Predicting the Next Position of a Disclosed Institutional Portfolio

Evidence from five disclosure regimes, and why a predictable trade is not a profitable one.

Howard Chan. PriorMoves, Hong Kong. Incoming undergraduate, University of Cambridge (from October 2026). ORCID 0009-0000-3463-3795. Version 2, 17 September 2026. Working paper.

Download the PDF (20 pages). Submission status: staged for SSRN and Preprints.org; the arXiv q-fin cross-post waits on an endorsement. This page is updated when a venue posts it.

Abstract

Institutional equity holdings are disclosed publicly and late. A United States 13F filing describes a portfolio as it stood up to forty-five days before the filing appears, and the large-shareholding regimes of Japan, Korea, Hong Kong, the United Kingdom and the European Union each impose their own statutory delay on their own trigger. The finance literature reads this record as an input to a return question: can an outsider profit by copying what was disclosed? We ask a narrower question that the same data answers far more cleanly. Given a manager’s disclosed history, which position does that manager add next?

We fit one gradient-boosted classifier per manager on a purged walk-forward split and evaluate with a pair-weighted stratified AUC whose strata are manager-quarters, so that no part of the score can be earned by learning which manager a row belongs to. On 49 United States managers across 35 quarters and 124,731 held-out observations, comprising 1,244,011 comparable pairs against a base rate of 12.46%, the model reaches an AUC of 0.6762 with a manager-clustered 95% bootstrap interval of 0.6563 to 0.7002. The top decile of the ranking is bought 3.17 times as often as the base rate.

The method transfers to four further disclosure regimes whose filing triggers, thresholds and deadlines differ from the United States and from each other. Against a control refitted on labels permuted within each manager, which preserves manager identity in full and destroys every other signal, Japan, Hong Kong, the United Kingdom and the European Union clear their controls. Korea does not, and we report it as a negative result rather than suppressing it or quietly dropping the market.

The paper’s second half is the part a practitioner should read first, and it is negative. The predictable component of institutional demand does not convert into excess return at any horizon we can measure. A pre-specified family of 8 portfolio transforms of the ranking, entered at the date the filing actually becomes public and held one quarter, produces a best absolute t-statistic of 1 against a family-wise threshold of 2.57 on 84,021 predictions. Decile monotonicity of subsequent return across 82,949 priced rows is 0.018. A quarterly 15-name book formed at quarter end returns 489.5% against a benchmark 239.4% over the same 35 quarters, and we report in full that only one of five estimators of its mean quarterly margin clears 0.05, that none does once the single largest quarter is removed, and that reaching t = 2 at the observed effect and dispersion would require about 81 quarters of a quarterly instrument.

We argue that this combination is coherent rather than contradictory, and that it is the honest reading of the disclosure record: the identity of the next buyer is predictable because it is driven by persistent, slow-moving manager characteristics, while the price impact of that buyer’s arrival is either absent, already impounded, or smaller than the transaction costs of harvesting it. The contribution is a clean, reproducible measurement of the first fact and an equally clean refutation of the second.

Keywords: institutional holdings, 13F, disclosure, portfolio prediction, machine learning, market efficiency, backtest overfitting. JEL: G11, G14, G23, C53.

The forward record the paper points at

85 forecasts are stamped and public before their outcome existed, 40 have matured and 4 hit. The per-call ledger is on the track record and the grading rule is fixed in the repository before each block of calls is stamped. The record becomes informative in 2027; today the correct statement is that it exists and is checkable.

Data appendix

Every figure in the paper is emitted by a named artifact and the document generator reads the artifacts. This page renders from the same files, so a number here and a number in the PDF cannot drift apart without the build failing its audit (scripts/gates/paper_number_audit.py).

The United States panel is built from SEC EDGAR 13F-HR filings; the other five registers are EDINET (Japan), the FCA National Storage Mechanism (United Kingdom), DART (Korea), the AFM register (European Union, Netherlands) and HKEX Disclosure of Interests (Hong Kong). Each regime is described on /regimes/. The held-out panels are not redistributed with the paper because the underlying filings are the registers’ own; the collectors and the artifacts that summarise them are in the repository named in the PDF.

AI disclosure

Parts of the manuscript were drafted with the assistance of a large language model. All numerical results are produced by the named artifacts in the project repository and were not generated by a language model. The author is responsible for the content, the method and the errors.

The plain-language version of these numbers, with the figures we have retired and when, is on methodology. Research, not investment advice; see the disclaimer.