QBUS2820 Chap.1 Predictive Modelling Foundations
Predictive Modelling Foundations
Predictive modelling converts a business question into a response, predictors, a learning task, a loss and a validated decision. This chapter distinguishes supervised from unsupervised learning, regression from classification, and prediction from inference.
It also explains the data-generating process, expected squared error, the regression function, parametric and nonparametric methods, interpretability, underfitting and overfitting. The central habit is to define what must be known at prediction time before choosing an algorithm. A careful framing also separates the population of future cases from the sample already observed.
The analyst should say whether a prediction concerns the same stores, new stores, later dates or a changed market, because each target requires a different validation unit. Response quality matters as much as algorithm choice: a delayed, inconsistently recorded or proxy outcome can teach the model the wrong objective.
Predictor timing must be audited one field at a time, including summaries that may quietly incorporate future transactions. The loss function then converts business consequences into a comparison rule; squared error is convenient, but asymmetric stockout and overstock costs may justify a different choice. Deployment completes the loop by introducing monitoring, recalibration and retraining decisions.
A model can remain mathematically unchanged while customer behaviour, data collection or operating policy shifts around it, so performance evidence has an expiry condition rather than permanent validity.
What this chapter covers
- 01
Business questions and prediction targets
- 02
Supervised and unsupervised learning
- 03
Regression and classification responses
- 04
Prediction compared with inference
- 05
Loss and expected prediction error
- 06
Regression functions and additive noise
- 07
Underfitting and overfitting
- 08
Parametric and nonparametric methods
Worked example · free
Frame a demand prediction task
- +1The marks used in this drill are a study allocation, not an official University marking scheme. Define the response as next-day units sold for each store and product.
- +1Use predictors available before the order, such as recent sales, price, promotion, weekday and forecast weather.
- +1Classify the task as regression because the response is numerical.
- +1Choose a loss that reflects stockout and overstock consequences, with squared error as a baseline.
- +1Exclude actual next-day traffic because it is observed after the replenishment decision and would leak future information.
Key terms
- Supervised Learning
- Learning a response from labelled training cases containing both predictors and observed outcomes.
- Unsupervised Learning
- Finding structure among inputs without an observed response variable.
- Prediction Loss
- A penalty measuring disagreement between an observed outcome and its prediction.
- Regression Function
- The conditional mean of a numerical response and the optimal squared-error target.
- Overfitting
- Learning accidental training details that do not generalise to future cases.
- Model Flexibility
- The capacity of a method to represent complex patterns in the predictor-response relationship.
- Decision Clock
- The point when a prediction must be issued, determining which predictor information is legitimately available to the model.
Predictive Modelling Foundations FAQ
How do I identify the learning type?
Check whether an observed response is available. No response suggests unsupervised learning. A numerical response gives regression, while a categorical response gives classification. Then state whether the goal is future prediction or explanation of predictor effects.
Why is training error not enough?
The model was fitted to reduce error on its training cases, so that error is optimistic. Generalisation requires genuinely unseen cases and an evaluation design that keeps them outside fitting, preprocessing, feature choice and tuning.
What is the difference between prediction and inference?
Prediction prioritises accurate future outputs. Inference prioritises the direction, size and uncertainty of predictor effects. The same model can be used for both, but the evidence and interpretation required are different.
How should the prediction population be described?
State whether future cases are later observations from existing groups, entirely new groups or cases under changed conditions. The split design should reproduce that target as closely as the available data permit.
What makes a response variable unsuitable?
A response can be delayed, inconsistently measured or only loosely related to the decision objective. Document its timing and meaning before fitting, because an algorithm cannot repair a target that represents the wrong outcome.
Exam move
Build a one-page task-framing template with six lines: decision, response, predictors, learning type, loss and validation unit. Apply it to every lecture example before reading the method name. Rehearse the distinction between regression and classification using new business questions, then explain why each candidate predictor is available at decision time.
Draw the learning pipeline from business understanding through data, modelling, prediction and deployment, including the feedback arrow that creates new data. For mathematical practice, write squared loss and explain why the conditional mean minimises expected squared error. Compare a rigid line with a flexible local method using bias, variance and interpretability language.
Finish each study session by inventing one leakage example and one reason a high training fit could fail after deployment.
Working through Predictive Modelling Foundations in QBUS2820? Sia is AskSia’s AI Statistics tutor — ask any QBUS2820 Predictive Modelling Foundations question and get a clear, step-by-step explanation grounded in how QBUS2820 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
