical form is profitable, and its empirical record does not settle what would happen in other markets or beyond the one end-to-end index walk. The document identifies itself as arXiv version 3, dated 20 September 2026.
Paper data and sources
Original title: The Axiomatic Trader: Latent Regularity, Information Budgets, and the Canonical Form of a Quantitative Investment System
Authors: Jiayu Li
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text