On the Links Between Forecasting Performance and Statistical Features of Time Series Applied to the Cash Flow of Self-Employed Workers

Proper cash flow forecasting is a complex task that can be done by modeling the cash flow data as a time series. Although parametric methods have been widely used to accomplish this task, they require some assumptions about the data that are difficult to hold. A well-founded alternative is the use of fuzzy inference systems, which have proven to be competitive in many practical problems. This paper presents a statistical study that compares the performance of fuzzy inference forecasting systems with that of a traditional parametric approach, in a cash flow forecasting problem based on the weekly income and expense data of 340 self-employed workers over a period of 338 weeks with 4 different time horizons (1, 4, 9, and 13 weeks). We also check for significant links between several statistical characteristics and observed performance, to determine which features might most affect the quality of the predictions. After finding that kurtosis is the most correlated feature, a more detailed exploration is performed on it.