Feature selection methods for financial engineering

Abstract

I experiment with a well-recognized filter-wrapper hybrid feature selection method –

minimal-Redundancy-Maximal-Relevance Criterion feature selection refined by a

wrapper using Support Vector Machines. I apply this hybrid method to predict the stock

trend on 10 indexes on Singapore’s own Straits Time Index.

The method consists of two stages: firstly the minimal-Redundancy-Maximal-Relevance

Criterion feature selection is performed with datasets of 60 features and nearly 2000

instances extracted from Straits Time Index of Singapore and the top features are selected.

Secondly, a wrapper using Support Vector Machines then further generates the “optimal”

subset. Experiments are performed with a time series binary classification model, where

output is the stock price trend in the following day, either rise or fall and features are

technical indicators calculated using data from Yahoo! Finance Singapore. Lastly, 10-

fold cross validation is performed with the selected feature subset and accuracy rate

reports are generated.

Similar procedures are conducted with the same dataset but using different filter selection

methods such as information gain filter, correlation filter and consistency filter.

Comparisons are done among all 4 methods and the filter-wrapper hybrid method

generally outperforms the rest based on its higher mean accuracy rate and lower standard

variation.