Forecasting approaches that exploit analogies require the grouping of analogous time series as the first modeling step; however, there has been limited research regarding the suitability of different segmentation approaches. We argue that an appropriate analytical segmentation stage should integrate and trade off different available information sources. In particular, it should consider the actual time series patterns, in addition to the variables that characterize the drivers behind the patterns observed. The simultaneous consideration of both information sources, without prior assumptions regarding the relative importance of each, leads to a multicriteria formulation of the segmentation stage. Here, we demonstrate the impact of such an adjustment to segmentation on the final forecasting accuracy of the cross-sectional multi-state Kalman filter. In particular, we study the relative merits of single and multicriteria segmentation stages for a simulated data set with a range of noise levels. We find that a multicriteria approach consistently achieves a more reliable recovery of the original clusters, and this feeds forward to an improved forecasting accuracy across short forecasting horizons. We then use a US data set on income tax liabilities to verify that this result generalizes to a real-world setting.