Numerai Metrics Cheatsheet: CORR, MMC, BMC, CORJ60 & More

Beginner forum threads on Numerai converge on the same question: "What is the difference between CORR, MMC, BMC, CORJ60, FNCv3, Alpha, MPC, and TC?" There are good answers scattered across years of Discord posts and docs revisions, but Numerai retires and renames metrics every 12–18 months, so half the answers are stale by the time you find them.

This page is the cheatsheet. It walks through every scoring metric Numerai has shipped, what each one is calculated from, which versions are current, and, using real data from this dashboard's database, how correlated they actually are.

The Four You See on Modern Dashboards

If you opened the Trends page today, you would see four metrics: MMC, BMC, CORJ60, and FNCv3. These are the primary scoring metrics for Numerai Classic (tournament 8) as of round 1260. Everything else is either a precursor, a sibling for another tournament, or retired.

The catch is that Numerai keeps issuing new versions of each metric without changing the display name. "MMC" today is MMC v6, but submission scores from older rounds were computed under MMC v1 through v5, each with slightly different math. The Scoring Versions post quantifies how much each version change shuffles the leaderboard.

Every Metric Numerai Has Shipped, by Round

The Numerai API records every (metric, version) pair Numerai has produced. Plotted as a Gantt chart, the lifetime of each version makes the rename-and-retire pattern obvious.

The timeline tells a few stories at once. MMC is the most-revised metric on the platform, with six numbered versions across rounds 168–1255. FNC went through two retirements (v1, v2) before settling on the FNCv3 family in 2022, which itself iterated to v4 around round 833. TC (True Contribution) was Numerai's most ambitious metric (it tried to score how much your model improved hedge-fund-side trading) and quietly disappeared after round 713. The cluster of new versions around rounds 833–846 (MMC v4, BMC v2, FNCv3 v4, CORJ60, CORR60) is the largest coordinated re-versioning on the timeline: a scoring overhaul roughly two years after daily rounds launched at round 340.

The CORR Family: Different Time Windows, Same Idea

All correlation metrics measure the rank-correlation between a model's predictions and the eventual target. They differ in three things: the time window of the target, clipping behavior, and whether they are raw or derived.

• CORR (now also called Alpha in Signals) — the daily-updated raw Pearson rank correlation between predictions and the canonical target. Lives in the -0.10 to +0.10 range for staked models. Most volatile day-to-day because each day shows the score on partial data.
• CORJ60 — CORR with the 60-day-return target, then winsorized to ±0.20. The "J" stands for Jerome (Numerai used internal mathematician nicknames for several scoring versions). CORJ60 is the primary correlation metric for Classic in the current payout formula and is the most stable visible scoring number.
• CORR60 — raw 60-day correlation without the winsorization. Reported since round 846. Largely tracks CORJ60.
• CORR20 / v2_CORR20 — 20-day target CORR. v2_CORR20 v4/v5/v6 are the modern revisions.
• CORT20 — CORR target 20, an alternate 20-day formulation. Tracks v2_CORR20 closely.

The common confusion: "Is CORR the same as CORJ60?" No: they use different target horizons, so the same model can post a high CORR and a mediocre CORJ60.

What Does MMC Mean? (MPC, "Meta-Model Contribution")

MMC stands for Meta-Model Contribution: it is the metric Numerai pays its biggest multiplier on. It measures the residual correlation between your prediction and the target after subtracting the meta-model's prediction. A model that perfectly matches the meta-model has MMC ≈ 0, regardless of how accurate it is. A model that consistently identifies signal the meta-model misses has positive MMC. See MMC vs Correlation for what this implies for staking.

Formula sketch (current MMC v6 / MPC v2):

The projection step orthogonalizes your prediction relative to the meta-model before correlation. In Numerai Signals (tournament 11), the identical formula is shipped under the name MPC ("Meta-Prediction Contribution"). MPC v2 values are byte-for-byte equal to MMC v6 values for any model that submits to both tournaments: they are the same metric with two names.

MMC's coefficient in the payout formula is 2.0 × MMC, four times the 0.5 × CORJ60 correlation coefficient. Realized payout dollars depend on both the coefficient and the distribution width: see The Payout Factor.

BMC: Beyond the Benchmark Models

BMC (Benchmark Models Contribution) generalizes MMC. Instead of subtracting only the meta-model, BMC subtracts a basket of Numerai's official benchmark models, the public baseline predictions you can download. A high BMC means your model adds signal the benchmark cohort misses, which is operationally what Numerai wants: predictions that aren't just a tuned XGBoost on the released features.

BMC was introduced around round 200 and split into v1 (rounds 200–842) and v2 (round 834 onward). The two versions agree closely on rank order: see the scoring versions analysis for cross-version stability.

FNCv3 and FNCv4: Feature-Neutral CORR

Feature-Neutralized Correlation orthogonalizes a model's predictions against the features themselves (Numerai's published feature columns) before computing correlation against the target. A model that is just a clever weighted average of public features scores near-zero FNC. A model that captures signal beyond the published feature space scores positively.

There have been four FNC families:

• FNC v1, v2 (rounds 168–458) — early versions, retired.
• FNCv3 v1, v3, v4 (rounds 200–present) — the current line. FNCv3 v4 is what the dashboard shows today.
• FNCv4 (rounds 833–present) — successor for a newer target schema. Listed alongside FNCv3 but not yet primary.

FNC is the only metric in the visible set with a meaningfully positive median in our distribution check: see the next section. Staked models tend to post FNC ≈ +0.005 by default, even when their MMC is near zero, because feature space orthogonality is comparatively easier to get right than meta-model orthogonality.

TC, IC, RIC, Alpha — the Cross-Tournament Family

Several metrics either live in one tournament only, or were retired:

• TC (True Contribution, v1, rounds 178–713) — Numerai's most ambitious metric. It attempted to score predictions by their effect on the actual hedge-fund-side optimized portfolio. It was retired after round 713, likely because the simulator-based scoring was too expensive and too sensitive to risk-model changes. Still appears in older docs and old leaderboards.
• Alpha — the Numerai Signals (tournament 11) name for CORR. Alpha v2 (rounds 911–present) is numerically equivalent to Signals' canonical CORR. Classic does not produce Alpha values.
• MPC — Signals' name for MMC. MPC v2 = MMC v6 values.
• RIC (Residual IC, v2 / v3) — Signals-specific. Measures the residual information coefficient after subtracting the meta-model's signal. Conceptually similar to MMC but with a different baseline.
• IC, IC v2 — early "Information Coefficient" metrics, retired.

For Classic stakers, none of these are scored. For Signals stakers, Alpha+MPC+RIC are the working set.

The Correlation Matrix: What's Redundant, What Isn't

The question that drives most "which metric should I optimize?" threads: how independent are these metrics, empirically? The cheatsheet answer is the matrix below, computed across 571,108 staked model-round observations in Numerai Classic, rounds 1100 through 1240.

A few patterns are worth pointing at:

• v2_CORR20 and CORT20 are nearly the same metric (ρ = 0.83). Same with FNCv3 and v2_CORR20 (0.81). The 20-day CORR family has a high internal redundancy: different formulas, mostly the same signal.
• MMC and BMC have ρ = 0.79 — the two contribution-style metrics agree more than they disagree. Models that beat the meta-model also tend to beat the benchmarks.
• CORJ60 and CORR60 (ρ = 0.77) are essentially the same 60-day measurement with different clipping rules.
• CORR (the daily, full-target version) is the least like everything else — its highest correlation with anything is 0.36 (BMC), and it sits at 0.25 with CORJ60. This is not because CORR is broken; it is because CORR is a fast-changing daily measurement while everything else converges over the 20–60 day target window. The same model's CORR moves substantially day-to-day before stabilizing.
• MMC is meaningfully independent from CORJ60 (ρ = 0.40). Of all the cross-family pairs, this is the one with the most strategy implication: optimizing one does not automatically lift the other. See MMC vs Correlation for the staking story behind this gap.

Metric Scale: A 0.05 in MMC Is Not a 0.05 in CORJ60

A common beginner mistake is to compare metrics by their raw numerical value. They live on different scales.

CORJ60 has the widest spread: staked models routinely post values between -0.032 and +0.037. MMC is tighter, with the bulk of the field between -0.020 and +0.017. FNCv3 is the only metric with a noticeably positive median (+0.006), reflecting that feature-neutrality is somewhat free to harvest from any sensible model. A +0.02 MMC is beyond the p95 of the field, while a +0.02 CORJ60 sits well inside the routine range. When comparing leaderboards across metrics, lean on percentiles rather than raw values.

Which Metric Most Predicts Payout?

For someone deciding what to optimize, the practical question is: which metric, by itself, best predicts NMR payout? Bucketing models into deciles of each metric and averaging per-round settled payout makes this concrete.

BMC has the steepest slope (-0.29 to +1.18 NMR, a 1.47 NMR range across deciles), with MMC close behind at 1.37 NMR. CORJ60 has the highest floor (bottom decile is only -0.13 NMR) but a noticeably flatter middle: its payout signal is concentrated in the top three deciles. FNCv3 mostly tracks the BMC/MMC shape with a slightly flatter top.

The pattern matches the payout formula's 2.0 × MMC + 0.5 × CORJ60 weighting: the contribution metrics (MMC, BMC) reward consistently across the rank, while CORJ60 mostly pays for the top.

Cheatsheet: Aliases and Translation

If you encounter one of these names in a forum thread, this is the modern equivalent in the Classic dashboard:

• CORR / "rapid CORR" / canon_corr — daily-updated raw rank correlation with the canonical target. Lives in Classic, Signals, and Crypto.
• Alpha / Alpha v2 — Signals' rebrand of the current CORR. Same numbers, different name. Signals-only.
• CORJ60 — 60-day CORR target, winsorized to ±0.20. The primary CORR metric for Classic and the one the payout formula multiplies.
• CORR60 — raw 60-day CORR with no winsorization. Reported in Classic since round 846. Tracks CORJ60 closely (ρ = 0.77).
• CORR20 / v2_CORR20 / CORT20 — 20-day CORR family. Multiple historical formulations, all in Classic. v2_CORR20 v6 is the current daily-round version.
• MMC / MMC v6 — Meta-Model Contribution. Orthogonal-to-meta-model correlation. Classic and Signals.
• MPC / MPC v2 — Signals' rebrand of MMC v6. Identical numbers. Signals-only.
• BMC — Benchmark Models Contribution. Orthogonal-to-benchmark-cohort correlation. Classic only.
• FNC / FNCv3 / FNCv4 — Feature-Neutralized Correlation. Classic only. FNCv3 v4 is the current line.
• TC / True Contribution — retired after round 713. Appears in old leaderboards but is not currently scored.
• IC / RIC — Information Coefficient family. RIC v3 is the residual variant; Signals-only today.
• Season Score — aggregate quarterly score, sums across rounds. Classic and Signals.

Takeaways

There are really only four scoring metrics for Numerai Classic today: MMC, BMC, CORJ60, FNCv3. The rest are either renames (Alpha = Signals CORR, MPC = Signals MMC), historical (CORR v1, FNC v1, TC), or alternate time windows on the CORR theme.

The CORR family is internally redundant within a time window. v2_CORR20, CORT20, CORJ60, CORR60: they all measure rank correlation against the same target with different windows or clipping. Same-window pairs run 0.77–0.83; across the 20- vs 60-day windows, agreement drops to 0.48–0.61.

MMC and BMC are the contribution twins (ρ = 0.79). If you optimize one, you mostly get the other. Both have the steepest payout slope of any single metric in this dataset.

MMC and CORJ60 are the most independent pair worth optimizing jointly (ρ = 0.40). The payout formula weights them as 2.0 × MMC + 0.5 × CORJ60. A blended strategy captures both.

Metric values are not on a shared scale. A +0.02 MMC is in the top tail; a +0.02 CORJ60 is unremarkable. Always read percentile alongside raw value.

Numerai has reshipped scoring every 12–18 months. This page reflects what is current as of round 1260; the next iteration could appear without ceremony. The trends dashboard and submission score history are the places to watch for the next change.