Numerai combines two things - a prediction competition, and yet another complicated and obfuscated crypto coin scheme.
The prediction competition is like the Netflix Prize [1]. They give you a data set and a prediction goal, you submit an algorithm, and there's a winner. OK, although if you can do well at that, you're wasting your time giving your results to Numerai. You're competing for a total payout of $15K/month.
Then, for no particularly good reason, this is combined with some smart contract / crypto coin scheme. Other than for PR purposes, it's not clear what the gain is. Everybody is dealing with Numerai; it's not a many-to-many system. It's more like airline reward points.
The hedge fund itself doesn't use Etherium; it's completely traditional.[2]
You don't submit an algorithm, you only submit your predictions. Your algorithm to generate your model predictions is not known to Numerai.
The data you work with from Nunerai is obfuscated, so you can't take it directly to the market.
It is an interesting no-risk way to evaluate models/algorithms. A lot of people are there to learn on a tough problem.
Edit: The reason for the coin is to allow data scientists to give confidence information about their models. There is an additional payout in USD associated with NMR staking (currently an extra ~$4000/mo for the most consistent models).
Very compelling. Without staking the hedge fund doesn't have any knowledge of how confident you are in your estimates. If you can put the value of a coin up against your confidence estimates then you are giving that information in a very real and binding way.
Do you have another way data scientists can give estimates of their model's ability to do better than chance and be held accountable for that estimate? Additional prize pool based on giving good reliability estimates seems like an excellent way to do it.
The tokens are not the value of the numerai fund itself. They are merely a tool for the staking process.
Yes...I remember a few years ago, with one of the algorithm testing sites, a group basically just submitted 20-30 pseudo-random predictions. One of them ended up winning, but it wasn't like the team actually had a useful algorithm, they were just playing the system because betting was free.
"Then, for no particularly good reason, this is combined with some smart contract / crypto coin scheme. Other than for PR purposes, it's not clear what the gain is."
Oh, it is entirely clear what it gains.
Cryto currency backed equity is just free money that you don't have to pay back if things go south.
There are basically zero protections for the "investors". Numerai could just walk away with the money and they'd probably get away with it.
Why wouldn't you turn down free money that the market is throwing at you?
> Although if you can do well at that, you're wasting your time giving the results to Numerai.
This is the fundamental break down in proposals like this (Numerai, Quantopian, etc).
For anyone who can consistently generate alpha, why bother with a crowdsourced model? They can just go join any one of the hedge funds that will pay half to several million each year in compensation for that acumen.
Sure. But that doesn't meaningfully negate my point. In fact it strengthens it - if you can generate alpha on the strange "homomorphically encrypted" data they've got, you can do it in the actual market with transparent data and more control far more profitably.
I'm not saying someone is going to do well and leave with the data. I'm saying anyone who can do well should probably just participate in the real market.
And as another point, Numerai lacks entire categories of data outside the conventional "core financial" that is useful for finding alpha.
> if you can generate alpha on the strange "homomorphically encrypted" data they've got, you can do it in the actual market with transparent data and more control far more profitably.
Sure. Now go get "transparent data"...oh wait, that's also time-consuming (and much less interesting). That's why, if one doesn't have to search for tradable data, it might be interesting to validate a signal/strategy on data that someone with stake says is tradable.
Agreed that if one starts making numerai money with numerai data, one should definitely go get some transparent data and start trading it. Trouble is, now one doesn't know what transparent data one has actually learned, so it's hard to know what to go and spend time on trading.
At that point, though, (I guess this is the theory, at least), numerai presumably know that one has a good signal, and one also knows one has a good signal for something, so both parties know that one's signals are valuable. That should be enough to get one a job at a place that does have good, transparent data.
Assuming that's a job which you want. Some people want to screw around with a project for a while and this is a good opportunity to test your skills on something interesting and real.
Worst case scenario you get some "free" money and some other people you don't know get rather more money out of it.
2 counterpoints:
1. buying data is expensive
2. running a fund is a completely different skill than building predictive models (back-office, compliance, etc.)
I would guess the crypto scheme is to cash out with ICO - those have been raising ridiculous amounts of money lately, for projects that sound like a joke.
My Mom suggested I start a coin where each coin is backed/tethered to one of the thousands of Beanie Babies in her closet. I told her the majority of cryptocurrency traders are too young to get the joke so it wouldn't get much traction. Then again Dogecoin started as an obscure joke and it now has $300Million+ market cap.
There is no ICO. Coin goes directly to competing data scientists. And that coin can be used to increase earnings of the data scientist depending on the consistency of their model.
People have been going to the slack asking for data scientist to sell them their tokens. Bidding was up to around $20/coin.
The main point, though, is that data scientists don't have to sell the coin. They can stake coin against their models to earn additional USD. If their model performs better than chance, then they get their NMR back to stake the next round. If they are also competitive wrt confidence in their consistency then they also get some additional prize money that is only available to staking participants.
If they are overconfident and do not do better than chance then the coin they staked is destroyed.
All these cryto ico things are premined or centrally allocated in a way where they get a cut. If you have access to the printing press, then you make all the money, no matter what happens with the currency.
Sure, maybe they gave away some of the coins to datascientists, but I am sure they kept a healthly cut for themselves.
One of my favorite things about the stock market is you know exactly how good you are at all times. Your bank roll is your track record and its very black and white.
Has anyone seen any returns for Numerai? They've been around for a while and I wish they would "put up or shut up" as the saying goes. I see that they have paid out money to developers but its not clear if the fund itself is generating alpha above their benchmark index or not.
One of my jobs is to review people who claim to have created algos that can generate alpha. Sadly it's very common for people to talk a big game about the returns they can generate, but then are unable to show any reliable track record when pressed.
Some of reasons I've heard:
- well we weren't actually trading, this is just all theoretical results
- well we didn't include the first 6 months of the year as we weren't profitable yet
- yes if you don't include theses particular trades we were profitable.
- yes we made money, its just that the market itself made more with less volatility.
I really would like Numerai to succeed, mostly because I'm interested in if the idea of mixing 1000's of models together and letting people vote on their weightings can succeed.
> One of my favorite things about the stock market is you know exactly how good you are at all times. Your bank roll is your track record and its very black and white
In the very long run, yes. But if you apply this to a short term (a week? a month?), that might lead to a rather fatal ego boost. I believe, that this is something that everyone interested in trading has to be aware of.
From my experience (i work in trading) that saying is more often used by people who had their ass handed to them by the market and don't want to admit they were wrong.
The market doesn't just go irrational. There are always reasons, just that sometimes they are non-obvious even in retrospect.
Yeah, I don't work near trading, but for me it has always meant that you can be "right" about the long term trend but really wrong about the timing. And the timing is just as important.
I think it's important to point out that this is their second attempt at making a hedge fund with crowdsourced investment strategies. Their first attempt failed because they would pay out rewards (bitcoins) based on the performance in a historical benchmark. This made the whole machine learning competition susceptible to intentional overfitting (common problem in the HF world). They outline the problem in their white paper: https://numer.ai/whitepaper.pdf
1) they always paid out USD values, but transfered equivalent Bitcoin.
2) they had problems with overfitting leaderboards based on validation sets, now they pay out based on live results resolved over 4 weeks. That was solved several months ago before the crypto currency went live or really had anything to do with it.
3) Payouts are now in both USD and NMR for the competition (based on live data).
So, they did have some problems at first, but that was solved by paying out on live results rather than a leaderboard based on validation data. That solution didn't have to do with the cryptocoin.
The purpose of the cryptocoin is that data scientists can earn more USD (in addition to the well functioning direct competition) by staking NMR. The NMR and conversion rate they require tells the fund how confident they are in their model. This additional confidence information, outlined in the paper you link, gives the fund a better idea of how to weight individual competition submissions for their meta-model that determines asset allocation.
It is an ETH coin. Read more carefully. The whole benefit of ETH over BTC is that you can build coins inside of ETH.
In this case it is basically just a colored coin, the point of which is for the people who hold them to be able make stakes on their long term performance (hence plain ETH would not be useful).
1) Bullshit marketing where they advertise the usage of "homomorphic encryption". If you've looked at their datasets they are most likely taking some fundamental and pricing factors and then performing min/max scaling. Nothing too crazy... and I think it would be possible to reconstruct which stock each observation corresponded to given some effort.
2) Professional quants care a lot about the way their features are constructed. The type of data you are looking at affects the type of model you would use... e.g. some kind of GARCH process for lagged time series data vs something different for sector betas, etc.
3) Given (1) there is an insider trading opportunity to reconstruct the underlying data and then submit predictions based on unknown or illegally sourced features.
#3 is actually a neat way for them to skirt insider trading laws. Insider trading is only illegal if you know that the information is material nonpublic (and a few other requirements). IANAL, so happy to be corrected if I'm missing something.
Well I know the guy that wrote it - he's one of the smartest and most talented people I know and he's not bullshitting. But by all means, try to break it!
What would be pretty cool is a DAO that takes in money and models, auto invests, and pays out to investors and data scientists. I don't really know enough about Ethereum to see if that is possible to encode into a smart contract.
Definitely possible. Conceptually you're just debundling the roles into discrete tasks with automated mechanisms that assign those tasks and reward completion of them.
The problem is that it's a complex undertaking to build it, because a lot of the building blocks have to be invented, and they have to be built to the highest standards, both in terms of efficiency, given the cost of Ethereum computing resources, and in terms of security, given the underlying platform is immutable and does not allow reversals of hacks.
That's like arguing the US dollar "struggles under the weight" of failed crowdfunded startups, ICOs being almost perfect analogues to them financially.
The only "sufferers" are the investors too dumb to do their homework. Thousands of failed real-world startups haven't killed SV's lustful search for the next big unicorn either; in both cases you have what's essentially a lottery and new suckers along every minute thinking they're going to make it big.
Maybe you could clear up something I've been wondering as someone new to cryptocurrencies. In the ICO, Bancor exchanges their own, newly created coins for Ethereum and the 'investors' are hoping that the new Bancor coins value will go up? So the Bancor team raised ~150 million dollars (in Ethereum) with no strings attached -- ie, no investors, debt, etc? Do I have this right?
This is accurate, all of the ether that bancor owns it can do whatever it wants with - the investors get nothing except a centralized token (BNT) that Bancor can control completely (create, delete, etc, see https://medium.com/unchained-reports/bancor-unchained-all-yo... ), the investors do not get equity or insurance or any other similar guarantees.
So, unless I'm grossly misunderstanding things, during the ICO, Bancor BNT tokens were purchased representing some fraction of the 150M contributed. Those tokens can now be exchanged by participants in Bancor's exchange, and purchased or liquidated freely to the underlying held cryptocurrency.
The ether raised during the ICO are used 1) fund various Bancor operations, and 2) fund an ether reserve to support liquidity.
But note, that reserve is just a fractional reserve. If there's a run on BNT, some people will lose if the reserve dries up! In their whitepaper example, of 200k ether raised, they mention a 30% fractional reserve.
You might ask yourself, why not just buy/sell ether directly? Well, the novelty, here, is that Bancor tokens could represent other cryptocurrencies or even baskets of currencies (though how the subsequent transactions on those other blockchains are handled, I have no idea, and their whitepaper doesn't go into it).
So yeah, I'd say your general description is accurate. If things go bad for Bancor, there's gonna be a lot of people losing out.
The prediction competition is like the Netflix Prize [1]. They give you a data set and a prediction goal, you submit an algorithm, and there's a winner. OK, although if you can do well at that, you're wasting your time giving your results to Numerai. You're competing for a total payout of $15K/month.
Then, for no particularly good reason, this is combined with some smart contract / crypto coin scheme. Other than for PR purposes, it's not clear what the gain is. Everybody is dealing with Numerai; it's not a many-to-many system. It's more like airline reward points.
The hedge fund itself doesn't use Etherium; it's completely traditional.[2]
[1] https://en.wikipedia.org/wiki/Netflix_Prize [2] https://www.forbes.com/sites/laurashin/2017/02/21/this-is-th...