Historical scope reviewed September 8, 2026. The historical residual subtracts mechanical payout effects from stake changes and uses the stated settlement anchor. It remains an allocation proxy, not verified voluntary deposits or withdrawals. Atomic claims, internal migration and allocation settings require separate event classification. See the current metric definitions.
Every Numerai burn round raises the same question in chat: will stakers pull their NMR now? The stake series supports a descriptive comparison. Anchoring each resolved round at the moment its result lands (resolution, roughly 30 days after close) and tracking the total stake pool over the following weeks shows an average decline of about 1.3% after worst-decile losses in this historical sample, fading within ten weeks. It does not identify an investor withdrawal or a causal response.
The setup: for every resolved Classic round with stake, compute its net payout ratio (earned minus burned, over stake at risk). Then sample the total NMR at stake at each subsequent weekly round open. Final settlement status is the study anchor; provisional scores can be visible earlier.
The Long View: Burns Barely Register
First, the raw picture — the stake pool with the worst 5% of burn resolutions marked.
The pool's big moves — the 2021-22 climb to near 1M NMR, the 2023-24 decline, the 2025 recovery — do not line up neatly with burn events. The dense cluster of red marks in mid-2023 sits inside a decline that had already started, and the deepest pool trough (433K NMR, late 2024) arrives more than a year after the worst burn cluster. For the burn history itself, see Anatomy of a Numerai Burn Streak.
The Event Study: A Dent, Not an Exodus
Averaging the pool's path after every resolution, split by outcome decile, compares the observed paths.
After a worst-decile burn (106 events), the pool drops roughly 1.3% over the next four weekly rounds, stays down through week six, and is back to flat by week ten. After a typical round, it drifts up 1.5% over four weeks. The gap between those two paths — about three percentage points — is a descriptive difference rather than an identified burn-driven outflow, and it is comparable in size to the burns themselves, which run as deep as 4% of stake in the worst round.
One caveat on the top line: the +18.6% ten-week path after best-decile earns mostly reflects when those rounds happened. Huge net-earn ratios cluster in 2020-21, when the pool was growing fast for reasons that had nothing to do with any single round. Read that line as era, not incentive.
Round by Round, the Signal Is Weak
The scatter below drops the hyper-growth era entirely and correlates each 2022+ round's outcome with the pool's change over the following four weekly rounds.
The correlation is positive but weak: r = 0.18 across 973 rounds. Outcomes explain about 3% of the variance in what the pool does next month. Everything else (NMR price, model launches and retirements, the individual whale decisions tracked in Stake Flows) swamps the burn signal.
What the decomposition cannot identify
The former figure labeling the payout-adjusted residual as voluntary flow has been withdrawn. Subtracting economic payouts from changes in round stake does not isolate external deposits or withdrawals. Model entries, missed submissions, internal transfers and differences in observation timing can all enter the residual.
These comparisons retain their original pre-atomic sample. Atomic migration changes the principal allocated to each round, so the same residual would also mix capital-basis changes with allocation decisions. Use reconciled balances and transaction events before interpreting capital entering or leaving the system.
The historical average path is useful as a description of that sample. It is not an estimated causal response, a forecast of atomic-era withdrawals or proof that participants ignore losses. See the stake-behavior correction for the records needed to test those claims.