Deep Learning and Trading - Outputs

Overview

One of the critical choices when desiging a neural network for trading is the decision of what the output variable is that is being modeled. The standard answer to this question is to model the assets return. That, is, the output is a continuous variable.

However, this is not the only choice. One could for example model a 0-1 type variable, which is 1 when the return is positive and 0 if it is negative. This type of modeling could justified by arguing that the exact return is impossible to know, but whether the market sentiment is positive or negative in general is more likely to be accurately measured, and it theory would require much less training data and/or time.

Alternatively, one could try to model a set of market regimes. In this case, the outputs would form a finite discrete set of outcomes.

Considerations