Bayesian Machine Learning

Discussion in 'Professional Trading' started by yimyom, Sep 17, 2011.

  1. yimyom

    yimyom

    Hi,

    I come from the academic world with a background in Bayesian Machine Learning. What firm would you recommend me to apply to in order to go in the world of automated trading (knowing I already have a professional experience in automated trading too).

    It seems, many firms are not aware of this type of new techniques that work pretty well in other industries. Even if it's not a proof that it can work too for automated high-frequency trading, I wonder which firm would be interested working on that or at least giving it a try, hoping for results (maybe maybe) as good as Google Ads System or Amazon's Recommendations.

    Thanks for your help
     
  2. Occam

    Occam

    Ask on the careers section of Nuclearphynance or Wilmotte.
     
  3. heech

    heech

    I don't think Bayesian pattern recognition is really a "new" field/technology... it's been used in various fields for decades, and I'm sure the low-hanging fruit in finance was snatched up long ago.

    I will say, my impression is that this is what RenTech does. I think this is a more interesting field than blindly training neural networks. I also think the various applied math quants out there that focus on theoretical modeling/pricing is at sort of a dead-end... not much room for marginal improvement there.
     
  4. yimyom

    yimyom

    That's indeed interesting to have an opinion from a practionner.
    Indeed, Bayesian techniques date back from.... Thomas Bayes in the 18th century even if they really started to develop in the 20th.

    What I'm talking about are very recent techniques like probabilistic graphical models. I believe they are plenty of applications to research in finance and especially in high-frequency trading. My reason is that the Bayesian paradigm allow for the construction of models with implicit knowledge, basically, the knowledge of the trader. When you think about very high-frequency data, it's hard to make stat-arb strategies due to the nature of the data. It makes more sense with low frequency trading. But you can incorporate your knowledge of a specific market or a specific asset by including some prior distribution on the variables you are modelizing. Instead of just thinking in term of a distribution on X with a parameter alpha, you think in term of the distribution of X and alpha and you can deduce P(X|alpha) or P(alpha|X). Okay, my explanations are really fast and Bayesian models are way more than that, but it's just a little post.

    Of course, this doesn't make a stragegy, as one has to make many other considerations in term of signals, limits and executions. This are crucial for a strategy to be succesful and it's more logic and common sense than Bayesian techniques.

    Here Bayesian models can help make predictions, in fact that's the basic idea of Bayesian models.

    I think it's also interesting at tick data, because it allows the modelization of discrete data because to be honest, at tick data, it's hard to assume the continuity of prices and time ! And because of the capability of Bayesian model to deal with huge data sets, there are plenty of techniques (sparse methods, penalization, just to cite a few, ....) that can help making good models with this huge database.

    I would love to see people's reactions and development on that topic and I'm happy to talk about my ideas
     
  5. toho

    toho

    Please do. I would be interested in knowing what you intend to model specifically.

    I would have to agree with heech, though, that this is not something new to the field. Also, the thing about financial data is that the signal to noise ratio is so low that even with vast amounts of data your estimates will tend to depend very much upon your priors and assumptions, rather than converge to the true distribution, whatever that is. That doesn't necessarily mean that Bayesians methods are useless, but I don't think good results will come easy.
     
  6. I can attest that Bayesian techniques have definite applications in this field. Our company has designed a few products that incorporate Bayesian methods. I am actually a big proponent of this approach. It is very instrumental in making decisions in the high level of uncertainty. Below is a useful link for those who are interested.

    http://www.ipam.ucla.edu/publications/gss2007/gss2007_7200.pdf

    Cheers,
    MAESTRO
     
  7. rosy2

    rosy2

    are you automated, profitable, and intraday? if so, almost any firm would do some kind of deal with you.