---
title: "Business Process Simulation"
type: concept
tags: [bpm, simulation, synthetic-data, evaluation, ppm]
sources:
  - "[[sources/2024-riess-synbps-simulation-framework]]"
  - "[[sources/2018-dumas-fundamentals-of-bpm]]"
created: 2026-04-20
updated: 2026-04-20
---

# Business Process Simulation

Business Process Simulation (BPS) is the practice of generating synthetic event-log data from a specified or learned process model, for *what-if* analysis, design-time comparison, and (more recently) methodology-level evaluation in [[concepts/predictive-process-monitoring|predictive process monitoring]].

## Two design philosophies

### Calibrated / ecological-validity BPS
Tools such as **Simod** (Camargo et al.), the **BPSim/WfMC**-standard work (Fracca et al. bpsimpy), **SimPT** and **PMSD** (Pourbafrani et al.), Grüger et al.'s SAMPLe, and classical DES tools (Arena, CPN Tools) aim to reproduce the behaviour of a *specific* observed process. Control flow is either discovered from event logs or hand-built from domain knowledge; activity-duration and arrival distributions are calibrated to data. The target deliverable is realistic simulated logs for answering what-if questions about *that* process (staffing changes, reroutings, SLA impact).

### Parametric / internal-validity BPS
[[sources/2024-riess-synbps-simulation-framework|SynBPS]] (Riess 2024) takes the opposite stance: the researcher is not trying to reproduce a specific process but to generate controlled synthetic event logs where data-generating-process factors can be independently manipulated. Control flow is a Markov chain of configurable order; entropy of the transition matrix, state-space size, activity-duration distributions (exponential / hypoexponential), arrival rate (Poisson) and stability (drift) are all user-specified. The target deliverable is factorial designs for PPM-method robustness assessment — analogous to controlled benchmarks in classical ML regularisation research (Lasso, LARS, elastic net).

## Why this matters for PPM evaluation
PPM papers typically evaluate methods on 3–9 public event logs (BPIC, Sepsis, Helpdesk, Hospital Billing, etc.). Each event log is *one* realisation of *one* process from *one* organisation. Comparative claims across methods are valid, but claims about which *process characteristics* drive performance cannot be causally tested without control variables ([[sources/2024-riess-synbps-simulation-framework]]). SynBPS provides the infrastructure for this second type of question.

Separately, simulation can be used to evaluate *prescriptive* process monitoring interventions (e.g. queue prioritisation) where counterfactual deployment on real systems is infeasible — see Paper III of [[sources/2023-riess-phd-thesis-ppm]] (with Scholderer), which uses agent-based calibrated simulation to test predicted-loyalty queue prioritisation.

## Related
- Method page: [[methods/process-simulation]] — discrete-event simulation fundamentals from [[sources/2018-dumas-fundamentals-of-bpm]] §7.3.
- Concepts: [[concepts/predictive-process-monitoring]], [[concepts/concept-drift]], [[concepts/behavioral-variability]].
- Author: [[entities/mike-riess]].