Exploiting information in vintages of time-series data

The data measurement process for many time series is complex with an extensive set of revisions leading to multiple vintages of the same variable. How should the data measurement process be taken into account in the modelling and forecasting process? For example, the accuracy of the forecasts depends essentially on which vintage of data is used to construct the model and which vintage is used to evaluate the forecasts. Using historical series to estimate the model but more recent data to evaluate the forecasts, can give rise to equilibrium mean shifts and forecast failure. Combining the data measurement process and the data generation process allows consistency of approach in model building, evaluation and forecasting. The case of nonstationary data is considered here with, in the most general situation, multiple vintages on several variables. Simpler systems are used to motivate the ideas, which are illustrated with data for US income and consumption.