研究誠信：本站對自己設下的標準

SurgeFlow 的五項公開承諾——只做研究不給投資建議、證據誠實、資料溯源、方法透明、AI 輸出必經查核——以明文寫下，讓讀者得以檢驗與監督本站。

Category: Data Quality

What research integrity means, and why a data site needs it

Research integrity is the discipline of never letting a claim outrun its evidence. In laboratory science it shows up as pre-registration, replication, and the honest reporting of negative results. On a market-data site it takes a more mundane shape: every number should trace to a named source, every model output should be labeled by how it was validated, gaps in the data should be visible rather than papered over, and descriptive language should never quietly harden into a recommendation. The pressures pushing against this are constant and familiar — confident copy reads better than hedged copy, smooth charts look better than gappy ones, and a statistic that survived many silent retries looks exactly like one that worked the first time. A standard that lives only in a team's heads is a preference, not a standard: it cannot be checked, so it cannot be trusted. This page turns SurgeFlow's internal working rules into explicit public commitments, so readers can verify them against what the site actually publishes — and call out the failures.

The mechanics: integrity as arithmetic

A surprising amount of research integrity reduces to small pieces of arithmetic, applied without exception.

Multiple testing. If a process runs N independent statistical tests, each at significance level alpha, the expected number of false positives is E[FP] = N * alpha, where N is the number of tests tried and alpha is the probability that a single test looks significant by pure chance. Run a hundred screens at alpha = 0.05 and about five will look like discoveries even if nothing real exists. Any pipeline that searches across many factors, parameters, or screens must therefore discount what it finds — or disclose the search.

In-sample bias. Write D_train for the data a model was fitted on and D_test for data it never saw. Performance measured on D_train is biased upward, because the fitting procedure has already adapted to that sample's noise; only performance on a disjoint D_test estimates behavior on genuinely new data. The label “in-sample” is not a technicality — it changes what a number means.

Numeric grounding. A factual claim about a figure can be checked mechanically: accept x_claimed only if |x_claimed - x_source| / max(|x_source|, e) <= t, where x_claimed is the number appearing in the text, x_source is the value in the underlying dataset, e is a small constant that prevents division by zero, and t is the tolerated relative error.

Fail-closed. The grounding rule generalizes to a publication policy: publish(claim) = verified(claim). If verification returns unknown, the claim is treated as unpublishable. A fail-open system publishes unless proven wrong; a fail-closed system publishes only what it can prove, accepting that some true statements will be withheld as the price of bounding the false ones.

How SurgeFlow applies it: five public commitments

SurgeFlow's working rules are codified as the five commitments below. One honesty note first: these are enforced constraints, not a certificate of correctness. The gates catch the classes of failure they were built to catch — target-price wording, ungrounded numbers, unlabeled in-sample results — and nothing more. Upstream vendor errors can still pass through, and human-written pages are held to the same rules by review rather than by code. The commitments: - Research, never investment advice. No price targets and no buy/sell/hold language anywhere on the site, including AI-generated reviews — a deterministic gate scans model output and blocks target-price wording before publication. - Evidence honesty. Cite evidence when it exists; neutralize the claim when it does not; fail closed when verification is impossible. In-sample results are always labeled in-sample. - Provenance. Every dataset names its source and its update cadence. Missing data is shown as missing — it is never silently imputed to make a chart look complete. - Methodology transparency. The formulas, screens, and gates behind the site's analytics are published in this appendix, so results can be reproduced or disputed. - AI-output grounding. Claims made by the analyst panel pass numeric and semantic grounding checks against the underlying data before they are published.

What SurgeFlow data this uses

These commitments apply uniformly across the eight markets SurgeFlow covers — the United States, China, Japan, Hong Kong, Taiwan, Korea, the United Kingdom, and India. Provenance is the concrete face of the standard: each market's data names its supplier, and each supplier has its own conventions and failure modes, which is precisely why the source must stay attached to the number. Where a source lacks a field for a given company, that field renders as missing on the site — the gap itself is information about the source. - US, UK, and India: fundamentals and prices via Financial Modeling Prep (FMP). - China and Hong Kong: via Tushare. - Japan: market data via J-Quants, with filings from EDINET. - Taiwan: prices via FMP and yfinance, with filings from TWSE/MOPS. - Korea: filings via DART.

Why fail-closed: the asymmetry of research errors

Why fail-closed rather than fail-open? Because the two error types are not symmetric. A true statement that goes unpublished costs a reader some insight; a false statement that gets published can silently shape decisions, and on a financial-data site the reader may act long before anyone notices the error. Fail-closed bounds the damage of the worst case at the cost of the average case. The same asymmetry explains the in-sample labeling rule. The literature on backtest overfitting shows how reliably a search over enough configurations produces an impressive historical curve, and how little that curve says about the future. SurgeFlow cannot make overfitting impossible — no one can — but it can refuse the one move that makes overfitting invisible: presenting an in-sample result in language that implies it was a prediction.

Explore it yourself

Every claim on this page is checkable against the same data the site uses. A free (beta) API key is available at /membership, with a limit of 2,000 requests per day, and the full interface is documented in the OpenAPI specification at /api/public/openapi.json. The data is provided for research, not as signals: query the raw series behind a page, recompute a figure, and compare it with what the site displays. If they disagree, that is exactly the kind of failure these standards exist to surface.

Sources and further reading

Primary, academic, and SurgeFlow data references for the claims above: - Ioannidis, J. P. A. (2005). “Why Most Published Research Findings Are False.” PLoS Medicine. - Leamer, E. E. (1983). “Let's Take the Con Out of Econometrics.” American Economic Review. - Lo, A. W., and MacKinlay, A. C. (1990). “Data-Snooping Biases in Tests of Financial Asset Pricing Models.” Review of Financial Studies. - White, H. (2000). “A Reality Check for Data Snooping.” Econometrica. - Harvey, C. R., Liu, Y., and Zhu, H. (2016). “… and the Cross-Section of Expected Returns.” Review of Financial Studies. - Bailey, D. H., Borwein, J. M., López de Prado, M., and Zhu, Q. J. (2014). “Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance.” Notices of the American Mathematical Society. - SurgeFlow routes: /research (research library), /methodology/factor-research (factor methodology and gates), /ai (analyst panel), /membership (API access).