Resumen
Many studies examine the use of Neural Networks (NNs) as a tool for business time series forecasting, but the findings have been mixed and inconsistent. This paper explores the conditions under which NNs can improve business time series forecasting based on studies that compare NNs with traditional statistical models. The findings are that NNs generally outperform alternatives when data are nonlinear or discontinuous, building effective NNs for time series forecasting, including designing and selecting the structure, simulation functions, stopping rules, training algorithms and evaluation criteria remains challenging. A case study is discussed to illustrate these findings, and implications for future research and practice are also provided.