How does staking NMR really work?



Hello all,

I’m trying to learn this system and get up to speed, but there is a glaring issue with the staking of NMR that I could use help understanding.

At the time of writing this, NMR is worth $28.57 on coinmarketcap. If I bought a single unit of NMR, I believe I could stake it against any model on the leaderboard this week. The formula for payout looks to be Amount of NMR / Confidence, so if I’m trying to turn a profit, I could set my confidence to 1/30 and the payout becomes $30. Lets say I expect the model to fail sometimes, so I place my confidence at 1/300. Now the model only has to succeed one out of every 10 weeks and I still turn a profit. What step in this process am I missing?

Thanks in advance


You can’t stake on anybody else’s predictions except your own. You also need to take into account that the payouts of the prize pool start with the highest confidence and works it’s way down the list. So if the first 10 people have depleted the prize pool before they get to your confidence level, the you will not receive anything.

Also a better place to check the price of NMR is on bittrex, much better than coinmarketcap.


What is the size of the prize pool? With so few people staking it seems easy for a low confidence stake to still get paid.

Also what counts as model success? Do I only need to place in the top 100, turn any kind of profit for numerai, or turn a specific amount of profit?


Prize pool is $3000 at this point.

I agree that there are very few stakes in the game at the moment, so anything is almost good to get some of the prize pool. But this is only the second week that staking is taking place, so a lot of people are not taking part yet, until they see how it all plays out. Remember that it’s 3 weeks from the close of the tournament until you find out how well you did.

Your predictions are evaluated on logloss on the live data. Anything worse than random (ie, logloss > 0.69314718056) will be considered as bad and the staked NMR will be destroyed.

The leaderboard has nothing to do with the staking tournament. In fact it doesn’t have anything to do with the regular tournament either. See the discussion here: Does the 4 week model evaluation affect the final leaderboard standings?


You’re being extremely helpful! Thank you so much!

Is that $3000 the same $3000 handed out each week? And my current prediction (linear regression in excel done in about 20 minutes) is 0.69247 so better than random, but fails the consistency check and the originality check. If I made no improvements would I have any chance to win a stake?
(I don’t intend to leave my prediction as this, I’m simply trying to understand how the system works)


The $3000 will be paid out every week to the winners of the staking tournament. We expect that number to increase in the future depending on how well the Numerai fund does. They have made adjustments to the payouts in the past and I have no doubt they will change those values in the future.

If your predictions don’t satisfy any one of the criteria, then they will not be considered for the regular tournament or the staking tournament. You have to pass all the checks (consistency > 75%, originality and concordance) to be considered for payout.


Linear regression isn’t really the right model to be using for a classification problem anyway, you’d want Logistic regression. But a simple logistic regression without doing much to tweak it or engineer new features has almost no chance of beating the originality check.

As themicon said if you don’t beat all the checks (originality, concordance, consistency >75%) you won’t be eligible to stake on your submission. Also, you can’t necessarily extrapolate from your log loss to actual live performance because of the high amount of noise and randomness in the data. The consistency check is really the more important bit when you’re dealing with staking, because it is the percent of eras in the validation data where your log loss is better than chance. If you’re not beating 75% on that, then you should have pretty low confidence that your model would beat random on the live data as well. If you can regularly get high consistency with your model, then you can be more confident of beating random on the live data.


@xmaslightguy some really good questions and even better responses from @themicon and @daenris, they’re spot on.

A consideration for me, in relation to your [quote=“xmaslightguy, post:5, topic:427”]
… is 0.69247 so better than random…

is to remember the staking payouts are based on an auction system. So even if you are better than random logloss (assuming you’ve also passed the other criteria), although you will not lose your NMR, you may not get a payout because the payout is exhausted before it gets to you. And even if you are confident you will do better than random, there is a real chance you won’t and you will lose your NMR.

So you could find yourself in a situation where you have no possibility of winning any $$, but you can lose your staked NMR. Someone could stake 2000 NMR at a confidence level that would put them at the top of the auction and sweep the $3000 pool for themselves, leaving all other stakers with nothing even if they did better than just random. I think (these are just my own thoughts) there is currently not much staking happening (less than 150 NMR currently and less than 2000 NMR previous tournament) because we are at a stale mate. The price of NMR is too high compared to payouts of the staking tournament / the payouts of the tournament are too low compared to the price of NMR etc.

When there are a couple of thousand NMR being staked, but ten’s of thousands up for sale on Bittrex, to me shows that people are more willing to cash in, rather than use their NMR for staking (at current price). I for one wouldn’t stake until the last day of the tournament to see what amounts and at what confidence levels others have staked, and assessed whether i’d even be in the running for a staking payout at my intended stake level. No point risking losing NMR if i can’t win a payout.


If the payout pool is exhausted before your stake is considered you don’t risk losing your NMR, it is just returned.


A good example of how NMR can fail big time just happened in tournament 61 staking results. A participant staked 220 NMR (market value at the time between $10,000 and $20,000) to win only $220 and failed being better than just random and had their 220 NMR destroyed. That’s big time gambling right there. I wonder if the participant knew the market value of NMR when they placed that stake.
The most staked in tournament 64 was 62 NMR. Who still thinks NMR is going to $800 each?


So I just learned something that wasn’t clear to me at the start and changes this whole dynamic for me. When a stake is successful, both the cash payout and NMR are returned. This means the value of NMR / confidence does not need to surpass market value since it can keep returning funds week after week (if the models are successful)