Conclusive Evidence on the Benefits of Temporal Disaggregation to Improve the Precision of Time Series Model Forecasts

Simulation methods are used to measure the expected differentials between the Mean Square Errors of the forecasts from models based on temporally disaggregated versus aggregated data. This allows for novel comparisons including long-order ARMA models, such as those expected with weekly data, under realistic conditions where the parameter values have to be estimated. The ambivalence of past empirical evidence on the benefits of disaggregation is addressed by analyzing four different economic time series for which relatively large sample sizes are available. Because of this, a sufficient number of predictions can be considered to obtain conclusive results from out-of-sample forecasting contests. The validity of the conventional method for inferring the order of the aggregated models is revised.