A specification of a forecasting model in which relationships (coefficients) change over time. It may be difficult to identify when parameters change and a time-varying parameter model might make changes in response to false signals. Some researchers advocate time-varying parameter models. Riddington (1993) systematically evaluated research on time-varying coefficients in forecasting. He  "concludes conclusively that the [time-varying coefficient models] approach significantly improves forecasting performance.” He reached this conclusion by summarizing results from 21 forecasting studies. However, Riddington’s evidence is based only on ex post evaluations of forecast accuracy. (Ex post forecast evaluation can be useful for assessing how well models might predict the effects of changes in policy variables.) If the time-varying procedure provides substantially better parameter estimates, it might also improve ex ante forecasts. However, a common finding in this area is that refinements in the estimation of the parameters in econometric models do not contribute to ex ante accuracy. Time-varying-coefficients procedures are harder to understand, expensive, and may reduce the reliability of the model. Evidence that the parameters will change, or that they have recently changed, is unlikely to be found in the time series itself. If the structural changes are recent, then it is important to capture the changes. However, when one has only small samples (with perhaps unreliable data) and no domain knowledge data, the procedure may lead to a false identification of changes in parameters. Given the evidence to date, and modern computer capabilities, the analyst should simply rely on successive reestimation of models as more data are obtained, unless it is possible to use domain knowledge. (See adaptive parameters.)