Numerai Submission Scores: From Day 1 to Resolution

Historical scope reviewed September 8, 2026. These figures are a historical score-path sample with its original coverage. Score horizons resolve separately, and current submission-score ingestion is paused. A mutable endpoint value alone cannot reconstruct the preceding daily path or establish a completed 60-day outcome. See the current metric definitions.

Day-one MMC correlates with final MMC at just 0.21 — barely above zero. By day seven that climbs to 0.55, day fourteen to 0.80, and only at day twenty does it lock at 0.97. Numerai scores evolve over a multi-week window, and the gap between “early read” and “settled” is wider than most participants act on. The practical question is when the signal is stable enough to react to.

This post tracks daily MMC evolution using historical submission-level scoring data. The dashboard’s live rounds chart follows the stake-weighted Meta Model’s CORR scores, a different population and metric. For background on the metrics, see How Numerai Works and MMC vs Correlation.

How Scoring Works

Each round’s predictions are scored against realized market returns over a multi-week window, and updated daily as new return data lands. Early scores reflect short-term moves; later scores incorporate longer-term patterns. The final score feeds the payout formula alongside stake size and the tournament-level payout factor.

Day one reflects a tiny slice of the information that will eventually set payout — informative, but unreliable on its own.

The Score Journey

How does the average daily score evolve from round open to resolution?

Per-round average daily MMC from day 0 to day 27, with a bold cross-round mean hovering near 0.003 — Enlarge chart

Each thin line is one round’s average daily MMC across all models; the bold line is the mean across rounds. In the first week, individual rounds swing between roughly -0.04 and +0.025 MMC before the cross-round average settles near 0.003. Two rounds with nearly identical final MMC can take wildly different paths to get there.

Do Early Scores Matter?

The cross-model correlation between day-one and final MMC tells you how much early information is worth.

Per-round correlation between day-one and final MMC across rounds ~720 to 1215, 10-round rolling average around 0.2 — Enlarge chart

The 10-round average sits near 0.2, with individual rounds swinging from about -0.6 to +0.75. Day-one scores carry some signal, but the relationship is far from deterministic. Plenty of strong day-one models regress by resolution, and plenty of weak starters recover.

The information curve compounds fast. Across 600+ rounds of submission data, mean correlation to final MMC is 0.21 at day one, 0.46 at day five, 0.55 at day seven, 0.80 at day fourteen, and 0.97 by day twenty — most of the signal arrives in the second week, not the first. A model sitting in the top 20% on day three has a 39% chance of finishing in the top 20%: nearly 2x random, but still a 61% chance it falls out.

Volatility Across Rounds

Rounds differ sharply in how much daily scores bounce around.

Median per-model score standard deviation for rounds 1124 to 1219, rising from about 0.014 to a peak near 0.040 at round 1184 — Enlarge chart

Across rounds 1124 to 1219, median daily-score standard deviation climbs from about 0.014 in the 1120s to peaks above 0.033, topping out near 0.040 at round 1184. Patience matters most in these high-volatility rounds, when daily swings are largest and least informative.

When Rankings Stabilize

How quickly do model rankings converge to their final order?

Spearman correlation between daily model rankings and final rankings for five recent rounds, rising from roughly 0 to 1.0 across the 28-day scoring window — Enlarge chart

Spearman correlation between daily and final rankings shows clear convergence. Rankings start noisy — some rounds dip below 0 in the first few days — then climb into the 0.5–0.9 range by day 14. By resolution they all pin near 1.0.

Absolute scores stay unreliable early, but relative rank within this study’s daily-score sample stabilizes well before the round closes. That is distinct from the published reputation rankings discussed in the rankings guide.

Takeaways

Day fourteen is the inflection point. Day-one MMC correlates 0.21 with final MMC; day-seven correlates 0.55; by day fourteen it hits 0.80. Wait for the second week before treating a round’s signal as informative.

Volatility is a round property, not a model property. Market conditions drive most of the daily swing, and recent rounds have been unusually volatile.

Rankings converge faster than scores. Within this daily-score sample, relative rank settles well before absolute scores do. That result does not establish the stability of published reputation rankings.

Patience reduces avoidable churn. Participants who restake based on daily fluctuations risk reacting before the signal stabilizes. For a broader view of how payouts actually land, see Round Economics.