# Prior Moves dataset sheet: SEC-only next-entry rankings

Version 2026-10-16 (identifier, period and latency rows corrected; deadline-roll limit added). Contact: howard@priormoves.com.

## What it is

For each tracked investment manager and quarter, a ranked list of securities the manager does not hold, scored by how likely each is to appear as a new position in the manager's next Form 13F. Built only from SEC EDGAR 13F filings.

## Coverage

| Item | Value |
|---|---|
| Managers | 58 tracked 13F filers, from concentrated long-only firms to large high-turnover filers (sample: 55) |
| Candidate universe | Top 250 issuers by number of tracked managers holding them in the same quarter |
| History | Walk-forward scores for every quarter from 2014-12-31 (46 quarters, 377,215 rows); the 25-quarter result uses 2020-03-31 onward; panel 390,421 manager-issuer rows |
| Frequency | Quarterly |
| Latency | Rankings available at the as-of quarter's 13F deadline (quarter end + 45 days, rolled to the next business day); outcomes at the next quarter's deadline, about three months later |
| Public sample | Last 4 quarters, 28,092 rows, identifiers one-way hashed |

## Fields

| Field | Type | Meaning |
|---|---|---|
| manager_id | string | Manager key: first 16 hex of SHA-256 of the manager slug. The licence key table maps it to name and SEC CIK |
| period | date | As-of quarter end. The forecast is for the next quarter |
| security_id | string | Issuer key: first 16 hex of SHA-256 of the issuer name as filed, upper-cased. Never a CUSIP |
| score | float | Model score; higher means more likely to appear |
| rank | int | Rank within manager-period (live cohorts only; in the sample, order score within manager and period) |
| prediction_ts | timestamp | When the ranking was frozen |
| baseline_score | int | Peer-count baseline: number of other tracked managers holding the issuer |
| label | int | 1 if the security appeared as a new position in the next first-public 13F |
| outcome_public_ts | timestamp | When that filing became public |

## Point-in-time rules

- Holdings: each manager's first 13F filed within 45 days of quarter end. Amendments and late filings are excluded.
- Features: computed from filings public at the forecast date. Peer features exclude the target manager.
- Universe: chosen from same-quarter filings only. No price data, no later information.
- Training: expanding window; a training row's target quarter ends before the test quarter.
- In the public sample, prediction timestamps are simulated (quarter end + 45 days). Live deliveries carry real timestamps with an OpenTimestamps proof.

## Evidence

| Test | Result | File |
|---|---|---|
| Walk-forward, 25 quarters 2020 to 2026 | Mean within-manager AUC 0.6635 | runs/gates/q25 (summary on /evaluate/) |
| Peer-count baseline | 0.6368 | same |
| Shuffled-label refits (5) | Mean 0.5365, best 0.5583 | same |
| Public sample, 4 quarters | Pair-weighted AUC 0.6865 (CI 0.6525 to 0.7324) against baseline 0.6458; difference +0.0407, paired CI 0.0157 to 0.0645 | results.json |
| Top-decile lift (sample) | 6.04x over a 1.57% base rate | results.json |
| Full history, 46 quarters 2014 to 2026 | Pair-weighted AUC 0.6766 (manager-clustered CI 0.6584 to 0.6925) against baseline 0.6281; difference +0.0485, paired CI 0.0252 to 0.0691; top-decile lift 5.54x over a 1.61% base rate | history/results.json |
| Sealed prospective cohort | Q3 2026: 10,448 rankings, 51 managers, hashed and stamped 2026-10-09; grades after 2026-11-16 | /evaluate/ |

## Known limits

- The public sample's quarters were evaluated before. It reproduces a result rather than providing a fresh holdout.
- Most of the ranking comes from peer popularity. The model's margin over that baseline is about 0.03 to 0.04 AUC.
- 13F shows net long positions at quarter end only: no trade dates, no shorts, few derivatives.
- The universe is the 250 most-held names. Smaller names are not ranked.
- The 45-day cutoff does not roll for weekends and holidays, so in 18 of 48 historical quarters only the 15 to 22 managers that filed early are present. The Q3 2026 grade uses the legal deadline, Monday 16 November.

## Rights

Built only from SEC EDGAR filings. The source-rights gate (`docs/rights/source_entitlements.json`) clears this lane for an external evaluation pilot. Redistribution through a hosted API is unresolved. No vendor, academic or scraped source enters this dataset.

## Delivery

| Stage | How |
|---|---|
| Evaluate | https://priormoves.com/evaluate/: CSV, Parquet, evaluator source, SHA-256 manifest |
| Full history | https://priormoves.com/data/sec-eval/history/MANIFEST.json (files predictions.csv.gz, outcomes.csv.gz, pm_evaluate.py, results.json beside it): all 46 quarters, 377,215 rows, gzipped CSV in the same schema, evaluator and SHA-256 manifest |
| Forward test | Buyer names managers and baselines; rankings stamped before the deadline; delivered by expiring signed link |
| Trial or licence | Versioned files in object storage (S3-compatible), or a data share on request |