This report "Trading With SVMs: Performance" sounds very interesting. The result on trading the S&P500 (without leverage, I think) since 1952 is 32.59% of annualized return while during the same period Buy&Hold is only 15.14%.

Here is Wikipedia's article on SVMs.

I wonder if anyone has used SVMs in real trading. Could you kindly share any experience about it?

Victor Niederhoffer writes: 

One of my very rare strengths counterbalanced by many much more glaring weaknesses is that I can usually look at any paper in statistical finance and find its weaknesses in 30 seconds or less. I looked at support vector machines in that context and could not figure out why it should do better than discriminant analysis or similarity analysis. I was turned off by the statement that armi garch results using the last 5 changes gave similar alluring results, as I don't believe they will work unless they are retrofitted during particular periods. The non-linear features that SVM tries to capture would not be predictable or repeatable I would think nor would it in any case be immune to ever changing cycles. I will look at this much more carefully with the doc before I can render a totally worthless opinion on this. 





Speak your mind

1 Comment so far

  1. douglas roberts dimick on December 30, 2012 2:21 am

    New Year Rules – to Govern Deterministic Systems?

    V, you might look at “C. Exchange Activity as Market Schematics?” (Pages 39-44) of my Quantitative Relativity sent to you, Christmas 2009.

    Note the section’s closing quip – “patterns that any given system may or may not realize during a
    given state.”

    Your hunch here with smv’s (i.e., nonlinear features and changing cycles) are two leads that might be followed based on the assumption that such aspects of learning machines do not escape the linear determinism — found in the generalization error — inherent during a stochastic process, such as with the kernel trick for subsequent transposition of sets within hyperplane(s)… Basically a multidimensional pattern recognition system but with dot products, yes?

    If one posits that dimensional space does not exist by a nonrandom system of coordinates, then a metaphysical (neither statistical nor quantum) architecture is necessary to extrapolate linear correlations from nonlinear states to include finite transitions and changes of relative states. The use of a linear classifier to decide (for example, linear regression of) mappings of (normal) vector space would be a case in point; eventually, one must correlate the functional margin (regardless of its size) to set classes that then define states.

    If such an instance, even in a corresponding set where there is an optimal maximum-margin hyperplane, a margin classifier places (in effect, arbitrary) decision boundaries… Hence your point on changing cycles as one reason why the nonlinear features may fail.

    It has occurred to me that program trading architecture, somewhere along the way, confused deterministic systems based on causality ( with deductive reasoning.

    I along with a control science engineer are presently constructing a logic sequence now. What one realizes, when attempting to define a system of coordinates relative to any given dimensional space(s), is that neither the physics nor the mathematics control the design and engineering of the machine.

    Happy New Year…


    Ps. Great article Leo… H NY.


Resources & Links