A methodology for time series prediction in Finance

Aims to predict the Return on assets (ROA)of the company for the next year correctly and efficiently, this paper proposes a methodology called OP-KNN, which builds a one hidden-layer feedforward neural network, using nearest neighbors neurons with extremely small computational time. The main strategy is to select the most relevant variables beforehand, then to build the model using KNN kernels. Multiresponse Sparse Regression (MRSR) is used as the third step in order to rank each k nearest neighbor and finally as a fourth step Leave-One-Out estimation is used to select the number of neighbors and to estimate the generalization performances. This new methodology is tested on a toy example and experiment on 200 French companies to predict ROA value for next year.