Performance of 109 Machine Learning Algorithms across Five Forecasting Tasks: Employee Behavior Modeling, Online Communication, House Pricing, IT Support and Demand Planning

This article puts the problem of forecasting in economic and business situations under scrutiny. Starting from the premise that accurate forecasting is now a key capability for analyzing problems of business operations and public policy, we investigate the performance of alternative prediction methods that include both traditional econometric approaches as well as novel algorithms from the field of machine learning. The article tests a total of 109 different regression-type algorithms across five pertinent business domains – employee absenteeism, success of online communication, real estate asset pricing, support ticket processing, and demand forecasting. The results indicate that forecasting algorithms tend to produce a set of widely dispersed outcome, with some methods such as random forecast and neural network implementations being able to consistently generate superior performance. We further argue that forecast accuracy is not necessarily predicated upon computational complexity and thus, an optimization decision between the costs and benefits of using a certain algorithm can feasibly be made.