Tuesday, May 09, 2006

Discussion board...

I'm thinking of creating a discussion board about NN and their application to financial markets... even though i got great feedback from readers of this blog i think we could try to discuss the issue together more often... could you please comment on this..
take care...

1 comment:

zendeni said...

Concerning the subject of NN for financial markets: what makes the financial markets apparently so much more difficult an application of NN? Almost every vendor (or free resources) of NN tools and frameworks always address the simple static classifier or function-approximation types of applications, and maybe a time-series application that is formulated as 'predicting' the next step in the data series. Time-series predictions might be appropriate for the econometics type studies, but if developing a 'trading system' for the financial markets is your goal then predicting the next price time-step is mostly futile or at least irrelevant to the problem. The so called 'reinforcement learning' paradigm is a much more relevent application or approach to building a 'trading system' for (possibly automatic) sequential (trading) decisions. The linear and stationary assumptions have to be thrown out (i.e.,ARIMA models do not apply). The 'markets' are notoriusly chaotic and non-deterministic, so it is futile to use any NN training procedure that forces a 'search' for fixed weights (that are supposed to model the markets over practical time periods). The 'reality-model- picture' of the markets must include multiple behavior modes (or multiple-dynamics that shift in dominance). In otherwords, the problem needs to be formulated in a way that will automatically identify maybe 2 or 3 different 'market conditions', and then separate NN weight-sets are determined (i.e., trained) for each of the different market conditions (or 'operating modes').

I believe that the 'Reinforcement Learning' paradigm is most correct way to formulate the financial market applications because this paradigm specifically separates the process-environment modeling from the trading-decision-policy. The market prices are the (changing-dynamic) environment, but the trading activity is only a calculated-decision 'response' to the markets. The two aspects should not be mixed up as just one 'trading model', such as typically handled. The trading-strategy should (ideally) be multiple strategies that are optimized for specific kinds of market conditions, but I see little evidence among existing NN-based 'trading systems' that this kind of problem formulation has been addressed yet. If so, I have yet to see it :>))

Has anybody seen any NN research that successfully models multiple modes of behavior? There has been some research like this by Faustino Gomez and Jurgen Schmidhuber at www.IDSIA.ch , that has been applied to non-linear control of 'robots', and uses genetic algorithms that are used to directly train Recurrent-NN weights, rather than depending on the back-propogation of the error gradient. This seems particularly important when the modeling involves both long and short-memory (modes?), because the gradient error signal vanishes after only a few time-steps and therefore makes long-term memory (inside the network topology) impossible, or unstable.

My skill level in mathematics and programming, and NN in general, are not (yet) up to speed to really carry these insights to the next steps towards implementation. If there are any people out there that can 'pitch in' some software-design and/or other related research-insights, we might be in a better position to attain some rather intersting financial goals :>)). What say?