---
title: Predictive Process Monitoring (PPM)
type: concept
tags: [ppm, process-mining, prediction, runtime, hub]
sources: ["[[sources/2018-dumas-fundamentals-of-bpm]]", "[[sources/2026-padella-llm-features-ppm]]", "[[sources/2020-baier-handling-concept-drift-bpm]]"]
created: 2026-04-13
updated: 2026-05-11
---

# Predictive Process Monitoring (PPM)

A branch of [[methods/process-mining-basics|process mining]] concerned with **forecasting the future of running process instances** from partial event-log traces.

Where traditional process mining is **retrospective** (discovery, conformance, performance on completed cases), PPM operates on **running cases** — at any point `k` within an in-progress trace, it answers questions such as:

## Prediction targets
- **[[concepts/next-activity-prediction]]** — which activity comes next?
- **[[concepts/remaining-time-prediction]]** — how long until completion?
- **[[concepts/outcome-prediction]]** — will this case end normally / violate an SLA / result in a positive outcome?
- **Full-suffix prediction** — the entire remaining activity sequence.
- **Event attribute prediction** — next resource, next data value.

## Typical pipeline
1. **Event log** (XES / OCEL) as input.
2. **Trace prefix extraction** — for each completed case, generate all prefixes of length `k ≥ 1` to use as training instances.
3. **[[concepts/trace-encoding]]** — convert variable-length prefixes into fixed-size feature vectors (or tensors), or — for [[concepts/llm-based-ppm|LLM-based PPM]] — into human-readable strings (ρ_seq).
4. **Model training** — classical ML (random forest, XGBoost), RNN/[[concepts/lstm-ppm|LSTM]], [[concepts/transformer-ppm|Transformer]], GNN, or — increasingly — pre-trained LLMs invoked via in-context learning.
5. **Online inference** — apply the model to live running cases.

## Design dimensions
- **Target** — next activity / remaining time / outcome / activity occurrence.
- **Encoding** — aggregate, index-based, one-hot, learned embeddings, LLM string (ρ_seq).
- **Data perspective** — control-flow-only vs data-aware (attributes, resources, timestamps).
- **Inter-case vs intra-case** — does the model consider contemporaneous cases?
- **Model family** — classical ML, sequence models (RNN/LSTM/GRU/Transformer), graph models (GNN), [[concepts/llm-based-ppm|pre-trained LLMs]].
- **Training-data regime** — full event log (typical for classical/deep methods) vs small-scale / 100-trace regime ([[sources/2026-padella-llm-features-ppm|Padella, de Leoni & Dumas 2026]] show LLM-PPM excels at the latter).

## Why it matters
- **Operational** — drive interventions (alert, reroute, reallocate).
- **APM-relevant** — feeds the "Recommend" enactment role of [[concepts/conversational-actionability]]; mentioned in [[sources/2026-calvanese-agentic-bpm-manifesto]] as a prototypical capability an APM agent invokes for time-to-completion reasoning.
- **Feeds PrPM** — predictions only become actionable when combined with interventions; the runtime recommender built on top is [[concepts/prescriptive-process-monitoring|prescriptive process monitoring]] (see [[sources/2022-kubrak-prescriptive-ppm-slr|Kubrak et al. 2022]]).

## Landscape
The `raw/Predictive process monitoring/` corpus spans a decade of PPM research — see [[syntheses/ppm-landscape]] for a cross-cutting synthesis and [[overview]] for domain positioning. For the authored research line of the wiki owner, see [[syntheses/riess-research-arc]] and [[entities/mike-riess]].

## Evaluation rigour
Beyond point accuracy, PPM models should be evaluated on **earliness**, **temporal consistency** ([[sources/2023-riess-temporal-loss-remaining-cycle-time]]), and their robustness across *data-generating-process* characteristics — the latter motivating controlled synthetic benchmarks via [[concepts/business-process-simulation|parametric simulation]] ([[sources/2024-riess-synbps-simulation-framework|SynBPS]]). Deployed models are subject to [[concepts/concept-drift|concept drift]]: [[sources/2020-baier-handling-concept-drift-bpm|Baier, Reimold & Kühl 2020]] document an industrial P2P case where ignoring drift halves accuracy over a 70 k-transaction window, recoverable only by Page-Hinkley detection + incremental retraining on the most recent pre-alarm batch. Survey: [[sources/2021-sato-concept-drift-pm-survey|Sato et al. 2021]] (45-paper SLR; flags lack of common evaluation protocol). Adaptation via metaheuristics: [[sources/2022-riess-metaheuristics-concept-drift-survey]].

For LLM-PPM, an additional concern is **benchmark contamination** — public event logs may be in LLM pre-training data. The [[concepts/semantic-hashing-probe|semantic-hashing probe]] from [[sources/2026-padella-llm-features-ppm|Padella et al. 2026]] is a concrete protocol to test for this.

## Philosophy-of-science framing
A PPM prediction is a probabilistic causal claim and inherits the commitments of its chosen interpretation of [[concepts/probabilistic-causation]]. A tendency/disposition framing ([[sources/2023-anjum-rocca-phi403-lecture-18-risky-predictions]]) fits the case-level nature of PPM better than a frequentist reading; applying aggregate benchmark metrics directly to an individual running case is the **[[concepts/rct-limitations|ecological fallacy]]**.

## Related
[[methods/process-mining-basics]] · [[concepts/self-modification]] · [[concepts/conversational-actionability]] · [[concepts/prescriptive-process-monitoring]] · [[concepts/intervention-policy]] · [[concepts/business-process-simulation]] · [[concepts/concept-drift]] · [[concepts/llm-based-ppm]] · [[concepts/beta-learner-distillation]] · [[concepts/semantic-hashing-probe]] · [[concepts/probabilistic-causation]] · [[concepts/rct-limitations]]