How it works

A mirror is one model portfolio

For each of 58 famous investors we publish a single impersonal model portfolio, their latest disclosed 13F holdings plus the stocks a model predicts they will buy next. It is identical for every reader. We never execute trades, hold assets, or tailor anything to your account.

Predicting the next buy

13F filings disclose what each institution held at the end of a quarter, about six weeks after the fact. The model learns each investor’s historical buying pattern and ranks a broad ~2,000-name candidate universe, known at the rebalance date, by the probability that the name becomes a new buy in the next filing. The ranking is the PriorScore. The published page never changes that ranking after the fact.

Tested walk-forward, and one leak we did not solve

Every backtest is walk-forward: at each historical rebalance the model only sees data available on that date, holds the top predicted names equal-weight for one quarter, and scores complete quarters only, gross of costs. As-of joins are availability-lagged so no future feature information leaks in.

One leak survives that, and we will not call these results leak-free because of it. Each quarter’s positions are formed from 13F holdings as of quarter end, and those filings are not public for up to 45 days. Re-run so nothing is bought until the filing is actually public, the same pipeline returns −3.4 pts/q (t = −1.74) over 34 complete quarters. The measured edge does not survive that constraint, and the track record page states this next to the number rather than in a footnote.

What the accuracy number means

There are two accuracy numbers on this site and they answer different questions. The one that describes the product is a measured hit rate on the live board: the top-ranked name is the investor’s actual next new buy 1.7% of the time, against 0.6% for a name drawn at random from the same board, which is 2.8x chance over 2,030 investor-quarters of walk-forward holdout. The second is AUC, and it grades a much narrower exam that we no longer treat as a skill claim. It ranks names already inside a filing and asks which was opened most recently, and a one-line rule read off the previous filing scores a perfect 1.000 on it, because the label is defined as “absent last quarter”. Our models score 0.47 to 0.91 on it, median 0.659, which is worse than that free rule. We keep the spread here because it is an honest description of how well each investor’s style is captured, and for no stronger claim than that:

The spread is not random. It tracks how patient the investor is: holdout AUC falls with portfolio turnover (Spearman −0.60, p < 0.001, 49 investors), so for the lowest-turnover third of the roster the median is 0.705 while the highest-turnover third sits at 0.634. Behaviour repeats where books are stable, which is the same gradient Cohen, Lu and Nguyen find in mutual fund managers. Practically: trust the mirrors of patient managers most, and read fast-trading books with wider error bars.

• Median 0.659 across the scored roster, on the stacked engine (behavioural model combined with a style baseline) served since 22 August 2026. This is the narrow in-filing exam described above, not the next-buy task; on the live board the engine scores 0.55. Read it next to the baseline in the bullet below, because the level on its own overstates what the model adds. It was 0.61 here until 13 August 2026, when we found that buy-backs of names an investor used to own were being filed as the wrong kind of event; correcting that added 51% more real buys to the test and the number fell to 0.5868 before stacking recovered it. The old figures are not reproducible and we do not stand behind them.
• The average hides two different tasks, and the split matters more than the median. On names the investor has never held, the model scores 0.65. On names they used to hold, sold, and are buying back, it scores 0.54, which is close to a coin flip, and those buy-backs are 42% of the new-money buys inside the scored window (29% across the full filing history, so the harder half is over-represented in exactly the window the accuracy is measured on). The features describe the company and the investor's current book. Nothing tells the model what happened the last time this investor owned this exact name. So read the product as calling first-time buys, not comebacks.
• A static style screen beats it on the discretionary investors. On 14 August 2026 we scored the model against a 24-feature baseline that knows only what kind of company this is (sector, size, momentum, volatility, price, and their deciles) plus how much of that style the investor already owned in earlier quarters. It knows nothing about news, filings, analysts, insiders or peers. On first-time buys across 44 discretionary investors the baseline scores median 0.63 and the full 294-feature model scores 0.61, and the model wins for only 12 of the 44 (sign test p=0.004). Buffett 0.55 against the baseline 0.63, Ackman 0.47 against 0.67. The 294 features do earn their place on systematic books, where the model leads by 0.06 (D.E. Shaw 0.76 against 0.66, Citadel 0.67 against 0.61). So for a concentrated human investor, most of what the ranking captures is standing taste rather than a read on the next decision, and the honest description of that surface is style-consistent anticipation. We publish the baseline next to the number so the increment is visible instead of only the level. Gate and per-investor table: runs/gates/q2_static_taste.json.
• Every AUC on this page is measured against the investor's own book, and Top Picks does not rank that universe. Scoring is done on candidates drawn from the names an investor has held, so a positive sits among a few dozen companies that manager already knows. The Top Picks board ranks a far wider set: every name the investor has ever touched plus the 2,000 most-held companies across the whole roster, which is roughly 1,000 candidates a quarter instead of 36. Scored on that served universe on 27 August 2026, the engine reaches AUC 0.5086, with a bootstrap CI of 0.4907 to 0.5222 and a Wilcoxon p of 0.57 over 58 investors, so it is indistinguishable from chance. The same models on the same investors score 0.5832 against the book. The gap is 0.075 and it is the honest size of the drop from the published number to the task the product actually performs. A pooled architecture, one model per quarter over every investor's rows, fixes the mechanical half of the problem (scorable rows rise from 23.2% to 92.6%) and does not move the ranking. So read the accuracy figures above as a measure of how well each mirror describes its own investor, and do not read them as the hit rate of the Top Picks list. Gates: runs/gates/E9_served_universe_auc.json and E9_served_universe_pooled.json.
• Every AUC on this page is also measured on a universe that already knows which companies survived. The price and feature caches were filled from a ranking taken in June 2026, so a name an investor bought in 2016 and which was delisted in 2019 never entered the cache and is never scored. Checked on 14 August 2026: of the names still held in the last four quarters, 59.9% are in the price cache, against 6.7% of names last held five or more years ago. Within the rows that do get scored, the model ranks survivors better than eventual disappearers by 0.056, and 28 of 33 investors show the gap. Holding the quarter fixed, so this is not just the model being weaker in older years, the gap is 0.039 in 14 of 15 quarters. So read the median as roughly 0.04 too high, and read that as a floor rather than a ceiling, because names excluded from the universe entirely cannot be measured at all. Gate: runs/gates/j1b_universe_selection.json.
• 0.69 to 0.83 at the top. D.E. Shaw 0.83 is a broad systematic book, where ranking a new entry against a book of hundreds of names is an easier task and the lift on a live universe is near zero. That fund and the other multi-strategy books are excluded from the consensus Top Picks, so our best AUCs are not the ones the product ranks on. The rest of the top is ordinary discretionary managers: Oakmark 0.73, Buffett 0.71, Akre 0.71, Marks 0.69.
• Nothing is below chance any more, and three models were until August 2026: Pabrai 0.36, ValueAct 0.48 and Ackman 0.49 are now 0.55, 0.60 and 0.66. Combining the behavioural model with the style model helped the weakest books most, which is what you would expect if their handful of high conviction bets are driven by private research the behavioural signals cannot see. The bottom of the roster is now Sachem Head 0.53 and Hound 0.54. A further nine of the 58 models have too few holdout events to score at all. We say so on their pages rather than hide it.

So the path to the number is plain: a separate model per investor, trained walk-forward on up to 294 public features per stock per quarter, calibrated on 124,731 out-of-sample calls across 49 investors. A label-shuffle canary collapses it to a coin flip, which rules out labels bleeding into features. Survivorship is a separate problem, and on 14 August 2026 we stopped calling it unmeasured and measured it. Names an investor still held in the last four quarters are in our price history 59.9% of the time; names last held five or more years ago, 6.7%. Comparing like with like inside each investor, the model scores about 0.05 higher on the survivors, in 28 of 33 investors, and about 0.04 higher within the same quarter, in 14 of 15 quarters. So at least 0.04 of the published median is selection rather than skill, and that is a lower bound, because a name dropped from the universe entirely never gets scored at all. The honest claim is a median 0.659 on a survivorship-favoured universe, not a single big percentage. The survivorship measurement was made on the previous engine and we have not re-run it on the combined one, so treat the 0.04 as carried forward rather than freshly measured.

The honest result

A research publication

Prior Moves is a research publication: one impersonal model portfolio per investor, identical for every reader. You place any trades yourself at your own broker. No execution, no custody, no individualised advice. You hold the trades; Prior Moves publishes the playbook. Nothing here is investment advice. Read the full disclaimer.