There are many disciplines that have contributed to how one model's risk and return. Physics introduced Brownian motion and RMT. Machine learning has helped to solve complex portfolio construction problems, operations research has contributed to market making & risk management, and aerospace engineering has contributed the Kalman filter.

I am looking for papers that use methods traditionally applied to bioinformatics and genetics to model the behavior of equities. For example, epidemiology makes extensive use of canonical regression -- Blackburn et al. apply canonical regression to tease out the co-movement of equities here.

I am not looking for links to neural network, SVM, or other machine learning papers. I am looking for creative and novel extensions of methods used traditionally in the bioinformatics and genetics space that are applied to generate alpha or risk models in U.S. equities?

For example, bioinformatics has developed a variety of greedy search alorithms that identify which types of traits give rise to certain phenotypes. Is there a paper that applies some of the algorithms from the bioinformatics domain to U.S. equities where traits might be financial statement variables and phenotypes might be return outcomes?

I know Carlos Carvalho (U. Chicago) has done both innovative work in the field of equities and also genetics and recall he had some research along these lines.

  • $\begingroup$ Pardon my ignorance, but what complex portfolio construction problems are you referring to that have been satisfactorily solved by applying machine learning methods? $\endgroup$ – neo123 Apr 30 '12 at 16:29
  • 2
    $\begingroup$ For instance non-convex utility functions (let's say you are maximizing alpha, minimizing CVAR, VAR, and transaction costs with a square root filter). Cannot solve this via QP, but can solve this via genetic algorithms. $\endgroup$ – Ram Ahluwalia Apr 30 '12 at 16:42
  • 1
    $\begingroup$ @QuantGuy genetic programming can really help in minimizing VAR, but minimizing CVAR of portfolio (surprisingly) turns out to be more computationally tractable $\endgroup$ – Alexey Kalmykov Apr 30 '12 at 17:39

You question is quite strange: so you do not want to use methods inspired by bioinfo and genetics (neural networks, GA, geometry of folding, etc) but methods that are used in these fields?

In terms of modeling, the problematics in bioinfo and genetics are mainly:

It is close to usual data mining, statistical or applied math problematics, so the main differences are heuristics, to take into account some specificities of the problems to avoid some local minima or increase the convergence rate.

I think that you not find anything new as you think to expect. It would be more efficient for you to identify your problems precisely, for instance trying to put each of them in one of these jars:

  • continuous optimization
  • discrete optimization (the state space has no obvious metric)
  • estimation in presence of uncertainty
  • continuous prediction (including backward and forward PDE problems, ARMA, GARCH, etc)
  • identification of outliers

Then review the classical methods (remind that even in Der Mann Ohne Eigenschaften, by: Robert Musil (1930), you have clear descriptions by Ulrich of his day-to-day job as a mathematician, you have descriptions of such modelings, so nothing is really new in all this), and try to develop heuristics adapted to the specificities of your real situations. My opinion that it is better to know very well the theoretical pitfalls and essential points of the methods to develop adapted heuristics rather than to read about heuristics developed for other fields.

  • $\begingroup$ I should have clarified my question as approaches inspired by these domains are certainly within the scope. Thank you for the thoughtful pointers $\endgroup$ – Ram Ahluwalia Jun 12 '12 at 18:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.