Remaining cycle time prediction: Temporal loss functions and prediction consistency

The usefulness of remaining cycle time models for predictive and prescriptive process monitoring depends not only on the overall accuracy predictions but also their earliness: should be as accurate possible, early possible. To give this criterion more weight in model fitting, paper evaluates three L1 loss functions with temporal decay. All have property increasing residuals from stages a relative to later do so different degrees. are used LSTM networks training throughout four business processes based publicly available event log data sets. Compared trained unweighted loss, suggested modifications yield small significant improvements earliness out-of-sample data. Neither nor led strictly monotonically decreasing time.