Decomposition by Causal Forces: An Application to Forecasting Highway Deaths

Time series are often subject to conflicting forces; we refer to these as complex time series. This paper uses of causal forces in order to decompose complex series. In particular, we hypothesized three conditions to be important for effectively decomposing a time series by causal forces: 1) the forecaster has domain knowledge that can not be applied directly to the target series, 2) the domain knowledge can be used to structure the problem so that individual causal forces can be specified for one or more component series, and 3) it is possible to obtain relatively accurate forecasts for each component (relative to the target series). We tested decomposition by causal forces using forecasts of highway accidents. For 150 forecasts from two series that met the conditions, the forecast error was reduced by 16%. For 75 forecasts from one series that did not meet the conditions, decomposition would have increased the error by 42%.