Model Archetypes: Five Species of Numerai Quant

Clustering 19,000 Numerai models by MMC, CORJ60, BMC, and FNCv3 reveals five archetypes. Contrarians retain at 71%; High-Volatility models burn out at 2x the rate.

Numerai's tournament hosts over 19,000 models, but they are not all playing the same game. Some chase MMC (meta-model contribution) by finding signals the meta-model misses. Others score consistently on every metric without excelling at any one. Still others spike in one dimension and collapse in another.

We averaged each model's four primary metrics -- MMC, CORJ60 (60-day correlation to the target), BMC (benchmark-model contribution), and FNCv3 (feature-neutral correlation v3) -- over the last 100 weekly rounds. Classifying by percentile signatures, five distinct archetypes emerge.

The Five Archetypes

Radar chart comparing median percentile ranks of five model archetypes across MMC, CORJ60, BMC, and FNCv3, with MMC Specialists extending farthest and Balanced models clustered near center
Radar chart comparing median percentile ranks of five model archetypes across MMC, CORJ60, BMC, and FNCv3, with MMC Specialists extending farthest and Balanced models clustered near center

Each archetype's median percentile on the four metrics produces a distinct radar shape:

  • Balanced (7,818 models, 40%) -- The largest group. Middle-of-the-pack on every dimension, with a roughly symmetric polygon near the center. These models avoid catastrophic weakness but do not stand out on any single metric. Many are likely ensemble approaches.
  • MMC Specialist (4,560 models, 23%) -- High on both MMC and CORJ60, with strong BMC and FNCv3. Their polygon extends outward on all axes, especially toward MMC and CORJ60. These models score well on meta-model contribution and correlation simultaneously.
  • Conservative (2,921 models, 15%) -- Above-median on all metrics but with the tightest polygon of the group. No extreme highs, no extreme lows. Optimized for stability rather than peak performance.
  • Contrarian (2,151 models, 11%) -- The most distinctive shape: high MMC but below-median CORJ60. These models contribute unique alpha to the meta-model because they diverge from the consensus prediction. The gap between their MMC and CORJ60 percentiles is the largest of any archetype.
  • High-Volatility (1,973 models, 10%) -- The widest spread across metrics. Some dimensions score at the 70th-80th percentile while others fall below the 30th. Concentrated bets on specific features or market regimes produce extreme scores in both directions.

Cluster Separation in PCA Space

PCA scatter plot of 19,000 models colored by archetype, showing MMC Specialists in the upper-left, Balanced models on the right, Contrarians in the lower-center, and a diamond-shaped boundary from percentile constraints
PCA scatter plot of 19,000 models colored by archetype, showing MMC Specialists in the upper-left, Balanced models on the right, Contrarians in the lower-center, and a diamond-shaped boundary from percentile constraints

Projecting the four metrics down to two principal components confirms these are not arbitrary labels. MMC Specialists (teal) cluster in the upper-left quadrant, Balanced models (blue) spread across the right side, and Contrarians (red) occupy the lower-center. The Conservative cluster (amber) sits in the middle, overlapping with Balanced.

The diamond-shaped boundary reflects a constraint of percentile-ranked data: no model can be extreme on all four metrics simultaneously. Specialists and high-volatility models populate the edges, while Balanced and Conservative approaches dominate the core.

Retention by Archetype

Horizontal bar chart showing percentage of models still active in the last 20 rounds: Contrarian 71%, Balanced 69%, Conservative 65%, MMC Specialist 38%, High-Volatility 37%
Horizontal bar chart showing percentage of models still active in the last 20 rounds: Contrarian 71%, Balanced 69%, Conservative 65%, MMC Specialist 38%, High-Volatility 37%

Of the models active in the last 100 weekly rounds, the percentage still staking in the most recent 20 rounds reveals which strategies persist. The spread is nearly 2x from top to bottom.

Contrarians lead at 71%. Their unique alpha and above-average MMC produce enough payouts to justify continued staking, even though their CORJ60 lags. The tournament rewards differentiated signals.

Balanced models follow at 69%. Generic approaches have low downside variance. Without dramatic losses, operators have less reason to withdraw.

Conservative models retain at 65%. Stability keeps them in the game, though their lower ceiling may explain the gap versus Contrarians.

MMC Specialists and High-Volatility models share the bottom at 37-38%. This is surprising for MMC Specialists, given their strong absolute scores. The likely explanation: these models concentrate stake on high-conviction predictions. When the payout factor or market regime turns against them, losses hit harder. High-Volatility models face the same problem amplified -- extreme bets mean extreme drawdowns, and drawdowns kill participation.

Archetype predicts model longevity as strongly as raw performance does.

Payout Reality

Horizontal bar chart showing median cumulative NMR payout over 100 rounds by archetype, all near zero with MMC Specialist slightly positive at less than 0.001 NMR
Horizontal bar chart showing median cumulative NMR payout over 100 rounds by archetype, all near zero with MMC Specialist slightly positive at less than 0.001 NMR

Median cumulative payouts over the 100-round window hover near zero NMR for all five archetypes. This matches the broader staking profitability picture: the payout structure produces a near-zero median return, with profits concentrated in the top quintile of each group.

MMC Specialists show a marginally positive median, consistent with their strong metric scores. But archetype alone does not guarantee profitability. It determines the shape of your returns distribution, not the center.

Practical Takeaways

If you operate a single model, check which archetype it falls into. Compare your average MMC and CORJ60 percentiles over the last 100 rounds on the models page. If your MMC is high but your CORJ60 is below median, you are a Contrarian -- historically the most durable strategy.

If you run multiple models, diversifying across archetypes may matter more than diversifying within one. A Contrarian paired with a Conservative has less correlated drawdowns than two models from the same cluster. The diversification paradox applies at the archetype level too.

If you are starting out, the Balanced archetype's 69% retention shows that generalist approaches survive without requiring a strong thesis. But they produce median payouts near zero. The Contrarian path -- building signals that deliberately diverge from the consensus -- offers better retention and marginally better payouts, at the cost of lower CORJ60 scores that can feel discouraging round by round.

The meta-model does not need more of the same. It needs models that see what others miss. The data confirms the tournament rewards that with longevity.