Forecasts of eruption are uncertain. The uncertainty is amplified when volcanoes reawaken after several generations in repose, because direct evidence of previous behaviour is rarely available. It fosters scepticism about warnings of volcanic activity and may compromise the success of emergency procedures. The quality of forecasts has improved over the past 50 years, owing mainly to a growing sophistication in statistical analyses of unrest. Physics-based analyses have yet to achieve the same level of maturity. Their application has been delayed by a view that volcanoes are too complex to share patterns of behaviour that can be described in a deterministic manner. This view is being increasingly challenged and an emerging line of inquiry is to understand how forecasts can be further improved by integrating statistical approaches with new constraints on possible outcomes from physics-based criteria. The introduction of deterministic reasoning yields rational explanations of why forecasts are not perfect and, as a result, offers new opportunities for increasing public confidence in warnings of eruption.