Forecasting with structural change: why is the random walk model so damned difficult to beat?
This paper demonstrates that a naive forecast of no change can produce lower mean square forecast errors than those from true models when structural change exists.