An Effective LmRMR for Financial Variable Selection and its Applications

Abstract

Financial variables are of primary importance in financial modeling, fraud detection, financial distress management, price modeling, credit and risk evaluations and in evaluating the return on assets and portfolios. There usually exist a large number of financial variables, where their exhaustive integration in a model increases its dimensionality and the associated computational time. We extensively tackle this problem in this paper. In this paper, we present a modified version of mRMR feature selection model to deal with financial features by ranking features first and then finding the best subset and uncertainty related to it using likelihood evaluation. The wellknown measurement formula of mRMR is considered for ranking financial features using correlation similarity measurement and the concept of minimum redundancy and maximum relevance of financial features and return of assets. Then, likelihood calculations inherently account for the mutual correlations between the variables as well as between the variables and the return on asset and result in a unique ‘likelihood’ value that has a correlation with the return on asset that can be maximized by adding and removing variables from the subset. We conducted experimental studies on Dow Jones Industrial Average to study the effectiveness and applicability of the proposed approach both in terms of financial variable selection as well as its application in Stock trading recommendation model and potential price forecasting. The performance is evaluated and the proposed approach shows promise.