Glacier-dammed lakes can yield subglacial outburst floods (jo¨kulhlaups) repeatedly.Predicting flood timing is crucial for hazard mitigation, but incomplete understanding of flood-initiationphysics makes this challenging. Here we examine the predictability of the timing of jo¨kulhlaups fromMerzbacher Lake, Kyrgyzstan, using five flood-date prediction models of varying complexity. Thesimplest model, which offers a benchmark against which the other models are compared, assumes thatfloods occur on the same date each year. The other four models predict flood dates using a floodinitiationthreshold approach and incorporate weather forcing (approximated by the output of twoclimate reanalyses) behind the meltwater input to the lake; the most complex of these models accountsfor a moving subglacial water divide beneath the glacier that dams the lake. Each model is optimizedagainst recorded flood dates to maximize its prediction ability. In terms of their flood prediction ability,our two best models are those that assume a variable outburst threshold governed by the rate ofmeltwater input to the lake and the rate of lake-level rise. They excel over the simplest and mostcomplex models and correctly predict flood dates to within �20 days 57.4% of the time. We alsoquantify the impact of weather uncertainty on prediction success. Our findings can inform practicalflood-forecasting schemes and future investigations of flood-initiation physics.