Remaining Time Prediction

A PPM task: given a prefix of an event trace, predict the remaining execution time until the case completes.

Formulation

Input: prefix ⟨e₁, …, eₖ⟩ with timestamps.
Output: scalar t_remaining (regression).
Training data: for each prefix extracted from a completed case, label = (case completion timestamp − prefix-end timestamp).

Canonical approaches

Transition system annotation — van der Aalst et al.; annotate states of a transition system with mean remaining time.
Regression trees / ensembles — feature-engineered. Includes the CatBoost regressor used as a benchmark in Padella et al. 2026.
LSTM regression head — often in multi-task setup with next-activity prediction.
Transformer regressors — recent state of the art (e.g. PGTNet — Process Graph Transformer Network — used as benchmark in Padella et al. 2026).
LLM-based prediction — emerging family, particularly competitive in data-scarce settings. Padella, de Leoni & Dumas 2026 document Gemini 2.5 Flash Thinking trained on 100 traces matching or surpassing CatBoost / PGTNet on the full BPI12 / Bac / Hospital event logs for Total Time MAE.

Evaluation — three axes

Accuracy — MAE / MAPE / RMSE in time units (hours, days).
Earliness — accuracy at early prefixes, when information is minimal and prediction is hardest (Verenich et al. 2019; Riess 2023). Operationally critical for queue prioritisation and dynamic resource planning.
Temporal Consistency (TC) — the degree to which consecutive predictions follow the natural monotonic decrease of true remaining time (Riess 2023). Proposed axis; none of the standard losses evaluated (unweighted MAE, exponential / power / moderate temporal decay) produce strictly monotonically decreasing predictions.
Error curves over prefix length — accuracy typically improves as more of the case has been observed.

Loss-function design

Standard: unweighted L1 (MAE) — robust to time-between-event outliers and typically outperforms MSE (Verenich et al. 2019). Temporally weighted L1 (Riess 2023): three variants — exponential decay (MAEEtD), power decay (MAEPtD), moderate decay (MAEMtD) — upweight early-prefix residuals. Exponential decay gave statistically significant earliness improvements on 2 of 4 public event logs; should be treated as a tunable hyper-parameter, with a potential trade-off against temporal consistency.

Challenges

Long-tail case durations — outlier long cases skew error metrics.
Case heterogeneity — different variants have very different remaining-time distributions.
Non-stationarity — process behaviour drifts; models need retraining (concepts/concept-drift).
Data-awareness matters — incorporating case data (attributes, resources) outperforms control-flow-only (sources/2017-navarin-lstm-data-aware-remaining-time and others).
Small-benchmark external validity — most published methods evaluated on 3–9 public event logs; controlled synthetic evaluation via parametric simulation (SynBPS) recommended for data-generating-process-factor analysis.

concepts/predictive-process-monitoring · concepts/next-activity-prediction · concepts/outcome-prediction · concepts/trace-encoding · concepts/concept-drift · concepts/business-process-simulation · concepts/llm-based-ppm