A trend estimate based on medians or modified means instead of arithmetic means. Thus, trends are estimated for a series of time intervals, say the trend from year 1 to year 2, then from year 2 to year 3, and so on in the calibration data. The median trend is then selected from these estimates for use in the forecasting model. Use of a robust trend requires three or more trend estimates. The forecast is the current level plus the estimated trend. The robust trend protects against outliers. Thus, it can be expected to be useful for noisy data. Little validation research has been done for the robust trend. However, Fildes et al. (1998) found that the robust trend produced fairly accurate forecasts (compared to other extrapolation methods) for some monthly telecommunications data (which, at the time, were characterized by declining trends). They also used a factor to take into account the size and sign of the differences between the individual trend estimates and their median. It is not known whether this adjusting factor contributes to accuracy. Following Occam's razor, the adjustment factor should be avoided until it has been tested.