Methodology, stated plainly, with every number sourced

Prior Moves is an independent research publication. It predicts the next position each tracked investor is likely to add, from that investor’s own disclosure history, before the filing that reveals it is public. This page is the one to read if you are deciding whether to trust the rest of the site, or to cite it. Every figure below names the file in the project repository that produced it, and the same files feed the working paper. Numbers on this page are read from those files at build time; nothing here is typed in by hand. Build: 2026-10-16T10:59:12Z.

Who we track, and why these investors

We track 58 investors in the United States, plus filers in Japan, the United Kingdom, Korea, the European Union and Hong Kong. The US roster is built around discretionary long-horizon managers whose 13F filings carry conviction: value and quality investors, activists, concentrated stock-pickers. Quant and multi-strategy books are carried for contrast, because filings that churn thousands of positions say little about what a manager believes. The five other markets use each register’s large-shareholding notices, filed within days of a threshold crossing rather than 45 days after a quarter end; the regime pages under /regimes/ explain each one.

How the prediction works

The product’s current Top Picks ranking and the historical US research panel below are separate evaluation contexts. The panel result is a historical pair-weighted ranking diagnostic measured walk-forward on held-out manager-security observations. It does not describe current board performance, predict investment returns, or establish alpha. The product ranks predictions separately from disclosed positions; see the track record for the distinct portfolio-return evidence.

Does it predict better than guessing

In this historical panel, AUC 0.6762, 95% investor- clustered interval 0.6563 to 0.7002. The statistic describes ranking between a held-out positive and negative observation. It is measured on 49 US managers over 35 quarters,124,731 held-out rows, and 1,244,011 comparable pairs across 1,482 investor-quarters, against a base rate of 12.46%. The top tenth of this historical ranking is selected 3.17 times as often as the base rate. None of these quantities measures returns or the current board’s performance. source: runs/gates/us13f_pair_weighted_auc.json

The estimator: pair-weighted AUC, stratum = investor-quarter, investor-clustered bootstrap. The older headline, a median across investors of a mean across quarters, reads 0.6589 on the same rows; it is kept as a secondary figure because that estimator lets investor identity carry part of the score.

Two measured biases bound this figure. Survivorship: the price history that decides which rows are scored was seeded from a 2026 ranking, so securities that survived to 2026 are over-represented. On scored names they carry +0.051 AUC per investor, paired (28 of 33), a lower bound, so read 0.6762 as an upper estimate. source: runs/gates/j1b_universe_selection.json Option rows: a 13F lists purchased calls and puts, and convertible bonds, as separate rows on the same security, and the current labels add them to shares. Scored on stock-only labels, the same frozen predictions give 0.6611 on 119,124 rows, all 49 managers retained, 31 with changed labels. A stock-only refit waits for the model freeze to end on 2 December 2026. source: runs/gates/q24_option_rows_in_labels.json

A rights-cleared SEC-only sample, the evaluator that scores it, shuffled-label and baseline controls and the point-in-time audit are on evaluate.

The other five registers

Only Japan clears the quotability bars (0.7325 against a shuffled control of 0.4968). The others are shown with the reason they are not quoted, because dropping a market in silence is the thing this page exists to prevent. source: runs/gates/claim_facts.json

Does it make money

A backtest is history, not a forecast. A 15-name book formed at each quarter end and held one quarter, net of costs, returned 489.5% cumulative against the index’s 239.4% over 35 quarters (2017-09-30 to 2026-03-31), which is 22.48% a year against 14.99%. Quarters in which the book finished above the benchmark: 23 of 35, one-sided sign test p = 0.0448. source: runs/gates/claim_facts.json

The average quarterly margin of 2.217 points is not significant on its own. Its t-statistic is 1.313 and its 95% interval runs from -0.674 to 5.86 points, which includes zero. Drop the single largest quarter (2026-03-31, +47.5 points) and the mean falls to 0.884 points at t = 0.828, the sign test to 22 of 34 (p = 0.0607), and the book compounds at 16.35% against 13.57%. Reaching t = 2 at this effect needs about 81.2 quarters, 20.3 years, against the 8.8 we have. That is a limit of the disclosure calendar, not of the model. source: runs/gates/A1_return_power.json

The book enters at quarter end, and the filing that reveals those holdings is not public until up to 45 days later. Entered at the date the filing actually becomes available, a pre-specified family of 8 portfolio transforms of the ranking produces a best absolute t of 1 against a family threshold of 2.57 on 84,021 predictions, and decile monotonicity of subsequent return across 82,949 priced rows is 0.018. The ranking predicts who buys, not what rises. No returns or outperformance claim appears on any public surface of this site. source: runs/gates/A1_rank_to_return.json

Has it predicted anything in real time

Forecasts are hashed and anchored in Bitcoin before their outcomes exist. The record holds 70 distinct (market, investor, issuer) forecasts from 2026-07-07 to 2026-10-14, collapsed from 97,424 daily receipt rows; 90 receipts carry a completed third-party attestation. 65 forecasts have matured and 4 hit. The per-call ledger, with each call’s stamp, window, outcome and the rule that graded it, is on the track record. The ledger also keeps each call’s grade under the superseded rule, so it holds more rows than this table counts; only the fixed-window rule feeds these numbers. source: runs/gates/R1_stamped_call_grading.json

The hypothesis register

Every question is entered with its verdict rule fixed before the run. The register holds 498 litigated hypotheses: 190 confirmed, 140 refuted, 168 underpowered. A refuted result must carry a counterargument naming the measurement that would overturn it. source: runs/hypotheses/INDEX.json

Numbers we retired, and when

A figure that was published and later found wrong is listed here rather than deleted, so a reader who saw it elsewhere can check what replaced it.

What this is not

It is not investment advice, and Prior Moves is not a registered investment adviser. It is a general-circulation research publication: one impersonal model portfolio per investor, identical for every reader, the way model-portfolio newsletters have worked for decades. We never execute trades, never hold your money, and never tailor anything to your personal situation. You decide and place everything yourself at your own broker. A predicted probability is a statement about a filing pattern, never a claim that the investor has bought a name, or will, or that you should.

Who runs this

Prior Moves is built by a solo founder, bootstrapped. The core research is free and stays free. The paid tier funds alerts and a watchlist, never the predictions themselves, and never touches anyone’s money.

Verify it yourself: the live boards are on top picks, the graded stamped record is on track record, the score components are on how it works, the calibration of the probabilities is on calibration, and the full write-up with reproduction commands is the working paper. Forward quarter: 2026-09-30. Research, not investment advice, see the full disclaimer.