Time-Based Split

Time-based splitting is a preprocessing principle that becomes essential in time series and event-stream settings. Random splitting can allow future information to leak into the past, producing artificially high performance. For that reason, training data should come from earlier periods, while validation and test data should come from later periods. A time-based split is not merely a technical choice; it is an honest simulation of real-world model usage on data.