Improving forecasts for noisy geographic time series

Abstract

This study compares the performance of several simple top-down forecasting methods for forecasting noisy geographic time series to the performance of the three methods most commonly used for this problem: naive methods, Holt–Winters (exponential) smoothing, and the ARIMA (Box–Jenkins) class of models. The problem of producing weekly burglary forecasts at the precinct and patrol sector level in the city of Pittsburgh over a five-year period provides a case study for performance comparison. All top-down forecasting methods improve forecasting performance while significantly reducing the modeling workload. These results suggest that simple top-down forecasting models may provide a general-purpose method for improving forecasting for noisy geographic time series in many applications.