I read Numerai’s atomic staking rollout and switch to 60-day payouts as an effort to buy predictions that remain useful for longer. That could support a larger hedge fund: a forecast that survives patient trading gives the manager more room to deploy capital. For NMR stakers, though, the change redistributes rewards. It does not establish that everyone earns more.
The distinction shows up in the data. In a matched historical cohort, 18% of models change the sign of their average contribution when measured over the longer horizon. A model that looks additive over one window can look redundant over another. Extending the target changes which research Numerai pays for, alongside changing when stakers recover their capital.
My capacity interpretation remains conditional. Public sources explain why Numerai values persistent predictions, but do not measure a capacity increase caused by this rollout. To assess the changes, I separate three questions: what the fund wants to predict, how it prices those predictions, and how much NMR must back the resulting commitments.
Three decisions arrived together in August
Classic began atomic staking with round 1343 on August 28, 2026. At the same time, payouts moved from Ender20 to Ender60. Numerai’s September monthly announcement confirms that the change reached production, including the move to 60-day diagnostics.
The rounds on either side of the boundary show the full change:
| Payout setting | Last legacy round, 1342 | First atomic round, 1343 |
|---|---|---|
| Opened | August 27 | August 28 |
| Ender target horizon | 20 market days | 60 market days |
| CORR multiplier | 0.75 | 3 |
| MMC multiplier | 2.25 | 15 |
| Payout factor | Approximately 0.0904 | 1 |
| Maximum gain or loss | 5% of shared round stake | 100% of separately funded position |
These are observed round configurations, checked September 9. CORR measures predictive correlation. Meta-model contribution, or MMC, measures a prediction’s contribution after removing its exposure to the combined prediction. The CORR and MMC guide explains why a model can score well on one and poorly on the other.
Ender itself was not new in August. Numerai had already changed the payout target to Ender20 in January, after describing the target’s purpose as rewarding predictions valuable to the business. The Faith II announcement separates that target redesign from the later extension of its horizon.
There is also a useful counterexample to treating blockchain as the cause of longer predictions. Signals adopted 60-day Alpha and Meta-Portfolio Contribution payouts in September 2025, under its earlier staking system. Its 2025 scoring announcement establishes that a longer target did not require atomic staking. The two choices can reinforce each other without being technically inseparable.
A persistent forecast can make a larger position worth trading
The strongest public support for the capacity argument predates this launch. In May 2025, Numerai said its hedge fund operated over a long horizon and that strong predictions across both horizons were easier to monetize. At that point it explicitly had no plan to change payouts to 60 days. The 60D scores announcement therefore suggests incentive alignment with an existing business need, rather than proof that the fund suddenly tripled its holding period in August.
Capacity is the amount of money a strategy can manage before costs and constraints consume too much of its expected return. Buying a larger position may move the price against the buyer. Spreading the order across more time can reduce that impact, but only if the prediction remains useful while the manager waits.
That is where a slower-decaying forecast can help. Its expected edge has more time to pay for entering and exiting the trade. The manager may also replace positions less frequently, reducing how much of the portfolio must pass through the market each day. These are two distinct benefits: more patient execution and less turnover.
Consider a deliberately simplified portfolio, replaced in equal daily slices:
| Assumption or calculation | Shorter holding period | Longer holding period |
|---|---|---|
| Portfolio value | $100 million | $100 million |
| Average holding period | 20 trading days | 60 trading days |
| Daily replacement purchases | $5 million | $1.67 million |
Replacement purchases equal portfolio value divided by holding period; corresponding sales are additional. In this illustration, the longer holding period cuts daily replacement volume by two-thirds. If daily trading capacity were the only constraint, the same trading budget could service a larger portfolio.
The condition matters more than the arithmetic. Gârleanu and Pedersen’s research on dynamic trading shows why optimal positions depend on forecast strength, decay, risk and transaction costs together. An extremely weak long-lived signal need not be more valuable than a strong short-lived one. AQR’s practical execution research makes the related point that waiting trades off lower impact against lost opportunity.
A longer target also cannot create stock liquidity or short availability. In his February 2026 Flirting with Models interview, Richard Craib described removing stocks from Numerai’s universe when better liquidity and borrowing data showed that meaningful positions were impractical. At about 41 minutes, his explanation connects target and universe design directly to what the fund can trade.
I therefore see a plausible route to more capacity, with several conditions still to test. The public record does not show how much additional money the new incentives can support after market impact, borrow costs, crowding and portfolio limits.
A 60-day target is not a 60-day holding instruction
Four clocks matter here: the return window used to score a forecast, the time its predictive advantage survives, the fund’s realized holding period, and the staker’s collateral lock. They need not match.
A model can predict returns over a long window while changing its rankings sharply between submissions. If yesterday’s preferred stocks become today’s weakest names, implementing that model may still require frequent trading. Conversely, a short-horizon model may keep selecting the same stocks and produce a relatively stable portfolio.
Numerai’s data documentation defines the targets in market days and explains their risk-adjusted construction. Those targets are training and scoring objectives. The portfolio manager must still decide how much to trade, when to trade it, and which exposures to neutralize.
This distinction sets a higher bar for the capacity thesis. I would want to see persistent predictive value alongside stable implementable positions. An improvement in CORR60 alone does not tell us whether turnover falls. Nor does it reveal how much of the forecast survives the fund’s risk controls.
Atomic staking makes the capital behind each round explicit
Numerai’s atomic staking announcement frames v3 as combining individually backed positions with automatic staking. The earlier continuous system reused stake across unresolved rounds. Atomic positions give each round its own funding, while allocation strategies handle repeated commitments and settlements.
That addresses a collateral-accounting problem even without a change in prediction horizon. As the number of unresolved rounds grows, separating their backing makes exposure easier to specify. Removing overlap also changes the meaning of a per-round return limit. A large percentage move on a small position cannot be compared directly with the old percentage move on the shared stake.
Classic’s documented allocation uses 64 concurrent positions. Funding 1 NMR per round therefore requires 64 NMR across a fully deployed allocation. Existing positions remain locked until settlement; setting future stake to zero stops new commitments. Constant mode targets the configured amount, while Compound reinvests settlements. Numerai still computes scores and publishes settlement proofs. These mechanics are described in the atomic staking documentation.
The migration immediately creates a measurement trap. Reported stake for a single round fell by about 98% at the boundary. The count of models with positive stake changed far less:
The first atomic round carried roughly 12,500 NMR. Multiplying that position principal by the documented overlap gives about 802,000 NMR, close to the preceding legacy round’s shared stake. This is an accounting illustration, not a measurement of total locked capital: migration timing, idle balances and changing allocations prevent an exact reconciliation.
The raw decline cannot establish mass withdrawals. A useful capital-flow analysis must follow funded positions and available balances through their lifecycle. Counting models also has limits: several models can belong to the same person, and continued submissions do not establish unchanged economic commitment.
The new coefficients place more weight on contribution
Before clipping, the atomic payout calculation is three times CORR60 plus fifteen times MMC60. The relative coefficient on contribution has increased. But a coefficient measures how the formula responds to a score increment; it does not establish how much reward the typical model receives.
To compare the incentives against fully funded capital, divide atomic coefficients by the position overlap. For the legacy comparison below, I use the payout factor from the final pre-cutover round:
| Effect of an equal score increment | Legacy coefficient per unit of capital | Atomic coefficient per unit of capital |
|---|---|---|
| Correlation component | Approximately 0.0678 | 0.0469 |
| Contribution component | Approximately 0.203 | 0.234 |
| MMC coefficient relative to CORR | 3 times | 5 times |
Holding score magnitudes fixed, the correlation coefficient is about 31% lower after the capital adjustment. The contribution coefficient is modestly higher. Historical payout factors varied, so these comparisons describe the boundary configuration, not every earlier round. The payout-factor note explains that dilution mechanism.
For illustration, a CORR60 score of 0.02 and MMC60 score of 0.004 produce a 12% return on one atomic position. With equal funding across the full allocation, that single settlement contributes about 0.19% of capital. Other positions have their own settlement dates and outcomes. Treating the position return as an immediately repeatable return on the whole balance would overstate the opportunity.
Greater emphasis on MMC fits an ensemble business. Numerai benefits when a model supplies information that the combined prediction lacks. Its MMC definition measures that residual contribution; it is not a direct accounting of the model’s dollar profit in the fund.
Setting the factor to one also removes the old formula’s automatic dilution as more stake enters. Under unchanged scores and coefficients, more funded principal would produce larger reward obligations. That gives Numerai a reason to calibrate the coefficients around the new capital mechanics and score distributions. A fixed factor does not fix the reward budget, and round-configured multipliers do not promise an unchanged future payout policy.
Longer scores change which models appear additive
I tested the horizon change on 273 Classic models with at least 10 NMR staked in every round of a fixed historical window. Each model has the same set of mature observations, so differences cannot come from comparing a new entrant’s favorable month with an established model’s full history.
The window covers round openings from January through early May 2026. Its 60-day scores finished by the end of July, well before this research date. These are historical predictions evaluated at both horizons, not settled results from the new atomic system.
The broad ordering survives. Rank correlation between models’ average contribution scores is 0.89, meaning the longer horizon does not simply replace the old hierarchy. Many relatively strong models remain relatively strong.
But 50 models cross zero on their average MMC. Most cross toward positive contribution; some move the other way. The finding changes how I would evaluate an existing model: a familiar reputation over the shorter horizon does not settle whether the same predictions deserve capital under the new formula.
The correlation panel also shows why identical numerical assumptions are insufficient. Many models’ longer-horizon correlation scores sit above their shorter-horizon scores. That can compensate for a smaller correlation coefficient per unit of capital. It must be measured rather than inferred from the multiplier change.
These observations do not establish that the fund earns more from those models. Average scores summarize a short historical period, multiple models may share an owner or training method, and overlapping return windows create common exposure. The exhibit identifies a change in scoring incentives, not an independent test of additional hedge fund capacity.
A payout improvement depends on whose history is counted
I next applied three payout schedules to the paired observations. The first uses the documented legacy formula and each round’s actual payout factor. The second changes only the coefficients, clipping and capital allocation while retaining 20-day scores. The third also substitutes the observed 60-day scores.
Every atomic position receives one sixty-fourth of fixed capital. I calculate each model’s mean payoff per eligible round, then compare the distribution across models. This is a formula diagnostic: it does not simulate a wallet, compounding, migration, funding shortages or the timing of cash receipts.
In the continuously staked cohort, switching the rules while retaining old scores lowers the median. Substituting the longer-horizon scores then raises it above the legacy result. That sequence supports a specific interpretation: the changed score distribution matters to the apparent improvement. Bigger multipliers alone do not explain it.
The lower panel deliberately relaxes the participation rule. It includes models with meaningful stake in at least 60 rounds, allowing a broader set of histories. The headline reverses: the median under atomic rules and longer scores falls below the legacy median. One basis point in the exhibit is 0.01% of the fixed capital denominator.
That lower median does not mean most individual models lose. About 60% of the broader cohort improve when I compare each model with itself. The median of each distribution can refer to different models; it is not the median change experienced by a matched model. Both measures belong in an honest account of who benefits.
The balanced panel offers cleaner date matching, but selects persistent participants. The broader panel captures more models, but different models contribute different subsets of dates. Neither can stand in for every staker. The comparison shows how a seemingly simple claim about higher payouts can depend on who qualifies for the analysis.
Numerai has published its own historical simulation over the same round range. Its payout histogram labels a much wider participant population, a fixed legacy factor and final returns on starting capital. My narrower cohort, round-specific legacy factors and fixed-capital per-round diagnostic answer a different question. The two charts should not be treated as competing estimates of the same return.
Most importantly, none of this demonstrates the return a new staker will realize. The first atomic Classic round is scheduled to settle in November. A historical exercise can expose the mechanics before that date; it cannot supply missing live evidence.
A longer scoring window also slows the feedback loop
The longer horizon creates a cost for researchers as well as an opportunity for the fund. A new model takes longer to build a fully resolved record. During that interval, a participant may keep committing capital while earlier submissions still have unfinished outcomes.
Separate positions do not make those outcomes independent. Adjacent 60-day targets share much of the same future market history. If a model has a persistent weakness, losses can develop across many open positions together. Spreading collateral across rounds changes its accounting; it does not eliminate common model risk.
That is also why I avoid annualizing these score comparisons or presenting a naive Sharpe improvement. Andrew Lo’s The Statistics of Sharpe Ratios explains how serial correlation changes performance inference and annualization. A longer score window can make a series look smoother without giving the researcher proportionately more independent evidence.
My concern is selection pressure. Longer funding commitments may favor operators who can finance the wait, even when another researcher has a useful model. Whether that reduces model diversity is an open empirical question. It would require tracking entrants, exits and independent contribution over mature cohorts, rather than reading confidence from the first week’s model count.
Signals makes the tradeability objective more explicit
The Signals history gives the clearest public example of Numerai changing rewards because a tournament metric was not delivering enough business value. In its 2025 Alpha/MPC announcement, Numerai said the Signals meta-model was not then sufficiently additive to Classic and its internal models. That was a statement about the system at that time, not a diagnosis of Signals in September 2026.
The forthcoming Supernova change addresses another part of implementation. Payouts for rounds opening from September 25 are scheduled to use neutral correlation, neutral contribution and a neutral churn penalty. The Supernova release says the penalty reduces upside for predictions that change too much to trade effectively. Its example halves payouts at neutral churn of 0.2.
This is more direct evidence of an implementation objective than extending a target alone. It makes the stability of submitted predictions affect what contributors earn. The new Jupiter target also incorporates changes to risk neutralization. Predictive accuracy remains valuable only insofar as it survives the transformation into a usable signal.
Classic received another relevant change in July: Quantum added 807 features while retaining the existing target list. The Quantum announcement explicitly said the payout target was unchanged at that release. Future improvements after August could therefore reflect richer data, changed model populations or market conditions alongside the new payout incentives.
As checked September 9, Signals’ atomic migration remained undated and USDC staking remained unimplemented. A scoring launch, a dataset release and a collateral migration have separate schedules. Combining them into one story of an already completed system would overstate the evidence.
A larger fund creates an indirect connection to NMR
There is a commercial reason to care about capacity. JPMorgan Asset Management secured $500 million of capacity in Numerai’s fund in 2025, according to Numerai’s announcement. Secured capacity is not the same as funded assets, and it does not establish a causal link to staking changes a year later.
Numerai subsequently reported approximately $700 million under management in its July 2026 update. That company-reported snapshot gives context for the scale question. Growth in assets can come from subscriptions or investment performance; neither by itself proves the strategy’s economically usable capacity expanded.
The token connection is more concrete on the purchasing side. Numerai’s July buyback announcement reported a completed $1.2 million purchase of NMR. It explained buybacks as replenishing the treasury used to reward contributors. Buying tokens for later payouts supports the incentive system, but those tokens can circulate again when recipients sell them.
I see the possible connection as a chain of business decisions: useful predictions support the fund; a successful fund can support spending on research incentives; treasury purchases create demand for NMR. Every step matters. The announcement does not establish an automatic percentage of fund revenue flowing to token holders.
Nor is staked NMR the capital used to buy the fund’s stocks. Numerai’s staking explanation says the stake remains locked and is not spent or traded by the fund. NMR provides a commitment behind predictions. Staking gives no ownership interest in Numerai or its hedge fund fees.
Burn accounting needs similar care. Atomic restaking can defer a model’s burns and offset them against later rewards. That debt is not liquid NMR, and a scored loss need not correspond to an immediate supply reduction. For token analysis, treasury purchases, reward distributions and finalized destruction are separate flows; the burn-history note provides the broader context.
The next evidence should test execution and participation
My strongest conclusion is that Numerai is tightening the connection between research rewards and predictions it expects to trade profitably. The capacity benefit is plausible, but unmeasured. The historical cohort results make the staker conclusion more conditional still.
The first atomic Classic settlement is configured for November 25, 2026. I would use the ensuing mature cohorts to compare net rewards against funded capital, including participants who stop or reduce stakes. Restricting the comparison to persistent survivors would miss precisely the selection effect that changes the historical result above.
The questions I would put to Numerai are more specific than whether the new system is better:
- How do Ender60 predictions change portfolio turnover and execution cost at comparable exposure, after controlling for Quantum and risk-system changes?
- How much additional capital remains profitable after market impact and borrow costs, and what assumptions support that estimate?
- How do staker outcomes change across starting capital, participation frequency and model contribution, including exits and incomplete funding?
- What policy determines treasury replenishment as reward obligations grow, and which parts remain discretionary?
Lower turnover would support only part of the capacity thesis if it came with materially weaker forecasts. Worse execution-adjusted performance as assets rise would challenge it. A sustained loss of independent contributors would also count against the idea that the new incentives improve the research supply. Those are the tests that can turn a plausible business explanation into evidence.
Sources and method
Research date: September 9, 2026. The migration exhibit covers every Classic round opening July 31–September 5; benchmarks are excluded and participation means any positive stake. The paired study covers R1172–1260, with all score definitions checked and all rounds resolved; 60-day scoring ended by July 31. It uses the explicit payout CORR20 field, not the obsolete empty column or CORJ60. Historical eligibility requires at least 10 NMR. Models and their eligible rounds are equally weighted; missing rounds are not filled with zero. The complete panel contains 24,297 model-round observations. Broader-cohort means have differing date coverage.
Legacy formula payouts reconcile for every observation in the complete panel. Five rows outside that panel report zero settled payouts despite nonzero formula values; they remain disclosed discrepancies in the theoretical comparison. No atomic live return enters the study. Neither the payoff diagnostic nor the turnover illustration is an annual return or fund-capacity estimate. Download model-level aggregates.
Numerai’s operating announcements and documentation are linked at each claim. The direct interview is Richard Craib with Corey Hoffstein, Flirting with Models, February 23, 2026; the implementation discussion begins around 41:23.
The outside research linked above explains mechanisms and statistical limits:
- Nicolae Gârleanu and Lasse H. Pedersen, Dynamic Trading with Predictable Returns and Transactions Costs (2013).
- AQR, Transactions Costs: Practical Application (2018).
- Andrew W. Lo, The Statistics of Sharpe Ratios, Financial Analysts Journal 58(4), 36–52 (2002).