A method is presented for dealing with a system where a signal is observed during a recurring event by a varying number of samples, which form a time-variant trend during the event. The main purpose is to predict the evolution of the signal from one event to the following events, while the event-internal trend is of secondary interest. On the basis of the observations, the method quantifies and tracks changes in the trend by a recursive algorithm suitable for on-line use. Applied as a filter, the method is seen to increase the (linear) predictability of the signals studied.