Most forecasting methods are based on equally spaced data. In the case of missing observations the methods have to be modified. We have considered three smoothing methods: namely, simple exponential smoothing; double exponential smoothing; and Holt's method. We present a new, unified approach to handle missing data within the smoothing methods. This approach is compared with previously suggested modifications. The comparison is done on 12 real, non-seasonal time series, and shows that the smoothing methods, properly modified, usually perform well if the time series have a moderate number of missing observations KEY WORDS Non-seasonal time series Missing observations Forecasting performance Smoothing methods Most forecasting methods are based upon time series with equally spaced observations. Frequently a time series will contain missing observations, and then the forecasting methods have to be modified. (The series may also be observed at completely irregularly spaced time intervals, but we do not consider this situation here.) Much work has been done in this field during the last few years. The methods under consideration in this paper are simple and double exponential smoothing and Holt's method. They are all smoothing methods typically without any solid theoretical foundation. There are several reasons for their popularity. They are easy to understand, easy to compute and Holt's method, in particular, performs satisfactorily for series with many different patterns. These properties make the methods well suited for quick and rough forecasts without spending too much time on model building. A detail description and discussion of these methods when the time series are fully observed may be found in Gardner (1985). None of the simple methods have a well-defined underlying model. Thus an optimal method of handling missing observations does not exist. Wright (1986a'b) suggests some ad hoc modifications of the smoothing methods in the case of missing data. His procedure is discussed in the next section, where we also suggest an alternative way to handle missing data which gives a unified approach to the choice of weights after a gap and which, to a certain extent, is theoretically supported.
In the third section we compare the forecast performance of the various ad hoc techniques.