Why this page exists
Pega’s Customer Decision Hub picks the next best action for millions of customers at banks, insurers and telcos. Most people who hold the certification can operate the screens. Far fewer can explain what the platform is doing — and that difference matters, because every tuning decision (levers, exploration, contact policies) is a decision about the machine underneath.
I’m a Pega Senior System Architect and Certified Decisioning Consultant with a mathematics degree, so this project is my lane: decisioning explained from the inside, in language both a Pega practice lead and a data scientist will recognize as their own.
The 60-second version of CDH
When a customer opens the banking app, a real-time container fires. CDH assembles every action the bank could take — cross-sell offers, retention nudges, service messages — then filters them through engagement policies (eligibility, applicability, suitability), scores what remains with adaptive models, and ranks the survivors with one formula:
Priority = Propensity × Context weight × Value × Business levers
The top-ranked action renders before the page finishes loading. Every customer response — clicked, ignored, converted — flows back through the Interaction History into the models. The system learns continuously; nobody retrains anything by hand.
What an adaptive model actually is
This is where precision matters, because it’s the part most often hand-waved:
Pega’s Adaptive Decision Manager (ADM) is an online naive Bayes classifier. Each action gets its own model. Incoming predictors are binned — numeric fields into frequency-based intervals, symbolic fields into grouped categories — and each bin accumulates positive and negative response counts. A customer’s propensity is assembled from the log-odds evidence of the bins they fall into, smoothed so that thin evidence doesn’t produce wild scores. Model quality is reported as AUC on Pega’s 50–100 scale — the same discrimination measure any data scientist uses, rescaled.
Cold-start is handled with controlled exploration: new models get deliberate uncertainty in their served propensities, which shrinks as response evidence accumulates — so new actions get a fair chance to prove themselves without swamping proven performers. In ML vocabulary: an explore/exploit strategy over a fleet of online classifiers.
What ADM is not: a deep-learning black box, a batch model on a retrain schedule, or (a claim I see often) “just a bandit.” The bandit-style exploration sits on top of an interpretable, inspectable naive Bayes core — and that interpretability is exactly why regulated industries accept it.
The Rosetta Stone: Pega ↔ machine learning
The table I wish existed when I moved between these two worlds:
| Pega concept | ML equivalent | What it really means |
|---|---|---|
| Adaptive model (ADM) | Online naive Bayes classifier, incrementally updated | One self-learning propensity model per action, no batch retraining |
| Predictor binning | Discretization / WOE-style encoding | Numeric fields become evidence bins with response counts |
| Model performance (50–100) | AUC (0.5–1.0), rescaled | Discrimination: how well the model separates responders |
| Propensity smoothing on new models | Exploration (explore/exploit) | Uncertainty injected while evidence is thin — cold-start handling |
| Arbitration (P×C×V×L) | Scoring policy / expected-value ranking | Model output is one input; business context and levers complete the ranking |
| Engagement policies | Hard constraints / filtering before scoring | Eligibility is never left to the model |
| Contact policies & suppression | Constrained optimization | Protecting customers from over-contact beats squeezing propensity |
| Interaction History | Label store / feedback loop | The training data pipeline — and where leakage hides |
| IH summaries in policies | Feature aggregation over event logs | E.g. “suppress if ignored 3× in 14 days” |
| Champion / challenger | A/B testing of policies | Controlled comparison in production |
| Prediction Studio monitoring | Model observability (drift, calibration, volume) | Watching AUC, success rates and predictor health over time |
If you can read both columns fluently, you can run a decisioning program: tune it, govern it, debug it, and explain it to a regulator. That fluency is what this portfolio is for.
Governance is half the job
The part of decisioning that rarely makes conference slides: a propensity model that is well-calibrated matters more than one with spectacular AUC, because propensity is multiplied with value and levers in arbitration — an overconfident model silently distorts every ranking behind it. Likewise contact policies, suppression rules and lever governance (“when is a business override a smell?”) decide whether a decisioning program survives contact with a bank’s compliance department. My platform-operations background (deployment audits, CVE handling, change control at an EU institution) is unusually good preparation for this half.
Health check: Pega’s public CDH sample estate
To make this concrete, I ran pdstools — the open-source analysis toolkit built by Pega’s own data-science team — against the public CDH Sample ADM datamart export: 68 adaptive models across SMS, Email and Web, with 195,565 captured responses. This is the same analysis a decisioning consultant runs on a real client estate. All charts are interactive.
The estate at a glance
The signature CDH health chart: every bubble is an adaptive model, positioned by performance (AUC) vs success rate, sized by response volume. A healthy estate shows bubbles drifting up-right as evidence accumulates. Here the story is more interesting — a visible column of models stuck at the left wall.
Finding 1 — one in five models has learned nothing, despite having data
14 of 68 models (21%) sit at AUC 0.50 — coin-flip discrimination — and crucially, most are not cold-start models: MasterCardGold (Email) has 875 responses, MoneyMarketSavingsAccount (Web) has 1,206, yet neither separates responders from non-responders at all. Volume without learning means the predictors carry no signal for these actions. Recommendation: predictor review for this cluster — add IH summary predictors and action-specific context, or fold these niche actions into a group-level model.
Finding 2 — Email is the weak channel, on both axes
Channel comparison from the same snapshot: Email models average AUC 57.3 (response-weighted) vs ≈62.8 for SMS and Web — and Email’s success rate is 2.4% against 8.7% (SMS) and 9.9% (Web). When a channel underperforms on discrimination and outcome simultaneously, the first suspects are treatment quality and response-capture configuration (is an “open” being counted where other channels count a “click”?) — an operational question before a modeling one. That instinct, checking the outcome pipeline before blaming the model, is the difference between decisioning analysis and generic ML analysis.
Finding 3 — a third of the predictor library is dead weight
The estate feeds 89 unique predictors into its models, but 34 of them (38%) are active in fewer than 1 in 20 of the models they reach — ADM’s own predictor selection keeps rejecting them. The top of the ranking is dominated by a handful of customer-profile and Interaction History fields. Recommendation: prune the persistently-inactive predictors (cost and clarity), and note what’s missing — IH summary aggregates, which are typically the strongest signal family in mature estates, are underrepresented here.
Method note: performance figures are AUC on the 0.5–1.0 scale as stored in the datamart (Pega’s UI shows the same values rescaled to 50–100). Analysis: pdstools 4.x · ADMDatamart aggregates on the latest snapshot per model · public sample data only.
Walkthrough: inside the platform
What I’d explore next
- Rebuilding the ADM learning loop in open Python (binned online naive Bayes + explicit exploration) against a simulated customer population, exported in ADM datamart schema so pdstools can analyze my simulator with the same reports it runs on real estates.
- A calibration study: reliability curves on adaptive propensities, and what miscalibration does to arbitration rankings downstream.
- Uplift vs propensity: when “who will accept?” is the wrong question and “who accepts because we asked?” is the right one.
Everything on this page reflects public Pega documentation and my own certified training — no client configurations or data.