Outcome Prediction

A PPM task: given a prefix of an event trace, predict a case-level outcome — typically binary or small-multi-class.

Typical targets

SLA compliance — will this case complete within the deadline?
Business outcome — will this loan be repaid / approved / defaulted?
Positive / negative closure — will this complaint be resolved favourably?
Constraint violation — will a DECLARE rule be violated?
Activity occurrence — will a specific (typically high-cost or high-rework) activity occur within the running case? Used as a classification KPI in Padella, de Leoni & Dumas 2026 (W_Nabellen incomplete dossiers in BPI12, Service Closure with BO Responsibility in Bac, LABORATORIO in Hospital).

Formulation

Input: prefix of a running case.
Output: categorical outcome y ∈ {0, 1} or y ∈ {c₁, …, cₘ}.
Training data: extract prefixes from completed cases, labelled with the known ground-truth outcome.

Benchmark (Teinemaa et al. 2016+)

A systematic literature review and benchmark by Teinemaa, Dumas, La Rosa, Maggi established the field's evaluation baseline — see sources/2016-teinemaa-outcome-ppm-review.

Evaluation

AUC-ROC / AUC-PR — the standard in imbalanced settings.
Accuracy / F1 — less informative under class imbalance.
Earliness-accuracy trade-off — a prediction earlier in the case is more actionable but harder; often plotted as accuracy vs prefix length.

Philosophical caveats

An outcome prediction is a probability — but which kind? Interpreted as frequentist it gives the population-average rate of the outcome for cases with similar prefixes; applied to an individual case this is the ecological fallacy. Interpreted as credence it expresses the model's epistemic uncertainty (see concepts/aleatoric-vs-epistemic-uncertainty). Interpreted as propensity it claims an intrinsic tendency of the case itself (concepts/probabilistic-causation). The choice matters whenever a prediction drives an intervention — see sources/2023-anjum-rocca-phi403-lecture-19-what-rcts-do-not-show and sources/2023-anjum-rocca-phi403-lecture-18-risky-predictions.

LLM-based outcome prediction

Padella et al. 2026 benchmark Gemini 2.5 Flash Thinking against CatBoost Classifier on the Activity Occurrence variant with 100 training traces. LLM (non-hashed) achieves F1 0.77 / 0.98 / 0.90 on BPI12 / Bac / Hospital — matching or exceeding CatBoost trained on the full event log. Performance degrades slightly under semantic hashing (-2 % to -7 %), confirming partial reliance on embodied prior knowledge for classification too.

concepts/predictive-process-monitoring · concepts/next-activity-prediction · concepts/remaining-time-prediction · concepts/trace-encoding · concepts/probabilistic-causation · concepts/llm-based-ppm