Tilting 1,171 Institutional Equity Portfolios Toward Current Research

We test whether a score derived only from the scientific literature improves equity portfolios that other people chose. The score uses no prices, fundamentals or analyst data. No input, embedding model, prediction or company description used in the evaluation draws on papers or filings dated after it. Each quarter from March 2020 we examine the arXiv papers available at that date, select groups of papers with a proprietary method, and score US-listed companies by how closely the business section of their latest 10-K matches the existing and expected discoveries in those groups. We then re-weight, by rank of that score, the reported long US equity holdings of every 13F filer that meets fixed data rules: 1,171 managers, including asset managers, pension funds, banks, insurers and hedge funds. Each tilt keeps the manager’s own holdings and prices; only the weights change, and only at report dates. The tilt’s one parameter, a similarity threshold, is chosen walk-forward from past results. Over 26 quarters, from March 2020 to October 2026, the tilt adds +1.31 percentage points a year to the average manager’s own book. Random orderings of the same holdings, put through the same procedure, add +0.33 (z = 8.0; no draw of 1,000 reaches the observed value). The tilt improves 85% of the books and beats random orderings at p < 0.05 for 458 managers. The active return leans toward growth; after market, value and size factors the pooled alpha is +0.59 pp a year (t = 2.3). On the S&P 500 and the Nasdaq-100, against the real index funds, the tilt adds +2.01 and +2.69 pp a year, and +1.16 and +1.95 after style. A forward test on holdings and prices that do not yet exist has been registered (doi:10.5281/zenodo.23692720).