Does it make sense to use a “moving window” of calibration data?
Sometimes. In work done related to Chesapeake Bay, an asymmetrical 10-year moving window was used. Because of management needs, preliminary load estimates were given for the current year — year “10” — based on data collected during years “1” through “10”. These were then made final one year later based on fitting the model to a different 10 year window: the last 9 years of data plus the year of new data. Final estimates (which are for year “9” of the new 10-year window) are made using data from years 1 through 10. In addition to meeting management needs, there was a statistical justification for this: The quadratic approximation to the time trend in the 7-parameter model is arguably not quite right, and it is desirable to consider the impact of a higher-order series. However, any omitted third-order term will have roots at {1.13, 5.0, 8.87} (assuming uniform sampling, etc.; this is a consequence of orthogonal polynomials), which means the omitted term will have minimal impact in year
