This project computes 40+ distinct scores (momentum, value ratios, quality, insider consistency, smart-money composites, macro/rate indicators, and more) and serves every one the instant it computes without error. Computing a score is not the same as it predicting anything - so each candidate goes through the same walk-forward validation gates before it's called "cleared": cross-sectional (or, for macro series, single-time-series) decile sorts, regime-stability splits including the 2020/2022 stress years, out-of-sample walk-forward testing, and a composite quality scorecard. Twenty factor constructions have been through this so far; nine cleared the bar - one at our strict "core factor" tier, eight as regime-dependent components. Sorted below by fold-to-fold sign consistency, our own gating statistic.
Purchase-only (sales excluded), trailing 90-day window - our largest real evidence base (200,522 observations, 2,131 companies) and the first result in our whole validation history to clear the strict "core factor" bar. Matches the academic literature's own prediction that insider buying specifically, not net buying/selling, carries signal. Sign consistency: 85.5%.
A 6-criterion quality/distress signal, tested against 12,242 real observations across 431 companies. Real, crisis-robust, and out-of-sample stable - but not a "core factor" by our own strict bar. Sign consistency: 80.0%.
Negative standard deviation of TTM net margin across 3 independently-sampled points in time - low volatility as a quality/safety signal, tested against 15,547 real observations across 498 companies. Positive-signed, matching the "quality minus junk" literature's predicted direction. The closest any candidate has come to a second "core factor" without quite clearing it. Sign consistency: 76.4%.
TTM earnings / price, tested against 82,188 real observations across 1,864 companies. The strongest out-of-sample stability of any candidate we've tested (100%), and a positive relationship matching the classic value-factor literature's predicted direction - short of "core factor" only on fold-to-fold sign consistency. Sign consistency: 75.9%.
TTM R&D spend / total assets, tested against 6,386 real observations across 615 companies. Positive-signed, matching the R&D-mispricing literature's predicted direction - fully SEC-derivable, no market-price data needed as an input. Sign consistency: 71.8%.
10-year Treasury yield minus year-over-year CPI inflation, tested against our own tracked universe's equal-weighted market-wide forward return. The largest effect size we've measured (0.413) and the first macro/regime candidate to clear a validation tier - positive-signed, the opposite direction its own motivating literature predicts. Fully public-domain (FRED-only), no Yahoo-sourced data as an input. Sign consistency: 70.1%.
ROC20/ROC60-based momentum, tested at 4 horizons from 1 month to 1 year. A real, if weak, mean-reverting (not continuation) relationship - most reliable during market stress. Sign consistency: 69.8%.
Year-over-year change in TTM R&D spend, tested against 2,288 real observations across 67 companies. Negative-signed - the opposite direction from both its own motivating literature and from R&D Intensity's own real positive result on the identical underlying concept, a genuinely informative divergence between "how much" and "how fast growing." Sign consistency: 67.3%.
Year-over-year change in total assets. A real, walk-forward-consistent relationship - positive-signed, the opposite direction the overinvestment literature predicts for this construction. Sign consistency: 66.8%.
A clean null result: no measurable relationship with forward returns, sign flips between the two tested horizons.
A second null in the "return on X" family, tested against the largest real evidence base of any candidate so far (42,926 observations, 1,447 companies) - ruling out data sparsity as the explanation.
A four-signal directional-consensus composite (insider consistency, dilution, capital allocation). Genuine null, though on a thinner evidence base than ROE/ROIC.
A third genuine null in the "return on X"/signal family - net Purchase-minus-Sale, on our largest evidence base (200,522 observations). Isolating purchases specifically (above) turned this exact null into our first "core factor."
(Revenue - COGS) / Assets, the classic Novy-Marx construction. Direction leans negative but too inconsistently fold-to-fold to call it a confident result either way.
Total debt / assets, testing the "low-leverage outperforms" safety-anomaly literature. A clean null - near-zero out-of-sample agreement, direction flips essentially at random fold-to-fold.
Our first single-time-series (macro, not cross-sectional) candidate, tested against our own market-wide forward-return proxy. Positive-signed as the recession-indicator literature predicts, but near-coin-flip sign consistency and zero out-of-sample stability - a genuine null.
Not enough real filers tag the specific XBRL concepts 3 of its 9 criteria depend on to run a meaningful test, not a finding that it doesn't work.
Collapsed to just 57 real observations - the same XBRL cash-flow-tagging gap that made classic Piotroski inconclusive, not a tested null.
Zero qualifying observations across the entire tracked universe - the most severe data-coverage collapse of any candidate tested. The validation script itself never produced a scorecard.
Only ~3 years of usable history in FRED's currently-published series vintage for this indicator - too thin to trust a walk-forward result either way, not a tested null.
We publish the tiering honestly, including the null and inconclusive results, not just the ones that passed. Full derivation of every number: FACTOR_VALIDATION_METHODOLOGY.md.
Full validation methodology: FACTOR_VALIDATION_METHODOLOGY.md.
Free early access to the API: Get API access →