Data mining in computational finance

Computational finance is a relatively new discipline whose birth can be traced back to early 1950s. Its major objective is to develop and study practical models focusing on techniques that apply directly to financial analyses. The large number of decisions and computationally intensive problems involved in this discipline make data mining and machine learning models an integral part to improve, automate, and expand the current processes. One of the objectives of this research is to present a state-of-the-art of the data mining and machine learning techniques applied in the core areas of computational finance. Next, detailed analysis of public and private finance datasets is performed in an attempt to find interesting facts from data and draw conclusions regarding the usefulness of features within the datasets. Credit risk evaluation is one of the crucial modern concerns in this field. Credit scoring is essentially a classification problem where models are built using the information about past applicants to categorise new applicants as ‘creditworthy’ or ‘non-creditworthy’. We appraise the performance of a few classical machine learning algorithms for the problem of credit scoring. Typically, credit scoring databases are large and characterised by redundant and irrelevant features, making the classification task more computationally-demanding. Feature selection is the process of selecting an optimal subset of relevant features. We propose an improved information-gain directed wrapper feature selection method using genetic algorithms and successfully evaluate its effectiveness against baseline and generic wrapper methods using three benchmark datasets. One of the tasks of financial analysts is to estimate a company’s worth. In the last piece of work, this study predicts the growth rate for earnings of companies using three machine learning techniques. We employed the technique of lagged features, which allowed varying amounts of recent history to be brought into the prediction task, and transformed the time series forecasting problem into a supervised learning problem. This work was applied on a private time series dataset.