Professor Hendry investigates the theory and practice of econometric modelling and forecasting in a non-stationary and evolving world, where the model differs from the economy.
When the processes being modelled are not time invariant, many of the famous theorems of both macroeconomic analysis and forecasting no longer hold. Conditional expectations cease to be unbiased predictors, and the mathematical basis of inter-temporal derivations fails, making dynamic stochastic general equilibrium (DSGE) models inherently non-constant and non-structural—failing when they are most needed.
A generalized taxonomy of forecast errors also reveals the central role of unanticipated location shifts in forecast failure, and helps explain the outcomes of forecasting competitions. Surprisingly, other potential sources of forecast failure seem less relevant. Co-breaking, corrections to reduce forecast-error biases, and model transformations all help robustify forecasts in the face of location shifts.
Fortunately, although model selection poses great difficulties, our recent research has revealed high success rates, and allows operational studies of alternative strategies. Automatic model selection algorithms can handle multiple shifts, embed theory insights, and avoid mis-specified models from omitting substantively relevant effects which would otherwise distort inference. Autometrics offers a viable approach to tackling more candidate variables than observations while controlling spurious significance.