Who Moves Their Stake? Two Years of Numerai Staking Behavior

When NMR moves into or out of a Numerai stake, one of two things happened: the protocol settled a payout into the balance, or a person decided to act. The pool-level event study showed that voluntary flows, not payouts, drive the aggregate pool. This post pushes the same question down to the individual model: who actually adds and withdraws, when, and what triggers it.

The setup: for every Classic model, sample its stake at each weekly round open from July 2024 through June 2026. Payouts settle into the balance when a round resolves, 33 days after it opens, so binning each model's settled payouts into the week they land splits every stake change into a mechanical part and a deliberate part. The split checks out: across 404,219 week-to-week transitions the residual spikes at zero (72.8% under 0.01 NMR), exactly what payout compounding predicts for stakers who never touch anything. One composition note before reading passivity off that number: over half of positive stakes are sub-1-NMR remnants of 2023–24 onboarding credits, so the behavioral analysis below keeps only stakes of 10+ NMR. A "deliberate action" means the residual exceeded both 1 NMR and 2% of the stake, well above measurement noise. Even among these real stakes, 95% of model-weeks show none.

Two years of flows, decomposed

Summing the three components across all models reproduces the pool's monthly change exactly.

Over the window the pool grew 34.5%, from 556,720 to 748,787 NMR. Protocol payouts contributed a steady +334,825 NMR. Deliberate flows added +168,148 net, but that netting hides 578,620 NMR of top-ups against 410,512 NMR of withdrawals. Churn ran negative: stake entering with new models trailed stake leaving with departing ones by 310,906 NMR.

The deliberate component moves in waves. The second half of 2024 was a withdrawal regime (November 2024 alone saw 37,530 NMR of gross withdrawals). February through May 2025 was a deposit wave, peaking at 99,979 NMR of gross top-ups in May, including the window's largest single move, one model adding 31,339 NMR in a week. August–September 2025 brought another ~40,000 NMR/month of withdrawals, and spring 2026 another deposit wave. One caution on the churn bars: most "exiting" stake is models skipping submissions and returning later. Of the 1.77M NMR that left the observed pool, only about 450K belonged to models never seen again in the window.

Losses freeze stakers; wins get trimmed

How does a model's own recent performance change what its owner does next week? Sorting model-weeks by trailing four-week settled return:

The frequency curve is not the "flee after losses" story. After the coldest streaks, stakers are at their least active: top-up probability bottoms out at 1.0% per week and withdrawals sit near 2%. Action peaks after the hottest streaks: 3.5% top-up, 4.5% withdrawal.

The sizes tell the other half. When someone does withdraw after a cold streak, the median cut is 54% of the stake — capitulation, not trimming. After hot streaks the median withdrawal is 3.6% — profit-skimming. Top-ups are largest for models with flat trailing results (+45% of stake), which is mostly ramp-up of young positions. Fresh stakes with no settled history yet are the most active of all: 6.8% top up and 5.1% withdraw per week, with median moves near ±90%: new stakers either commit fast or bail fast.

Small stakers chase, large stakers harvest

The same hot/cold split, cut by stake size, shows two opposite playbooks.

After a hot streak, small stakers (10–100 NMR) add at 3.4% per week and rarely cut (1.8%): classic performance chasing. Large stakers (>1K NMR) do the reverse: 7.4% withdrawal probability against 3.2% top-up. They treat a hot streak as a chance to take NMR off the table, consistent with the volatility tax logic that oversized stakes compound badly through burns. After cold streaks everyone freezes; no tier's top-up rate exceeds 1.3%.

Weight matters here: the >1K tier moves 64% of all gross deliberate NMR in this panel, the middle tier 28%, small stakers 8%. Aggregate flow behavior is essentially what large stakers do.

Whose results move the money?

A staker sees three signals: their own results, the tournament's overall payout regime, and the NMR price. Splitting each into quintiles:

Deposits respond most to the pool's trailing return: top-up probability climbs from 1.4% in poor payout regimes to 4.8% in the best, the steepest slope on the chart. In a joint regression, one standard deviation of pool-wide trailing return moves weekly deposit probability by +0.91pp, versus +0.16pp for the model's own return. Withdrawals are more evenly driven (own +0.27pp, pool +0.32pp per SD). The payout factor mechanism makes this collectively self-defeating: capital arriving after good pool regimes dilutes the very payouts that attracted it.

The NMR price is the weakest signal, with a contrarian tilt: both actions spike after four-week crashes of 20%+, and top-ups are least likely after pumps. Numerai stakers buy dips; there is no visible take-profit-on-pump behavior, echoing the weak price–activity link in six years of price data.

Method notes

Stakes come from per-round selected_stake_value at weekly opens; settled payouts are binned into the weekly window containing each round's resolution (uniformly 33 days after open). The residual attributes a change to the staker, so a transfer between two models of one account shows up as a withdrawal plus a top-up. Withdrawal timing reflects when the balance changed, not when the request was made. Benchmark models are excluded, and the behavioral panels drop stakes under 10 NMR (89,269 model-weeks across 2,700 models), a floor that also removes the 0.1 NMR onboarding-credit stakes handed to 2023–24 signups, which would otherwise pad every passivity statistic.

Takeaways

• Passivity is the norm: among stakes of 10+ NMR, 95% of model-weeks show no deliberate move, and 52% of those models never acted once in two years.
• Loss responses are binary: freeze or capitulate. If you're cutting a cold model by half, know that's the modal exit, and it usually happens after the damage settled.
• Hot streaks trigger more withdrawals than deposits among large stakers. The people moving the most NMR harvest wins; the smallest accounts chase them.
• Deposit flows chase the pool's payout regime, which the payout factor then dilutes. A staking plan indexed to your own model's evidence, not the tournament's recent mood, avoids the crowd's timing.

Live staking data is on the models and tokenomics pages.