Two Pre-Registered Nulls in Alternative Data: Insider Purchases and Legal-Risk Language in SEC Filings, 2014–2026

Abstract

We report two pre-registered tests of popular alternative-data signals built from U.S. Securities and Exchange Commission (SEC) filings, each rebuilt end to end from public sources. Insider purchases: across 31,971 tradeable officer and director open-market purchases (Form 4, 2014 to mid-2026), the 63-day return in excess of matched comparables averages +0.85%, but the median is −0.36%, the hit rate is 49%, and with overlap-robust inference the t-statistic is 1.89 rather than the 2.30 reported in our exploratory run. A calendar-time portfolio is essentially levered small-cap beta (market beta 1.02, small-minus-big (SMB) size loading 0.76). Its Fama–French six-factor alpha is 0.29% a year net of 30 basis points (t = 0.18) and turns negative under buy-and-hold weighting or realistic small-cap costs. The retail refinements “insider buying below the 200-day average” and “before earnings” lower alpha, and so does the Cohen–Malloy–Pomorski “opportunistic” filter. A same-stock placebo detects a modest timing effect (+2.5% a year relative to the same stocks 3–12 months later, t = 2.1). That effect only lifts a universe with negative factor-adjusted returns to zero. Legal-risk language: quarter-on-quarter changes in litigious language in 10-K and 10-Q filings have no detectable relationship with subsequent returns. This holds for 79 large caps (3,383 filings) and for a pre-specified random sample of 366 small and mid caps (15,383 filings), for the original lexicon proxy, the Loughran–McDonald Litigious list and a “Lazy Prices” similarity measure: Fama–MacBeth rank information coefficients (ICs) are within ±0.02 with |t| < 1.1 for the pre-registered signal. A pre-set gate therefore stopped us from building a costly language-model scorer. A direction-free check suggests the gate was right: once volatility is controlled for, the proxy does not even predict the size of the next move. The rebuild also exposed a pagination truncation and a look-ahead normalisation in our own exploratory code. Power analysis shows the tests could detect rank ICs of about 0.05 (large caps) and 0.025 (small/mid caps) and annual alphas of about 4.5%. Declaration of generative AI use: All research questions, hypotheses, study designs and conclusions are the author’s own, including the decision to re-audit the author’s earlier exploratory results, which identified data errors in them. Under the author’s direction, generative AI tools (Claude Code, Anthropic) wrote most of the analysis code and ran it, and assisted with drafting, editing and formatting the text, and with typesetting. The author reviewed all code, results and text and takes full responsibility for the content of this paper.