Predictive Models Overview
Key Takeaways
- Predictive analytics estimates future or unknown outcomes to reduce decision uncertainty—not to impress with algorithms.
- CBDA competency 3.4 awareness level covers regression, logistic regression, trees/random forests, and neural nets as model families with different strengths.
- Supervised learning uses labeled outcomes; unsupervised learning finds structure without a predefined target label.
- Train/test (or holdout) awareness and overfitting concepts protect teams from trusting models that only memorize history.
- Mathematical models (competency 3.3) are simplified formal representations of relationships; CBDA expects conceptual literacy, not derivation of proofs.
Predictive Models Overview
Quick Answer: Predictive analytics builds models that estimate future or unknown outcomes so leaders can plan, prioritize, and intervene with less uncertainty. On CBDA, Domain 3 (Analyze Data, about 16% of exam weight) expects you to select model families at awareness level, respect train/test discipline, recognize overfitting, and treat models as tools for the research question—not ends in themselves.
Earlier Domain 3 work (analysis foundations, descriptive statistics, exploratory analysis, and diagnostic thinking) answers “what happened?” and “why might it have happened?” Predictive work answers “what is likely to happen?” or “what is the unknown label for this case?” Prescriptive work (often later or separate) asks what action to take under constraints. CBDA scenario stems frequently jump straight to “build a model”; strong candidates first confirm that the decision truly needs a forward-looking estimate and that descriptive baselines and metric definitions already exist.
Purpose of Predictive Analytics for Business Questions
A predictive model is useful only when it reduces uncertainty for a decision. Classic business uses include:
| Business need | Predictive framing |
|---|---|
| Who will churn next quarter? | Probability or risk score of churn |
| Which claims are likely fraudulent? | Classification of high-risk claims |
| What will demand be next month? | Continuous forecast of units or revenue |
| Who is likely to convert if contacted? | Propensity / response score |
| How long until an asset fails? | Survival or risk-of-failure estimate |
Notice the pattern: each row ties a population, an outcome, a time horizon or decision window, and a use (target outreach, capacity, audit sampling, inventory). CBDA items punish technique-first thinking (“we should use deep learning”) when the research question is still descriptive or diagnostic.
Predictive work sits in the middle of the CBDA lifecycle:
- Domain 1 framed the decision and research question.
- Domain 2 sourced and prepared data that can represent the outcome and candidate drivers.
- Domain 3 estimates relationships and scores cases (this chapter).
- Domains 4–5 interpret, communicate, and influence action—where model limits matter as much as model accuracy.
If the business only needs a clear baseline of last year’s conversion rate by segment, a predictive model is optional—and may be a distraction. If the business must allocate limited retention offers before customers leave, a well-scoped predictive score can be decision-critical.
Model Families at CBDA Awareness (Competency 3.4)
CBDA does not require you to code gradient descent or derive likelihood functions. It does expect you to recognize major families, what kind of outcome they typically support, and rough tradeoffs among interpretability, complexity, and data appetite.
Linear / multiple regression
Regression (ordinary least squares and extensions) estimates a continuous outcome (revenue, days to close, cost, blood pressure) as a weighted combination of predictors. Coefficients describe average change in the outcome associated with a one-unit change in a predictor (holding others constant—conceptually). Strengths: transparent, fast, good baselines. Limits: linear assumptions, sensitivity to outliers, struggle with complex interactions unless engineered.
Logistic regression
Logistic regression estimates binary (or sometimes multi-class) outcomes—churn yes/no, default yes/no, convert yes/no—using a logistic link so predictions stay between 0 and 1 as probabilities (or odds). Strengths: familiar odds framing for business and risk teams, relatively interpretable coefficients. Limits: linear relationship on the log-odds scale; interactions and nonlinearities need deliberate design.
Decision trees and random forests
Decision trees partition the feature space with if-then splits that map cleanly to business rules (“if tenure < 6 months and tickets > 3, high risk”). Single trees are highly interpretable but can be unstable and overfit. Random forests (and similar ensembles) average many trees to improve predictive performance at the cost of a denser, harder-to-explain model—though feature-importance summaries remain available.
Neural networks
Neural nets (including deep learning) learn layered nonlinear representations. They can excel with large data and complex patterns (images, free text, highly interactive tabular signals) but are often opaque, data-hungry, and expensive to govern. On CBDA, expect awareness that neural nets are a family option—not a default upgrade—and that business explainability and data readiness may rule them out even when accuracy might improve.
| Family | Typical outcome | Interpretability | Data / complexity appetite |
|---|---|---|---|
| Linear regression | Continuous | High | Moderate |
| Logistic regression | Binary / categorical | High–medium | Moderate |
| Decision tree | Continuous or class | High (single tree) | Moderate |
| Random forest | Continuous or class | Medium–low | Higher |
| Neural net | Continuous or class | Low (often) | High |
Exam mindset: match family to outcome type, stakeholder need for explanation, and available data—not to fashion.
Supervised vs Unsupervised (Conceptual)
Supervised learning trains on examples where the target outcome is known (labeled history): past churn flags, past claim fraud labels, historical demand. The model learns a mapping from features to that target, then scores new cases.
Unsupervised learning has no predefined label. Clustering, association rules, and dimensionality reduction find structure—segments, co-purchase patterns, latent factors. These techniques support discovery, segmentation, and feature construction; they do not, by themselves, predict a business KPI unless you later attach labels or rules.
CBDA trap: calling a clustering project “predictive analytics” when the research question was “who will default?” Without a default label or a validated rule mapping clusters to risk, you have structure, not a prediction of the decision outcome. Conversely, unsupervised methods can feed supervised models (cluster membership as a feature) after validation.
Semi-supervised and reinforcement learning exist in the wider field; for CBDA, master the supervised/unsupervised distinction and when each advances the research question.
Train/Test Awareness and Overfitting
A model that perfectly recapitulates the data it was built on can still fail tomorrow. Overfitting means the model captured noise and idiosyncrasies of the training sample rather than generalizable signal. Classic symptoms: excellent training metrics, weak performance on new data; overly complex trees; too many features for the sample size.
Train/test (holdout) discipline is the conceptual fix every CBDA candidate should own:
- Training set — used to fit parameters / learn splits.
- Test (holdout) set — held back to estimate performance on unseen data; not used for fitting.
- Validation / cross-validation (awareness) — used during model selection and tuning so the final test remains honest.
Related ideas at conceptual level:
- Underfitting — model too simple; high error on train and test.
- Data leakage — information from the future or from the label sneaks into features (e.g., “cancellation reason” predicting churn). Metrics look magical and real-world scores collapse.
- Temporal splits — for time-based decisions, random row splits can leak future information; training on past periods and testing on later periods often matches the business better.
On scenario items, if a team reports 99% accuracy with no holdout and dozens of engineered features on a small sample, your best next step is often validate on held-out data or simplify and re-check, not celebrate.
Mathematical Models at Basic Knowledge Level (Competency 3.3)
A mathematical model is a formal representation of relationships among variables—equations, probability distributions, optimization formulations, or algorithmic rules that map inputs to outputs. “Mathematical” does not mean “incomprehensible.” A linear equation y = β₀ + β₁x₁ + … is a mathematical model. So is a logistic probability formula. So is a decision tree’s partition of the feature space.
CBDA basic knowledge means you can:
- Recognize that models simplify reality (assumptions always exist).
- Distinguish descriptive summaries (means, rates) from models that generalize to new cases.
- Accept that model form should fit the measurement scale of the outcome and the decision use.
- Communicate that parameters (coefficients, split thresholds, network weights) are estimated from data and therefore uncertain.
You are not expected to prove Gauss–Markov theorems. You are expected to know that choosing an inappropriate form (e.g., unconstrained linear probability for a rare binary event without care, or a black-box model when regulators need reason codes) is an analysis design failure, not a software failure.
How Predictive Overview Connects to Validation
This section sets vocabulary. The next sections deepen regression and classification choices, interpretation of outputs, and—critically—whether results still answer the business question. Predictive success on CBDA is never “highest leaderboard score”; it is decision-useful, validated estimation with known limits.
Exam traps to avoid
- Jumping to neural nets when logistic regression meets the decision and explainability need.
- Confusing unsupervised clustering with prediction of a labeled outcome.
- Trusting training metrics without holdout evaluation.
- Building models before the outcome definition and population are locked (Domain 1–2 debt).
- Treating the model as the deliverable instead of an input to recommendation and decision support.
A product team asks, “Which customers are most likely to cancel in the next 60 days so we can prioritize retention offers?” Which analytics category and learning setup best match this need?
An analyst reports near-perfect accuracy on the same customer file used to train a complex random forest and recommends immediate production use. What concept should a CBDA practitioner raise first?
At CBDA awareness level, when is a neural network often a poorer first choice than logistic regression for a binary business outcome?