BSAN2201 Chap.4 Prediction and Optimisation Methods
Prediction and Optimisation Methods
Predictive methods estimate a number or category, while prescriptive methods select an action. Regression suits numeric targets, classification suits categorical targets, and time-series methods preserve temporal order. Machine learning learns patterns from training data and must be evaluated on unseen observations.
Optimisation then combines an objective, decision variables and constraints to choose the best feasible alternative. Simulation answers a related but different question by exploring how a system behaves under repeated uncertainty.
The chapter keeps forecasts, decisions and business costs separate so that a technically accurate model does not become an infeasible recommendation.
Model evaluation should mirror the operating decision. A useful baseline shows whether complexity improves on a simple rule, while held-out evidence tests performance beyond the training records.
Error measures must reflect business consequences because over-forecasting and under-forecasting rarely cost the same amount. Prescriptive results need the same discipline: every budget, capacity, service or policy limit must appear in the feasible set. Sensitivity analysis then reveals whether the selected action remains stable when forecasts or constraints change.
A robust recommendation may be preferable to a fragile optimum whose advantage disappears after a small input revision.
What this chapter covers
- 01
Regression and numeric targets
- 02
Classification and categories
- 03
Time-series prediction
- 04
Machine-learning evaluation
- 05
Optimisation objectives and constraints
- 06
Simulation under uncertainty
Worked example · free
Forecast and allocate a warehouse shift
- +1Divide 420 orders by 35 orders per picker to get 12 fulfilment pickers.
- +1Add the two constrained returns roles to reach 14 pickers.
- +1Classify the demand estimate as predictive and the allocation as prescriptive.
- +1State that prediction error, breaks and order complexity can make the minimum fragile.
Key terms
- Regression model
- A predictive model whose target is numeric.
- Classification model
- A predictive model that assigns an observation to a category.
- Training data
- Observations used to fit or learn a model's pattern.
- Optimisation
- Selection of the best feasible choice under an objective and constraints.
- Simulation
- Repeated modelling of uncertain system behaviour across possible scenarios.
Prediction and Optimisation Methods FAQ
How do I choose between regression and classification?
Start with the target. Use regression when the desired output is numeric, such as demand or cost, and classification when the output is a category, such as churn or no churn. Similar predictors can feed both, but the outputs and evaluation measures differ.
Why should a model be tested on unseen data?
A model can fit its training observations closely without generalising to new cases. Evaluation on held-out data estimates how it behaves beyond the examples used to build it and reduces the risk of mistaking memorisation for reliable prediction.
What makes an optimisation recommendation feasible?
Feasibility comes from constraints that represent the operating system, such as budgets, capacity, service commitments or route restrictions. If an important limit is missing, the solver may return a mathematically optimal plan that cannot or should not be implemented.
What is the difference between optimisation and simulation?
Optimisation selects the best feasible choice for a stated objective under constraints. Simulation explores how a system behaves across repeated uncertain scenarios and produces a range of outcomes. They can work together: simulate candidate policies under variable demand, then choose one using an explicit performance and risk criterion.
Assessment move
For every model, write the target, its type, the cost of each error and the action that follows. For optimisation, list the objective, decision variables and constraints before doing arithmetic. Create one counterexample where a missing constraint produces an impossible recommendation. This makes the distinction between predictive accuracy and operational usefulness concrete.
Compare the proposed model with a simple baseline and state why any improvement matters to the business. For time-ordered data, draw the training and evaluation periods so future information cannot leak backwards. Change one forecast or capacity input and observe whether the recommended plan changes.
Finally, invent one market or policy shift that could weaken the learned pattern, and name the monitoring signal that would reveal it. This connects initial model choice to responsible use over time. Build a loss table showing false-positive and false-negative consequences for two different thresholds. Explain why the preferred threshold follows from the decision cost rather than accuracy alone.
For a prescriptive example, solve the central forecast, then change capacity and demand separately to find when the recommended plan changes. Record the feasible set before ranking alternatives. End by assigning a review owner, realised outcome and recalibration trigger. This turns a one-time model selection exercise into an operational learning loop and makes fragility visible before implementation.
Working through Prediction and Optimisation Methods in BSAN2201? Sia is AskSia’s AI Business and Economics tutor — ask any BSAN2201 Prediction and Optimisation Methods question and get a clear, step-by-step explanation grounded in how BSAN2201 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.