Prior Moves · Mirror the greats · US · Japan · UK · Korea · EU

Mirror the world's best investors, before their next move is public.

A per-investor ML model builds a copy sheet for each great investor across five markets, their disclosed holdings plus the stocks they're most likely to buy next. The same next-action model is now validated in three markets: the US (SEC 13F), Japan (EDINET, walk-forward AUC 0.84) and the UK (FCA, 0.80); Korea and the EU are tracked on disclosure today. The disclosure-anchored coverage is the moat: US-only trackers never see Buffett's Tokyo trading houses, the Murakami funds, Oasis or Cevian. You place the trades at your own broker.

Prior Moves · priormoves.com Founder: Howard Chan · Tokyo Stage: live, bootstrapped, pre-incorporation Surface of a broader quantitative-research system

What it does

Copy-the-greats apps mirror an investor's most recent 13F, public data, up to 45 days stale, and auto-execute, which makes them registered advisers that hold your money. Prior Moves anticipates the next buy before it files and hands you a copy sheet you implement yourself: a target portfolio of each investor's holdings plus their predicted next moves, with calibrated conviction. You keep custody and control; it stays a research publication, not a managed account.

Anticipation, not lag

A per-investor model predicts the next buy before it is public, so the mirror leads the disclosure rather than copying a stale filing. ~100 investors across five markets (Buffett, Burry, Ackman, plus Murakami, Oasis, Cevian, Silchester …); 218 signals per stock per quarter in the US, a filing-history action model abroad.

You own the execution

The output is a copy sheet, weights and a dollar split you take to your own broker. No execution, no custody, no advisory fee on your assets. Impersonal and general-circulation, which keeps it publisher-exempt rather than a registered adviser.

Honest, calibrated odds

Isotonic calibration on 72,268 out-of-sample predictions cuts calibration error from 0.020 → 0.001; a label-shuffle canary collapses accuracy to a coin flip, confirming no leakage. The full backtest and confidence interval are published openly.

The signal stack

218 features per stock per quarter, fused across sources, retrained walk-forward every quarter.

1
IngestSEC 13F filings, insider + congressional trades, activist disclosures, fundamentals, sentiment, and macro data, normalised into a per-stock-per-quarter feature matrix.
2
Train (per investor, walk-forward)One model per investor; each quarter trained only on earlier quarters. 26-investor validated roster × 47 completed quarters of out-of-sample evaluation.
3
Rank + calibrateScore the full equity universe for each investor; isotonic-calibrate conviction so the displayed % is honest.
4
Publish the mirrorEach investor's copy sheet, holdings plus predicted next buys, weighted, with a dollar split, free to export. Plus the consensus mirror, the full backtest, and a live paper-trading account. Readers execute at their own broker.

Track record

+1.8ppper quarter vs S&P 500 (top-15, equal-weight, net of costs)
26 / 47quarters beating the index (2014–2026)
72,268out-of-sample predictions calibrated
t ≈ 1.45~94% probability the edge is positive

Backtest, not a live trading record. The net edge is small and not statistically distinguishable from zero, the 95% bootstrap confidence interval (≈ −0.4 to +4.3 pts) includes zero (t ≈ 1.45) on 47 completed quarters; a directional research signal, not proven. The unfinished current quarter is excluded and completed-quarter marks are frozen so the figure does not drift. A separate leak-free re-score over the same broad ~2,000-name universe the live product uses (instead of each quarter's filed holdings) gives a comparable +1.7 pts/q (t ≈ 0.7, also includes zero), so the figure is not an artifact of the backtest's candidate set. Not financial advice; educational and informational only. Predictions are statistical estimates of public SEC 13F filings, derived from public data.

How Prior Moves uses AI & cloud

Training25 per-investor models × walk-forward across 35+ quarters; retrained every quarter as new filings land, heavy, bursty batch compute.
Feature pipelineMulti-source ingestion (filings, insider/congressional/activist, fundamentals, sentiment, macro) into a 218-feature matrix per stock per quarter.
Inference + vector searchScore the full universe per investor each quarter; similarity search over the signal/holdings corpus (Chroma vector DB).
Serving + observabilityFree public app + weekly digest; metrics + traces on the model and data pipeline.

Cloud + AI credit converts directly into research depth: more investors modelled, more signals fused, faster quarterly retrain cycles at the same unit cost.

Prior Moves

Prior Moves is the public surface of a broader quantitative-research system built by Howard Chan (Tokyo; incoming Peterhouse, Cambridge, HSPS). Free SEC 13F intelligence today; the research engine underneath is the longer game.

Machine learning5 markets (US·JP·UK·KR·EU)SEC 13FWalk-forwardIsotonic calibrationChroma vector DBQuant research