Prescriptive Analytics
Descriptive statistics summarise what happened. Predictive models (the classifiers, regressors, clusterers and anomaly detectors covered elsewhere in this guide) score what is likely to happen to a given record. Prescriptive analytics asks a different question: given a model or a network of relationships, what should change, and what follows if it does — across a whole population of hypothetical cases, not one real record at a time.
Scope: explanatory, not a recall path
Monte Carlo simulation reuses a model that already has a real AdvancedAnalytics recall path (see
MonteCarloSimulation.md’s own mapping notes), so it is prescriptive in purpose but runs on production infrastructure
end to end.
The Bayesian network material does not have a native tool-layer path. AdvancedAnalytics has one shipped
Bayesian-classifier-shaped model, Tree-Augmented Naive Bayes (AdvancedAnalyticsTreeAugmentedNaiveBayesClassifierTrainingService,
see BayesianAnalysisWithBnlearn.md); it has no general Bayesian network as an
[AdvancedAnalyticsTool] — arbitrary structure, hand-set or learned CPTs, multi-node querying and intervention —
because general Bayesian-network structure search is out of scope
(docs/AnalyticalMethodology/Reference/ModelCatalogue.md). BayesianNetworks/ gives explanatory power rather
than a native tool-layer path: the conceptual and mathematical grounding for reasoning about causal structure,
conditional independence and intervention, worked entirely in bnlearn, applied on top of the tooling that does
exist (TAN, and the descriptive/inferential tool groups).
The R script execution job pattern (Jube.R, the sandboxed R server, plus the RScriptTools submit/status/result
agent tools — see R Server) is the real way to run this: an agent can submit an arbitrary
bnlearn script (structure learning, hand-set or learned CPTs, cpquery in either direction, mutilated()
intervention) as a job, poll it, and read back the result — exactly the operations BayesianNetworks/ catalogues as
bnlearn-only with no native tool-layer path. Governance, KPIs and Model Validation
carries the full compliance guidance naming Bayesian networks as “the primary probabilistic framework for both
anomaly detection and prescriptive analytics,” specifically as the fallback methodology for weak or absent labels:
unsupervised, probabilistic detection via conditional probabilities over feature states, refined by supervised
Bayesian models once reliable labels exist. That is a batch, exploratory, expert-reviewed use — not a per-transaction
recall requirement — and the R script job pattern is a genuine way to run it. See
Explanatory Role Inside AdvancedAnalytics for how this
sits alongside TAN. What remains genuinely absent, and is not needed for this fallback-methodology use case, is a
live, per-transaction recall path — that stays a bnlearn-only, job-pattern operation, not a real-time one.