Praxis · methodology
MarketGraphs Working Paper II · The Praxis Engine

How a claim becomes a verdict.

“The first principle is that you must not fool yourself — and you are the easiest person to fool.”— R. P. Feynman, Caltech commencement, 1974

Abstract

The pipeline behind every Praxis verdict: a crawler reduces the research firehose to finance-relevant papers (§1), extraction distils each into a structured spec (§2), the consequential ones are rebuilt end-to-end under documented deviations, known-value tests and negative controls (§3), then pushed through the C0–C7 gate ladder (§4) to a fixed-function verdict (§5). Nothing here forecasts. The engine describes what a published rule did: out of sample, net of costs.

1 · The corpus§1 2 · Extraction§2 3 · Reconstruction§3 4 · The gate suite§4 5 · The verdict§5 6 · Access§6 References·

1The corpus

Everything downstream depends on knowing what exists. The crawler runs daily across the open research firehose: 415,961 documents scanned, 42,984 strategy papers mapped, ≈480 new every month.

Figure 1. Finance-relevant papers per month. The field is accelerating, from a ~320/mo steady state toward ~480. January spikes are annual date-backfill (arXiv/SSRN year-only stamps), shown honestly and clipped at 650.
Figure 2a. Economic bets, top 12 of 74 ≈ title-based. Area ∝ papers, relative density.
Figure 2b. Computation, coloured by family ≈ title-based: classical · neural · RL · LLM.
Figure 3. Granular economic bets × asset class: the twelve most-published of 74 bets, title-tagged across 43,018 papers ≈ title-based. Fill ∝ papers, multi-label by design. A dashed cell is fewer than three.

2Extraction

A paper is only comparable to another paper once both are the same shape.

Definition 1 (Structured spec).

A paper reduced to fixed pillars: asset · data · predictor · method · universe · validation. Each field is either quoted from the text or marked undisclosed. 143 papers carry a full spec to date.

The pillars are the engine's coordinate system: they make every spec searchable over MCP, rank which claims deserve a week of reproduction, and expose the white space no single paper can see.

Figure 4. Asset × method coverage, a three-row excerpt of the corpus map, title-tagged across 43,018 papers. Fill ∝ papers. Dashed cells are combinations no paper has run. The full grid, the recombination map, and the white space they expose ride with the desk. Request a seat ↗

3Reconstruction

The rule, rebuilt from the paper's own description and nothing else.

Each reproduction is built from the published text alone. Where the paper is silent, the gap is filled with a documented estimate and flagged as such; where its data cannot be sourced, the closest survivorship-clean proxy is substituted and recorded. The deviation ledger ships with every report: the reader never has to trust us on what changed.

Algorithm 1. Faithful re-run, expanded
  1. extract the spec · quote every disclosed parameter
  2. source survivorship-clean reference data · record the vintage
  3. rebuild the strategy · fill undisclosed gaps, flag each
  4. test every block against known values · run negative controls
  5. validate through C0–C7 · publish claimed vs. measured

The controls are adversarial by construction. A look-ahead trap feeds tomorrow's data through the pipeline: if the honest run were peeking it would light up (on the network-momentum paper the trap explodes the Sharpe to 22.3; the honest run reads 0.62). A signal shuffle scrambles the predictor and the edge must collapse to zero (it does: 0.01). Closed-form cases pin the numerics.

# deviation ledger · 361154 · excerpt data: Pinnacle 64 ratio-adjusted futures (1990–2022) -> the project's roll-adjusted reference panel, 40 contracts window: paper OOS 2000–2022 -> 2020–2026 (the paper's pre-2010 window isn't publicly reproducible) solver: CVXPY on MOSEK (commercial) -> CLARABEL (open) cross-checked to <1e-2 · exact on the 2-node form controls: peek trap SR 22.3 · shuffle SR 0.01 · honest 0.62

4The gate suite

Eight gates, fixed before any paper is run. A claim that survives all eight has nowhere left to hide.

Table 1. The gate suite: eight fixed tests, set before any paper is run.
GateWhat it tests
C0ReplicationFaithful re-implementation on the paper's own period, universe (or closest free proxy), and cost assumptions. Does the mechanism exist at all?
C1Trading costsRealistic transaction costs, fees and implementation frictions applied: the numbers the paper should have printed.
C2SignificanceStatistical reality: multiple-testing corrections (Bonferroni / Holm / BHY across the paper's own configuration grid) and the Deflated Sharpe Ratio [1].
C3Crisis regimesThe worst regimes: 2008, Volmageddon 2018, COVID 2020, the 2022 rate shock. Advertised protection is tested, not assumed.
C4GeneralisationOther universes, countries and asset classes, where the claim implies them.
C5CapacityCapacity, borrow and tradability stress, with cost sweeps to breakeven.
C6PersistenceTrue post-publication out-of-sample: the window the authors could not have fit.
C7OriginalitySpanning regressions on standard factor models (HAC / Newey–West errors): alpha, or relabelled beta [3]? Probability of Backtest Overfitting via CSCV assessed alongside [2].

4.1The deflation arithmetic

Gate C2 is where most claims die. The engine treats a reported Sharpe as the maximum of a search the authors ran, disclosed or not, and asks what survives once that search is priced in.

deflated Sharpe (Bailey–López de Prado): the claim, haircut for luck · gate C2
probabilistic Sharpe: P(true SR > benchmark) given skew & tails · gate C2
expected max Sharpe under N trials: the overfit baseline any search must beat · gate C2
minimum track-record length: how much history a claim needs before it means anything · gate C2
alpha decay & half-life: how fast the edge dies post-publication · gate C6
probability of backtest overfitting (CSCV): is the in-sample winner an out-of-sample loser? · gate C7

5The verdict

A fixed function, applied after the gates. Never an opinion.

(1)
Table 2. The verdict: a fixed function of measured against claimed, Eq. (1).
VerdictMeaning
validatedThe claim replicates cleanly on honest data.
overstatedThe effect is real but materially smaller than claimed.
failsThe claim does not survive honest reproduction.

Across the audits to date, the median Sharpe haircut on Sharpe-denominated claims is ≈ 60% of the published number, before anyone traded a dollar against it.


6Access

Every verdict is public. Readers open ten exemplar reproductions in full (gates, deviation ledgers, notebooks) and query them over MCP where they already work. The full corpus rides with the desk. The engine never emits position-level output.

$ subscribe --tier reader

References

[1]Bailey, D. & López de Prado, M. The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality. J. Portfolio Management, 2014.
[2]Bailey, Borwein, López de Prado & Zhu. The Probability of Backtest Overfitting. J. Computational Finance, 2017.
[3]Harvey, C., Liu, Y. & Zhu, H. …and the Cross-Section of Expected Returns. Review of Financial Studies, 2016.
[4]Ioannidis, J. Why Most Published Research Findings Are False. PLoS Medicine, 2005.
[5]López de Prado, M. Advances in Financial Machine Learning. Wiley, 2018.
← paper I · the findings © MMXXVI MarketGraphs · descriptive research, not investment advice typeset in the browser