Comparing the Modeling Performance of Regression and Neural Networks as Data Quality Varies: A Business Value Approach

Abstract

Under circumstances where data quality may vary (due to inaccuracies or lack of timeliness, for example), knowledge about the potential performance of alternate predictive models can help a decision maker to design a business-value-maximizing information system. This paper examines a real-world example from the fteld of ftnance to illustrate a comparison of alternative modeling tools. Two modeling alternatives are used in this example: regression analysis and neural network analysis. There are two main results: (1) Linear regression outperformed neural nets in terms of forecasting accuracy, but the opposite was true when we considered the business value of the forecast (2) Neural net-based forecasts tended to be more robust than linear regression forecasts as data accuracy degraded. Managerial implications for fmancial risk management of mortgage-backed security portfolios are drawn from the results.