Video Of The Week: Richard Craib: Numerai

In this episode of This Week In Startups, Jason Calacanis talks to Richard Craib, founder of our portfolio company Numerai.

#blockchain#machine learning#stocks

Comments (Archived):

  1. awaldstein

    Gym opens at 8 this morning. This goes with me as my morning listen.Thanks Fred and have a great day.

    1. jason wright

      i can’t do that on my bicycle.

      1. Arnold Waldstein

        I ride every afternoon weather permitting and listen to podcasts most of the time

        1. jason wright

          yes, and i see that behaviour all the time, but i’m not participating. it feels too risky to me.…the strong speculation is that he was fiddling with his ipod and rode through the stop sign at the junction.

          1. awaldstein

            I live on the west side of manhattan.If I’m running chores on my bike or citibikes, not often.if I decide to ride up to the gw bridge on the bike path along the river or ambitiously around the island, most certainly.

  2. jason wright

    is the NMR token truly native, is it the immutable atomic unit of the Numerai process? i’ve spent some time on the Numerai Slack channel and the debate is roaring about the role of the token.

  3. Tom Labus

    Jason is such a great interviewer. He got this guy to explain the basics which never happens.The 2 of 2 and 20 covers a lot of incompetence.What if you get a lot of people to send in similar data and trade off that>? Or Some competitor pays you to do that?

    1. jasoncalacanis

      thanks pal!

      1. Twain Twain

        Jason,Here’s the most cutting-edge Machine Learning with sentiment analysis from Stanford applied to Twitter. That shows “state of the art” Natural Language Understanding is nowhere close to being solved, to support transaction and trading decisions. That was true in 2015 and remains true in 2017. It was also true before the 2008 financial crisis ($22 trillion in post-effect losses to US households): https://uploads.disquscdn.c…* https://www.technologyrevie…* https://www.technologyrevie…The inability to solve Natural Language Understanding is why Silicon Valley has NO WAY of dealing with the “fake news” problem as well as how ads are biased and irrelevant and impact the inclusion problem:*….

      2. Twain Twain

        Craib’s comment at 13:18 on “Even if you were to have all this data, you would still need to have HUMAN INTUITION about what to do with it. It’s sort of not really a quant thing” is the crux of the machine learning gap.These articles are worth reading:*…*…* https://weaponsofmathdestru…I’m a maths grad who worked for Professor John G Taylor (…, a pioneer who applied Neural Networks to the hedge fund sector long before Craib et al. We had models that ingested data from the stock exchanges and did pattern recognition.There is almost nothing in today’s ML models for hedge funds that I didn’t see when I was 22.The quants CANNOT take over the world for a very simple reason. Even the best quant+Machine Learning crew in the world at Google haven’t been able to and can’t solve the “Can Deep Learning help with Formal Reasoning (text understanding)” problem.Quants need the MBAs, the linguists and the liberal artists. https://uploads.disquscdn.c

        1. Twain Twain

          @jasoncalacanis:disqus @sigmaalgebra:disqus @wmoug:disqus — Everyone can see me asking Christian Szegedy of Google Research about state of the art Machine Learning for reasoning and text understanding, including sentiment analysis. Watch from 46m20s onwards.They literally threw the kitchen sink of maths, thousands of machines and DeepMind’s Wavenet at it.Yes, this is why I needed to fly 8000km to be in Silicon Valley: TO ASK THE RIGHT QUESTIONS TO MOVE THE ENTIRE INDUSTRY FORWARD IN SOLVING NLU.The answer is not in any existing quant method. An inventor would need to re-engineer about 2000+ years of data and science, that’s all — LOL.

      3. Twain Twain

        On the sentiment piece and whether emotions matters for trading predictions …Brian Uzzi, Northwestern University: “What we found was this. When traders are low in emotional states, they’re very cool-headed, they tend to make bad decisions. They’re too slow in taking advantage of an opportunity in the market, and they tend to hold on to bad trades too long. Exactly what you don’t want to do. We also found that when they were in a very high emotional state, they did the same thing. When they were at an intermediate level of emotion, somewhere between being cool-headed and being highly emotional, they made their best trades.”So it is NOT the case that the autistic AI will make better investment decisions. The best human decision-making has an element of emotional investment in it. Call it intuition / gut instinct / internal context etc., it adds perspective to good decision-making.https://insight.kellogg.nor…Moreover, where there is no financial context about the data and it is treated purely as a mathematical exercise, there are vital signal losses. Even being told that a data set is a time series rather than the delta changes in yields changes the choice of mathematical model used.

  4. Twain Twain

    Out & about so adding to must-watch list.Tangentially, interested to see what will happen now Google owns Kaggle.Quants & data scientists, indeed!

  5. Vendita Auto

    Thanks, Liking that a lot opens the fintech quants door to much ? Was for whatever reason reminded of the JJ Ulysses “Hanging by his heels in the odour of sanctity. Bell scared him out”

  6. Mark

    Andreessen said “software is eating the world.” The trustless digital scarcity enabled by blockchains has opened up a whole new part of the menu. Numerai is holding the spoon here. Machine learning is going to gobble up everything that Oracles expose, first via centralized Oracles like Numerai, then decentralized.

    1. sigmaalgebra

      > Machine learning is going to gobble up everything that Oracles expose,That’s just from hype: Machine learning is very narrow stuff, is basically empirical curve fitting. To look very smart, it needs too much data, more than is usually at all realistic. In no realistic way is machine learning, anything like it is now, at all a significant approach to real learning. Machine learning is about done with its gobbling. For the A16Z “software is eating the world”, software alone is about done. People have run out of important problems to solve that they can solve with just software or software with machine learning or artificial intelligence.Basically so far commercial software has written code to automate work we knew how to do, in principle and/or in practice, just manually. That’s too limiting.

  7. avip

    Great interviewee.The interviewer is abysmal though. The “I’ll speak more than you even though I know nothing about it” style hurts my ears. Please stop that. Let him speak.

    1. jasoncalacanis


  8. William Mougayar

    That was a very smart interview .

  9. sigmaalgebra

    The interview is a good example of how the Internet is enabling new media that often will be better than and an alternative to old media.

  10. sigmaalgebra

    Maybe machine learning can find models of the stock prices that will make a lot of money.Sure, regression can be improved on. For time series of stock prices, there was auto-regressive, integrated, moving average (ARIMA) processes. Sure, there is Brillinger at UC Berkeley. Then for a different approach, look for events, anomalies, triggers. IIRC James Simons mentioned such. For anomaly detection, I have some work, both multi-dimensional and distribution free — that I never intended for time series of stock prices.But I’m skeptical: To me just curve fitting, for an important example, (A) takes us back to Ptolemy and his epi-cycles, which apparently did fit the data on the motions of the planets but (B) totally sets aside Kepler’s observation about ellipses and, much worse, doesn’t lead to Newton’s second law of motion, law of gravity, and calculus. That is, Ptolemy was just curve fitting, empirical curve fitting, and not getting at the much, Much more powerful IDEAS of Kepler and Newton.After this example, and more, my view is that some good ideas, followed by some corresponding number crunching, stands totally to knock the socks off empirical curve fitting. For the origin or basis of the ideas, we know a lot about the stocks and their prices from difficult to impossible to see just in the time series data.I wanted to be a quant on Wall Street, but it appears that I tried too early. I got some interviews, but the idea that math could lead to better trading profits was about as welcome as proposing drinking snake oil.There is a fundamental problem: How will the old guard, still rich and getting richer from the best of the old ideas, be successful at, or even want to be successful at, picking some people, topics, techniques, approaches, etc. to replace the old ideas?Really, need a James Simons who comes in from outside, quickly makes some good money for himself, friends, and family, and grows from there — as he said in an interview “We made a LOT of money.” I can believe that. And I can’t imagine that any of the well known old houses on Wall Street have even a snowball’s chance in hell of competing with Simons. And, heck once I was in an NSF summer math thingy, met a guy, a student of A. Gleason, giving some lectures on differential geometry, got a copy of some S. S. Chern notes! Gee, that might have been close! Right, to fill in for most of the AVC audience, there is a famous result — Chern-Simons.I didn’t hear about Simons until basically too late. E.g., in E. Thorpe’s book on the market, in the back he gives an argument and says that it is routine “measure theory”. Okay, I already had a copy of Royden and plenty of prerequisites for reading it. And later in grad school I took a terrific course, right, measure theory, functional analysis, through martingales, etc. from a star student of, right, E. Cinlar. Wall Street never noticed. They might now! I enjoy Karatzas and Shreve — maybe if I’d pushed harder in that direction, but everytime I tried to talk to anyone associated with Wall Street about math for trading all I got back were dirty looks.So, I missed out on Wall Street!But now I have a better project — much cheaper to start, make much more money, easy enough for me to own it all, and much less risky! Less risky? Sure: For attacking Wall Street, I’d have to do some research, e.g., beat Simons and his people. That’s risky! For my project, I’ve already done the research, and the risk now is quite low.If my startup works, then I’ll be eager to dig into differential geometry and maybe cover the Chern-Simons result! Then I’ll get back to what I really wanted to do in college — mathematical physics! Seems to me that a lot of the stuff in physics is a mess; maybe I could clean up some of it.I have a conjecture: For positive integer n, random variables X_i, i = 1, 2, …, n, independent, random permutation Q on {1, 2, …, n}, and Y_i = X_Q(i), the Y_i are approximately independent and identically distributed. Here need to make clear the “approximately”; that’s part of the work to be done!. So, get into approximate independence. There’s some nice work by M. Talagrand that might be relevant and helpful. If some such is true, then maybe it could justify a lot of new applications of old statistical techniques to a lot of time series data. Ah, this is not related to my startup, so I just toss it out!

    1. jason wright

      Will you be issuing a token? 🙂

      1. sigmaalgebra

        Naw! Too complicated. So is a Delaware C-Corp.And a BoD? I could spend a lot of time important for my startup on, really, useless care and feeding of a BoD; there’s a lot from JLM and at AVC on how to do that, and the lesson I take away is that I don’t want to do that at all if I can avoid it. I don’t want to try to learn how to win at mud wrestling; instead I just want to stay away from the mud!Sometimes the simplest approaches are the best; here just remain 100% owner. At one time, some equity funding would have helped speed the work, but now it’s too late for that.There’s just no way to reach across the table: The latest I’d be willing to accept an equity check is too early for the equity check writers.E.g., I’m a sole, solo founder, a team of one (1), with a tiny burn rate. I don’t have a team of four co-founders, each with a maxed out credit card, each with a pregnant wife. And I don’t have a pregnant wife. I don’t even have a wife at all.And somewhere I get the strong impression that equity check writers have an iron-clad rule: Never but never invest in a one-person startup. Well, I have an iron-clad rule: Don’t have co-founders don’t really need. So far I don’t really need any co-founders, and later if I have some employees, I still won’t need any co-founders! And I have another such rule: Since co-founder disputes are one of the biggest threats to a startup, try to avoid having any co-founders!Today is normal — full of nonsense: The video signal cable on my VGA (analog signal) video display has an internal failure — the blue signal doesn’t get through. So, I get some goofy screen colors. For years I’ve been able to bend the cable just right, have the blue signals get through, and have normal screen colors, but yesterday the bending didn’t work. I have another display; this one takes a DVI (digial) signal. But, opening the box for the first time (got the display cheap), this display also comes with a cable for the 15 pin VGA signal! So, maybe that display also takes the old VGA analog signal. I was up late, and it’s a little early in the morning on Sunday to be working with video cables, so I’ll leave that for later today. So, that nonsense is too often the nature of the real work on my startup!But, no, I won’t be offering any tokens! I don’t even understand tokens very well although apparently AVC is a world center of token thinking and expertise! I’m just going to keep it simple — go live, get publicity, run ads, make money, try to make it grow. But, to go live, it will help to have my video working!Actually the nonsense was worse than that: I got into the back of my computer to get an inventory of USB 2.0 sockets so that I can plan a new keyboard and mouse since my old ones are failing and the new keyboard and mouse I’m considering both use USB 2.0 connections. The keyboard has several keys that for about a month worked only some times and now not at all. And the mouse too often gives double clicks when I want only one click.So, now that I took inventory, without a USB hub, I will be short of USB 2.0 sockets. So, looks like I need a USB hub, my first, a small one.USB inventory? Yes, I am using 1 USB 2.0 connection for each of a CD/DVD reader/burner, a B/W laser printer, a mouse, and an external 2TB Western Digital disk drive — that’s 4, and my motherboard offers only 4 USB 2.0 hubs. Besides, sometimes I download to my computer images from a little USB digital camera — unplug the CD/DVD player for that or just use the same USB port for the Western Digital since I use it only for backup and keep it unplugged in case of a ransom virus. So, yes, I need a small USB hub!Yup, system management and administration nonsense mud wrestling instead of the direct work on my startup!Heck, if I were a lone taxi driver, then I’d need to know how to change a tire! If I were running a grass mowing service, then I’d need to know how to clean dirt out of the gas tank. Well, I know all of those, but I’d rather work just on the startup itself!

    2. Twain Twain

      Barclays Capital offered me a Quantitative Risk role, Citigroup a Credit Risk role and MSCI a role with them so I’d have been a banking quant. But I opted for UBS because they offered me the Strategic Investments and corporate strategy role which was more wide-ranging and interesting.I learnt some business frameworks from that experience which has helped to develop more of a telescopic+microscopic view on the missing pieces in data structures and AI, and how to solve for them.Frameworks that the maths alone can’t provide.

  11. Twain Twain

    Basically, the prediction inputs from participants in Numeral’s competitions get fed into their Deep Learning model (as probability weightings, w, in layer 1). It then backwards and forward propagates across Numeral’s model. https://uploads.disquscdn.c… There are a few issues with the binary limits (between 0 and 1) of the originality and concordance criteria. They lend themselves to perceptron activation functions but there may be some signal losses. https://uploads.disquscdn.c…Oh and here’s what Craib meant when he talked about paying out to winners only if their logarithmic loss errors are the lowest. https://uploads.disquscdn.c

  12. duniajilbab

    Itu adalah wawancara yang sangat cukup cerdas.