Numerai Classic model performance correlation
Pairwise historical score correlations among the top 100 or 200 staked Classic models, with median, quartiles and measurable-pair coverage.
Coverage & dates
- Record date range
- to
- Date meaning
- Opening dates of the rounds selecting the model cohorts, including unavailable estimates. The range can include gaps; it does not imply complete observations throughout.
- Collected
- · 01:31:04 UTC
- Prepared by
- nmrdash, from the public sources below.
Sources & methods
Pairwise Pearson correlation of prior finalized 20-day CORR score histories within each round's top 100 or 200 positive-stake models. Each window uses 26 Saturday rounds with one metric version and universe; each pair needs 20 shared varying scores. Median and quartiles describe available distinct pairs; stake weights are products of reference-round stakes. Self-pairs and missing histories are excluded. Pair-weight coverage is reported. This measures performance co-movement, not correlation between stock predictions. Current historical scores can contain revisions or backfills.
Each window uses 26 finalized weekly rounds with one score version; each pair needs 20 shared varying scores. Incomplete stake rosters and insufficient histories remain gaps.
This describes co-movement of model score histories, not correlation between stock predictions. Published historical scores can contain revisions or backfills.
Source fields: Numerai roundDetails: historical selectedStakeValue and v2Corr20; round score configurations.
JSON fields
Paths use [] for array items and * for metric keys such as corr20 and corr60. Absent fields or null values mean unavailable information; a numeric zero remains a real observation.
| Field | Meaning | Unit |
|---|---|---|
fetched_at | Collection timestamp in UTC. This does not establish the observation date or score finality. | Timestamp |
definition_version | Version of the dataset’s measurement and export definition. | Text |
points[].round | Reference round selecting the staked model cohort. | See definition |
points[].time | Reference round opening timestamp in UTC. | See definition |
points[].cohort_size | Requested number of top-staked models: 100 or 200. | models |
points[].selected_models | Number of eligible models selected. | models |
points[].score_version | Shared CORR20 score definition for the historical window. | See definition |
points[].history_rounds[] | Prior finalized weekly rounds used in the window. | See definition |
points[].median | Median Pearson correlation across measurable distinct model pairs. | correlation (−1 to 1) |
points[].p25 | 25th percentile of measurable pair correlations. | correlation (−1 to 1) |
points[].p75 | 75th percentile of measurable pair correlations. | correlation (−1 to 1) |
points[].stake_weighted_mean | Mean pair correlation weighted by products of reference-round stakes. | correlation (−1 to 1) |
points[].measured_pairs | Number of distinct measurable pairs; self-pairs are excluded. | pairs |
points[].possible_pairs | Total distinct pairs in the requested cohort. | pairs |
points[].pair_weight_coverage | Fraction of potential pair stake weight represented by measurable pairs. | fraction (0–1) |
points[].stake_share | Fraction of reference-round stake held by the cohort. | fraction (0–1) |
points[].stake_roster_complete | Whether the reference-round stake roster is complete. | See definition |
points[].missing_reason | Explanation for an unavailable estimate. | See definition |
Download
Download classic-model-correlation.json · application/json
Definition version: numerai-classic-intra-model-performance-correlation-v1. Decimal strings preserve exact amounts; use decimal arithmetic when working with NMR or USD totals.
See the release manifest for this file’s checksum and byte count. The download is public and needs no authentication or JavaScript.
Interactive charts can load newer snapshots. Compare collection and observation dates before combining a download with chart values or another release.