<p>This paper develops a new prediction framework for the extreme episodes of cross-border capital flows in a mixed-frequency binary choice panel setting. The four well-established episodes are redefined and are modelled not only individually, as routinely assessed, but also jointly, in pairs. The model predicts quarterly event probabilities using daily and monthly macro-financial predictors. The time series of out-of-sample predictions is assessed using formal forecast verification techniques in full information and pseudo-real-time settings. The panel model predicts with significant forecast skill with respect to a random classifier and an intercept-only benchmark – more prominently so for three of the six episodes. The pseudo-real-time analysis also shows that the predictions can beat the benchmarks and thus generate meaningful early warnings a quarter in advance of the target quarter. Accuracy remains stable as time elapses and more information becomes available.</p>