This paper evaluates a forecasting procedure, which was based on the data for the period 1926–1993, in and around Japan. The period of the experiment is from 1994 to April 2011. According to the procedure, the probability that the first earthquake will be a foreshock varies in a range between 1 and 10-odd per cent, depending on its location. This location-dependent forecasting performed significantly better than the unconditional foreshock probability (3.7 per cent average) throughout the Japan region. Furthermore, when multiple earthquakes were observed as a cluster, the foreshock forecast probabilities varied in a much wider range, depending on the space–time distances and magnitude increments, between the earthquakes. Such forecasts performed better than the average foreshock probability, and the forecast probabilities were basically consistent with the outcomes. The improvements of the forecasts were objectively evaluated by using the log likelihood score. It is also shown that the forecasting procedure was robust enough and can be adopted for automatically determined earthquakes in real-time.