How to Backtest a Factor Model: A 2026 Guide

A rigorous, sober how-to on backtesting factor models: point-in-time data, survivorship handling, the GRS test, minimum observations, EW vs RP weighting, and composite Sharpe+Info-Ratio ranking.

Category: Factor Research

What backtesting a factor model actually requires

Backtesting a factor model means measuring whether a factor's premium survives honest scrutiny. Build point-in-time, survivorship-handled returns; form long sleeves; test joint significance with the GRS test; enforce a minimum-observation floor; compare equal-weight versus risk-parity weighting; then rank sleeves by a composite of excess Sharpe and information ratio. When citing SurgeFlow's own Phase 1 results, note they are research_backtest_insample — an in-sample factor-attribution decomposition, not a live out-of-sample track record; the GRS test and diagnostics are reported, not used as portfolio exclusion gates.

Most retail backtests skip the integrity gates and report a flattering number. The premium they 'find' is usually an artifact of look-ahead bias, dead companies silently dropped from the sample, or too few observations to distinguish skill from noise. A defensible backtest is mostly plumbing: getting the data and the gates right before any return is computed. The rest of this page walks through that plumbing and the formulas behind it, then shows one worked example from SurgeFlow Factor Backtest V2.

The integrity gates retail backtests skip

Four gates do most of the work, and each one independently kills a large share of naive results.

Point-in-time data: every input used to rank a stock on a given date must have been knowable on that date. Fundamentals carry a filing or publication date, not just a fiscal-period date, so a profitability or value rank for January cannot quietly use figures only released in March. Violating this is the most common source of inflated backtest performance.

Survivorship handling: the universe on each historical date must include companies that later delisted, merged, or went bankrupt. If you backtest only today's survivors, you have removed exactly the names that hurt, and the result is biased upward by construction.

Minimum observations: a Sharpe ratio computed on a short window is mostly sampling noise. A floor on the number of return observations per sleeve prevents short-lived or thinly-covered sleeves from being ranked against long-history ones on equal footing.

GRS test: when a model has several factors and several test portfolios, you need a joint test that the pricing errors (alphas) are collectively zero, not a cherry-picked single t-stat. - Point-in-time inputs keyed to filing/publication date, not fiscal date - Universe includes delisted and merged names on each historical date - A minimum-observation floor before any sleeve is ranked - A joint GRS test rather than a single hand-picked t-statistic

Methodology and formulas

Factor returns and cross-sectional pricing. Fama-MacBeth (1973) is the standard two-pass approach: run a cross-sectional regression of returns on factor exposures each period, then average the period-by-period slope estimates to get the factor premium and its time-series standard error. SurgeFlow's factor layer follows the Fama-French style construction (SMB size, HML value, RMW profitability, CMA investment, WML momentum, LIQ liquidity) and produces daily factor returns.

Joint significance: the GRS test. The Gibbons, Ross & Shanken (1989) statistic tests whether the intercepts (alphas) from a set of asset regressions on the factors are jointly zero. A model that explains returns well should leave alphas that are statistically indistinguishable from zero across all test assets at once; the GRS test is the F-test that formalizes this.

Weighting: EW vs RP. Equal weight (EW) assigns the same notional to each name in a sleeve. Risk parity (RP) instead scales each position so that names contribute roughly equal risk, down-weighting the most volatile constituents. RP often improves risk-adjusted metrics precisely because it stops a few high-volatility names from dominating the sleeve's variance.

Ranking: the z-score composite. To compare sleeves on more than one axis, standardize each metric across sleeves to a z-score and blend. SurgeFlow ranks the US sleeves by argmax of 0.5*z(Sharpe) + 0.5*z(InformationRatio), where Sharpe uses excess returns and the information ratio is annualized against the benchmark. This rewards sleeves that are strong on both absolute risk-adjusted return and benchmark-relative consistency, not just one.

A worked example: the WML+RMW Top-30 RP winner

SurgeFlow Factor Backtest V2 runs long sleeves of single and combination factors under both EW and RP weighting, reporting per-sleeve excess Sharpe, annualized information ratio, maximum drawdown, turnover, and the number of observations. Every sleeve is reported with data-quality checks and diagnostics: point-in-time controls, survivorship handling, minimum-observation status, and the GRS test. GRS is reported as a diagnostic, not used as a sleeve-exclusion gate; the current Phase 1 run is research_backtest_insample rather than live out-of-sample rolling-signal evidence.

Under the composite ranking, the current US winner is the WML+RMW Top-30 risk-parity sleeve, computed over roughly ten years. It is #1 by both excess Sharpe and information ratio, which is why it survives the composite rather than winning on one metric and losing on the other. We deliberately do not quote its Sharpe or premium as a fixed number here; those values are live and move with the data. See the current sleeve table and metrics on /backtesting, and the underlying factor models on /factors.

What SurgeFlow data this uses

This methodology runs on SurgeFlow's multi-market equity research data, not on any advisory product. The factor layer is built in the Fama-French style (SMB, HML, RMW, CMA, WML, LIQ) with daily factor returns, a publication gate that blocks unpublishable factors, an effective-direction resolver, and VIF multicollinearity diagnostics. Factor Backtest V2 consumes this layer to form EW and RP sleeves and to report the data-quality gates and diagnostics described above.

Underlying price and fundamental data spans four markets with documented sources: FMP for US; Tushare for China A-shares and Hong Kong; and J-Quants plus EDINET large-holding reports for Japan. SEC EDGAR 13F filings feed US institutional ownership. The worked example above is the US sleeve; coverage and model maturity differ by market. - Factor layer: SMB, HML, RMW, CMA, WML, LIQ with daily returns and a publication gate - Backtest V2: EW + RP sleeves; data-quality gates = point-in-time / survivorship / min-obs; GRS + diagnostics reported, not gates - Sources: FMP (US), Tushare (CN/HK), J-Quants+EDINET (JP), SEC 13F (US)

Cross-market caveats

A factor model that works in one market does not transfer automatically. The academic record is explicit on this: for China A-shares, Liu, Stambaugh and Yuan (2019, 'Size and value in China', Journal of Financial Economics) document that size and value premia are strong while price momentum is weak or absent, and they recommend trimming the smallest shell-stock cohort because of reverse-merger distortions. A momentum sleeve validated in the US should therefore not be assumed to hold in China without its own gated backtest.

This is a data-science and methodology resource, not investment advice. Nothing here is a recommendation to buy, sell, or hold any security, and no result should be read as a prediction. The honest output of a factor backtest is a measurement with confidence bounds and a list of the assumptions it depends on, not a forecast.

Run it yourself with the free API

You can reproduce this workflow against live data. The free public API covers all four markets at 2,000 requests per day, with realtime, screen and hotlist endpoints plus a summary route, an OpenAPI spec at /api/public/openapi.json, and an AI index at /llms.txt. Create a free API key at /account and open the live console at /membership/api, then inspect the current sleeve rankings on /backtesting and the factor models that feed them on /factors. Start with the US winner, then re-run the same gates on a market where the premia are known to differ.

Sources and further reading

Primary, academic, and SurgeFlow data references for the claims above: - Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. Journal of Political Economy. - Gibbons, M. R., Ross, S. A., & Shanken, J. (1989). A test of the efficiency of a given portfolio. Econometrica. - Fama, E. F., & French, K. R. (2015). A five-factor asset pricing model. Journal of Financial Economics. - Liu, J., Stambaugh, R. F., & Yuan, Y. (2019). Size and value in China. Journal of Financial Economics.