Stake-Weighted Age: Is Numerai's Meta-Model Getting Older?

Stake-weighted model age has climbed past 1,700 rounds. Here is what that means for newcomers, veterans, and the quality of Numerai's meta-model.

The Numerai meta-model is a stake-weighted average of all participants' predictions, so who holds the stake shapes the signal. Stake-weighted model age has now climbed past 1,700 rounds while the equal-weight average sits near 1,000. Veterans clearly steer the meta-model, but that does not automatically mean newcomers are locked out.

For background on how the meta-model is built, see The Meta-Model. For the live numbers behind the charts below, see the Models page.

How Old Is the Meta-Model?

Two averages tell different stories. The equal-weight version treats every model the same. The stake-weighted version gives more influence to models with more NMR at stake, so it reflects who actually shapes the meta-model.

Line chart comparing stake-weighted and equal-weight average model age in days, by round. Stake-weighted age climbs from roughly 500 days to over 1,700, consistently above the equal-weight line near 1,000.
Line chart comparing stake-weighted and equal-weight average model age in days, by round. Stake-weighted age climbs from roughly 500 days to over 1,700, consistently above the equal-weight line near 1,000.

Stake-weighted age sits roughly 600–700 days above equal-weight age for most of the tournament's history. Older models carry larger stakes, which is expected: veterans have had more rounds to compound NMR and more reasons to trust their approach.

Both lines trend up, but stake-weighted age is rising faster. Part of this is mechanical — surviving models simply get older — and part is selection, since the models that last tend to be the ones worth staking heavily on.

Does Age Equal Skill?

If older models score better, their dominance is justified. If not, the tournament has an incumbency bias problem.

Rolling 10-round average MMC broken out by model age cohort: under 90 days, 90–365 days, 1–2 years, and 2+ years. All four cohorts oscillate near zero within roughly ±0.006 MMC.
Rolling 10-round average MMC broken out by model age cohort: under 90 days, 90–365 days, 1–2 years, and 2+ years. All four cohorts oscillate near zero within roughly ±0.006 MMC.

The cohort picture is less dramatic than the stake gap suggests. All four age buckets oscillate around zero MMC within roughly ±0.006, and no cohort sits persistently above the others. The under-90-day cohort is visibly the most volatile — the cyan line swings widest on almost every wiggle — which matches the intuition that new models include both promising ideas and a lot of noise.

What does not show up is a clean monotone ranking by age. The 2+ year cohort is steadier but not systematically higher. Experience in Numerai buys consistency, not a scoring premium.

Can Newcomers Compete?

Every prospective participant asks this. If the leaderboard is locked up by entrenched veterans, there is little reason to join.

Rolling 10-round median MMC for newcomers in their first 10 rounds versus veterans with 50+ rounds. The two lines track each other closely, trading leads from round to round.
Rolling 10-round median MMC for newcomers in their first 10 rounds versus veterans with 50+ rounds. The two lines track each other closely, trading leads from round to round.

The score gap is small and not consistently in the veterans' favor. The newcomer line (orange) trades leads with the veteran line (teal) across the full history, and the two track each other closely through most of the tournament. Newcomers are noisier — their cohort is smaller in any given round — but they are not systematically worse scorers.

The gap has not widened over time either. Numerai is not becoming impenetrable to new entrants on scoring. What has changed is that both groups are operating against a higher baseline than they were 800 rounds ago.

Who Holds the Stake?

Scoring parity only matters if it translates into stake. The meta-model is shaped by stake weight, not raw score, so control of NMR is what drives influence.

Stacked area chart showing share of total stake by model age cohort over time. The 2+ year cohort (red) dominates, filling roughly 70–90% of recent rounds, with 1–2 year models second and sub-90-day models a thin sliver at the bottom.
Stacked area chart showing share of total stake by model age cohort over time. The 2+ year cohort (red) dominates, filling roughly 70–90% of recent rounds, with 1–2 year models second and sub-90-day models a thin sliver at the bottom.

The 2+ year cohort now holds somewhere around 70–90% of total stake in recent rounds. The 1–2 year group is a distant second, and sub-90-day models are a thin sliver at the bottom of the stack. The red band has been expanding steadily since roughly round 600.

This concentration of stake is not automatically bad. If older models are more proven, weighting them more heavily produces better aggregate predictions. The risk is the feedback loop: new entrants need stake to influence the meta-model, but stake accrues slowest to the models that have not yet proven themselves.

Takeaways

The meta-model is getting older. Stake-weighted average model age has climbed past 1,700 days and keeps rising faster than equal-weight age.

Age buys consistency, not scoring edge. All four age cohorts oscillate around zero MMC within similar bands. Older models are steadier, not systematically higher.

Newcomers score competitively. Median MMC for first-10-round models tracks 50+ round veterans closely, with no widening gap over time.

Stake concentration is the real moat. The 2+ year cohort holds the large majority of stake, so newcomer influence on the meta-model grows much more slowly than newcomer skill. Compounding stake takes time, which is why surviving long enough matters as much as any single round's score.