Bootstrap and nonlinear models applied to financial data

The bootstrap technique is a well-known method to generate multiple versions of predictors with the same structure. In this paper two different nonlinear structures are considered: neural networks and regression trees. They are both applied on real data related to the problem of predicting state bond price on the basis of the value of the previous auction and some financial indicators. Bootstrap is applied to the estimation set and the prediction abilities of both models improve quite significantly. The aim of this paper is to evaluate how the bootstrap features can be best exploited in order to improve the predictions. Experimental results show that by resampling the estimation set, nonlinear predictors outperform linear ones but for regression trees good results are achieved by merely resampling the estimation set while neural networks give the best results when the number of bootstrap repetitions is high.