MMC vs Correlation: What Numerai Actually Pays For
We compare high-MMC and high-CORJ60 models across every Numerai round to see which metric drives payouts and how stake concentrates.
The Numerai tournament rewards two things: raw predictive accuracy and original thinking. These map to CORJ60 and MMC, and the tension between them shapes every staking decision. Optimize for only one and you may leave NMR on the table.
Which metric actually pays better? The data has a clear answer.
What the Metrics Measure
CORJ60 (Correlation, journeyed over 60 days) measures how well your predictions align with actual market outcomes. High CORJ60 means your model predicts returns well. The distribution across staked models clusters near zero with a slight positive skew.
MMC (Meta-Model Contribution) is subtler. It asks whether your model still adds value after subtracting the crowd's combined prediction (the meta-model). A high-MMC model knows something the rest of the tournament does not. You can post strong correlation and near-zero MMC if your predictions are redundant with everyone else's.
CORJ60 rewards being right. MMC rewards being right in a way nobody else is. Both metrics are charted over time on the trends page, with per-round distributions on any round detail view.
The Payout Formula
Numerai's payout formula makes its priorities explicit:
MMC carries four times the weight of correlation. Numerai is a hedge fund, and redundant predictions do not help it trade. The meta-model needs diverse, uncorrelated signals, so the payout formula pays a steep premium for originality.
A model with mediocre correlation but strong MMC can outperform a high-correlation model that tracks the meta-model closely. Does that play out in the data?
MMC vs CORJ60: The Scatter View
The scatter plots each model-round observation with CORJ60 on the x-axis (-0.10 to 0.13), MMC on the y-axis (-0.08 to 0.04), and color encoding NMR payout. A few patterns stand out.
First, CORJ60 has a much wider spread than MMC — roughly 0.23 across versus 0.12. Raw correlation varies more model-to-model than meta-model contribution does. The two metrics are only weakly correlated: plenty of models sit at high CORJ60 with near-zero MMC (accurate but unoriginal) and a smaller group lives at modest CORJ60 with positive MMC (contributing something new).
Second, the highest-payout observations (greenest points) sit in the upper-right, where both MMC and CORJ60 are positive. Pure contrarianism in the upper-left — high MMC, negative CORJ60 — is thinly populated. Models that contribute uniquely to the meta-model tend to also clear a baseline level of accuracy. Numerai's benchmark models anchor what that baseline looks like in practice.
Do Big Stakers Optimize for MMC?
This chart compares stake-weighted averages against simple medians for both MMC and CORJ60 across rounds 200–1200. The same comparison is available interactively via metric distribution on the models page. The gap between the two lines shows how capital is allocated relative to skill.
For both metrics, stake-weighted values sit meaningfully above the median and swing more violently round-to-round. The median MMC and median CORJ60 both hover close to zero the entire time, while stake-weighted CORJ60 spikes as high as ~0.06 in several windows and stake-weighted MMC regularly runs 0.01–0.04 above its median. NMR is disproportionately concentrated in models that are both more accurate and more original than the typical submission — whether from sophisticated allocation or survivorship (weaker stakers burn down faster), the capital base leans toward skill.
As stake flows toward high-MMC models, the meta-model absorbs more of their signal, raising the originality bar over time.
Payout Distribution: Top-10% MMC vs Top-10% CORJ60
To make the comparison concrete, we isolated the top 10% of models by MMC and the top 10% by CORJ60 in each round, then plotted their NMR payout distributions.
The top-CORJ60 cohort has the wider spread and the fatter right tail, with individual observations reaching past 1000 NMR in a single round and outliers down to roughly -350 NMR. Raw correlation, when it hits, can produce extreme payouts — but it can also produce extreme drawdowns.
The top-MMC cohort is strikingly tight around zero. Originality delivers a steadier payout profile: most observations land in a narrow band a few NMR above or below break-even. High MMC rarely produces monster single-round paydays on its own, but it also avoids the deep negative tail.
The practical takeaway: correlation is the boom-and-bust axis, MMC is the steadier one. A large stake on a high-CORJ60 model captures the biggest upside rounds but wears the biggest drawdowns. Blending is what unlocks the 2x MMC multiplier without giving up CORJ60's ceiling.
Is the MMC-Payout Relationship Linear?
Bucketing models into MMC deciles and charting average payout per decile makes the relationship between MMC and pay direct. The bottom three deciles (0–30%) all average negative payouts, bottoming out near -0.2 NMR. Decile 40–50% is roughly break-even. From there the climb is steady and roughly linear, reaching ~0.55 NMR in the top decile.
The shape matters. Crossing from a below-median to an above-median MMC flips the sign of your expected payout — that is the single biggest jump. After that, every additional decile adds another ~0.05–0.10 NMR in average payout. Half-measures on originality do not pay much; committed differentiation does.
Strategy Implications
A few takeaways from the data:
- MMC flips the sign of your expected payout. The bottom four MMC deciles lose money on average; everything above the median earns. That is the single most important line to cross.
- CORJ60 is where the extreme rounds come from. If you want the occasional 1000+ NMR round, you need strong raw correlation. If you want to avoid the deep negative tail, correlation alone is not enough.
- Stake is already concentrated in skilled models. Stake-weighted MMC and stake-weighted CORJ60 both run well above the field median. You are not competing against the median submission, you are competing against the capital-weighted one.
- Originality compounds. As stake flows toward high-MMC models, the meta-model absorbs their signal, raising the bar for newcomers. Early movers on novel features and architectures have a structural advantage.
The tournament's incentive structure is unusually transparent: the formula is public, the data is available on every round page, and the outcomes are observable. Models that ignore MMC subsidize the ones that do not.