Numerai Burn Rounds: When Models Fail Together

Round 464 burned 85% of staked models in a single week. The recent cluster around rounds 1160–1200 put three separate rounds (1164, 1169, and 1194) above a 72% burn rate. Burn rounds don't arrive alone; they arrive in clusters, and most models lose money in every one.

This post measures mass burn events in Numerai: how often they happen, whether they cluster, and whether NMR price warns you in advance. For background on how payouts and burns work, see Is Staking Profitable? or the live earned vs burned chart.

The Burn Question

Every round has winners and losers, and that spread is normal. A round where 45% of models burn is ordinary variance. A round where 75% burn is a field-wide stress event.

The burn rate (the share of staked models with negative payout in a round) swings from under 10% to 85%. Most rounds sit in the 30–60% band. Red-shaded periods mark rounds where burn rates spiked above 60%, and those spikes visibly cluster rather than scatter evenly across the timeline.

That clustering suggests common external pressure, rather than independent model-level errors.

The Worst Rounds

Which rounds inflicted the most widespread damage?

The worst 15 rounds all posted burn rates above 71%, topping out at round 464 (85.3%) and round 288 (83.6%). In the worst of them, fewer than one model in five earned a positive payout. You can pull up any of them on the rounds list.

The recent stretch is telling too: rounds 1164, 1169, and 1194 all landed in the worst-15 list, part of a dense cluster of 60%+ burn rounds around rounds 1160–1200. The peak rounds aren't isolated disasters, they're crests of multi-round storms. The same regime shifts bleed into round economics and the payout factor.

Do Burns Cluster?

Does a bad round predict the next one? If burns are serially correlated, participants should consider cutting stake after a burn round. If they're independent, each round is a fresh draw.

The serial correlation is strong: Pearson r = 0.77 between a round's burn rate and the next round's. Bad rounds really do predict bad rounds, and the hostile market conditions tend to persist for several weeks rather than resetting instantly.

r=0.77 still leaves real variance. Plenty of high-burn rounds are followed by normal ones, and moderate rounds occasionally precede surprise wipeouts. Treat it as a regime indicator, not a clean trading signal.

NMR Price as a Signal

Is there any relationship between NMR price and tournament difficulty?

Large NMR price declines sometimes line up with high-burn periods, but the relationship isn't causal in either direction. NMR price tracks crypto market conditions; burn rates track equity market conditions. The two are loosely correlated through shared macro exposure, but they're driven by different mechanisms.

The implication for stakers is uncomfortable: during the worst stretches, you can lose stake while the NMR you're staking simultaneously drops in USD value. That double hit is why position sizing matters more than chart-watching: see NMR token economics for the supply-side view.

Takeaways

Mass burn rounds are rare but real. Most rounds sit in a 30–60% burn band. Rounds above 70% (round 464 topped out at 85.3%) are better read as field-wide stress events than ordinary round noise.

Burns cluster hard. Serial correlation between consecutive rounds runs r=0.77. A burn round materially raises the odds the next one burns too, though the link is still noisy enough that you'll get surprises in both directions.

NMR price and burn rates move on different clocks. Crypto drawdowns don't reliably warn of tournament pain, and strong NMR weeks don't shield you from it.

Position sizing beats timing. Burn rounds can't be reliably predicted in advance, so the only durable defense is staking a size you can survive losing.