Difficulty: Intermediate | Prerequisites: Familiarity with classification, regression, clustering, and time-series models (Modules 1–16).
This topic pulls together multiple analytics model types you have studied individually and applies them to a single, real-world business problem. The power company case is a capstone-style exercise: you need to pick the right model for each sub-problem, justify it, and explain what data it requires. If you are comfortable with classification, regression, time-series forecasting, clustering, optimisation, and simulation as standalone concepts, you are ready for this material. If any of those feel shaky, revisit them first.
A power company wants to shut off electricity for customers who can pay but refuse to. The analytics challenge breaks into three linked sub-problems: identify which customers will not pay (classification or logistic regression), estimate the cost of leaving power on versus shutting it off (time-series or regression models), and select which shut-offs to prioritise given limited workers and travel time (optimisation, clustering, or simulation). The answer framework throughout is: given {data}, use {model} to {result}.
Analytics discussion framework
The structured sentence "given {data}, use {model} to {result}" used to state an approach clearly. Think of it as a fill-in-the-blanks template that forces you to name your inputs, your method, and your output in one line.
Classification (SVM, KNN)
Supervised learning models that assign each observation to a discrete category. In this case, the categories are "will pay," "can pay but won't," and "can't pay." In simple terms, classification draws a boundary line between groups so new data points land on one side or the other.
Clustering (k-means)
An unsupervised method that groups observations by similarity without pre-labelled categories. Think of it as letting the algorithm find natural groupings in the data, though those groups may not map neatly onto the categories you actually care about.
Logistic regression
A model that outputs a probability (0 to 1) rather than a hard yes/no. In simple terms, it tells you how likely a customer is to not pay, and you then choose a cut-off threshold to convert that probability into a decision.
Exponential smoothing / ARIMA
Time-series forecasting methods that use a customer's past usage data to predict next month's usage. Think of them as weighted averages of historical values, where more recent data counts more heavily.
GARCH
A time-series model that estimates how much a value fluctuates (its variance), not just its average level. In simple terms, it tells you not only what a customer's bill is likely to be, but how wildly it might swing.
Vehicle Routing Model (VRM)
An optimisation model designed to find the best set of routes for a fleet of vehicles visiting multiple locations. Think of it as a sophisticated version of the travelling-salesman problem, applied here to scheduling shut-off crews.
Expected cost formulas
E[cost of keeping power on] = p(no pay) x E[cost of power next month]. E[cost of shutting power off] = (shut-off cost) + (1 - p(no pay)) x (turn-on cost). These let you compare the financial trade-off of acting versus waiting.
Protected characteristics
In the US, race, sex, age, and other demographic factors (or proxies highly correlated with them, such as location) that are illegal to use as model inputs. This is a legal constraint on feature selection, not a modelling choice.
When approaching a case study in this course, structure every answer with the sentence: given {data}, use {model} to {result}.
This forces clarity on three fronts: what inputs you need, which analytical method fits, and what output you expect. There is rarely a single correct answer, and the goal is a reasonable, defensible approach rather than a fully trained model.
Example applications
Given {child's age, sex, current height, parents' heights}, use {linear regression} to {predict the child's adult height}
Given {crude oil available each day at supply ports, demand forecasts, tanker capacities}, use {an optimisation model} to {determine the best pickup, delivery, and routing schedule}
Given {minute-by-minute traffic data on twelve websites}, use {ARIMA and GARCH} to {predict mean and variance of traffic five minutes ahead}
A power company has customers who do not pay their bills. Some of those customers genuinely cannot afford to pay; the company runs a donation fund for them and is not concerned about this group. The company is concerned about customers who can pay but choose not to.
The company wants to shut off power for the second group, but faces two constraints:
Accuracy constraint – they only want to shut off power for customers who will never pay. Customers who simply forgot or fell behind should be left alone.
Capacity constraint – shut-offs are manual. An employee must travel to each location to physically turn off the power. There are more locations needing shut-off each month than there are workers to handle them.
This creates three linked analytical sub-problems:
Customer identification – which customers can pay but won't?
Cost estimation – what does it cost to leave power on versus shutting it off?
Shut-off selection – given limited crews and travel time, which shut-offs should be done first?
The goal is to identify customers who can pay but won't.
Useful data features
Credit score
Income
Past history of defaults
Past power bill payment history
Value of home
Rent vs. own
How long they have lived at the address
Marital status
Number of people in household
Legal constraint on features
Race, sex, age, and other demographic factors are illegal to use in US models. Proxies that are highly correlated with protected characteristics (such as zip code or location) must also be excluded. This is a feature-selection constraint you must acknowledge in any answer.
Approach 1 – Classification (yes/no answer)
Use SVM or KNN to sort customers into discrete groups: will pay, can pay but won't, or can't pay. This gives a clear, binary decision on who gets targeted. The trade-off is that you lose nuance: a customer near the boundary gets the same treatment as one deep inside a category.
Approach 2 – Clustering (yes/no answer)
Use k-means to let the algorithm find natural groupings. This is quick to implement, but the resulting clusters may not map cleanly onto the three categories you actually care about. You would need to interpret each cluster after the fact.
Approach 3 – Logistic regression (probability answer)
Logistic regression outputs a probability of non-payment for each customer. This is more informative than a hard yes/no, but requires a second step: choosing a threshold (e.g., p > 0.7 = "won't pay") to convert probabilities into decisions. Tree-based approaches can also produce probability outputs.
Approach 4 – Hybrid
First cluster customers, then run a separate analysis (classification or regression) within each cluster. This can capture different dynamics across customer segments but adds complexity.
There is no single correct model. The point is to justify whichever approach you choose and explain why it suits the data and the business question.
Students often think there is one correct model for each sub-problem. There is not. The course is testing whether you can justify a reasonable approach, not whether you pick the "right" algorithm.
Students sometimes include age, gender, or zip code as features without acknowledging the legal constraint. In the US, protected characteristics and their proxies must be excluded from the model.
Logistic regression is sometimes confused with linear regression here. Logistic regression outputs a probability bounded between 0 and 1; linear regression does not.
Students forget that a probability model (logistic regression) requires a threshold to produce an actionable yes/no decision. The probability alone does not tell the company what to do.
⚠️ The "given {data}, use {model} to {result}" framework is the expected format for discussion answers. Practise writing your responses in this structure.
⚠️ Know the legal constraint on features. If an exam question lists demographic data, you should flag which variables cannot be used and why.
⚠️ Be able to name at least two different modelling approaches for the customer identification problem and explain the trade-off between them (e.g., classification gives a hard boundary; logistic regression gives a probability that needs a threshold).
⚠️ Understand that the power company case has three linked sub-problems, not one. An exam answer that addresses only identification without mentioning cost estimation or shut-off prioritisation is incomplete.
True or False: K-means clustering is a supervised learning method. (False – it is unsupervised.)
True or False: Zip code is always a safe feature to include in a US predictive model. (False – it can serve as a proxy for race or income and may be legally problematic.)
Fill in the blank: Logistic regression outputs a ______, which must be converted to a yes/no decision using a ______. (probability; threshold)
True or False: The power company wants to shut off all customers who have not paid. (False – only those who can pay but choose not to.)
Fill in the blank: The analytics framework sentence is: given {}, use {} to {______}. (data, model, result)
Q: Using the analytics framework, how would you identify which customers can pay but won't?
A: Given {credit score, income, past default history, past payment history, home value, rent/own status, length of residence, marital status, household size}, use {logistic regression} to {estimate the probability that each customer will not pay their bill}. Then apply a threshold to convert probabilities into a yes/no shut-off decision.
Q: Why can't you include zip code as a feature in a US model for this problem?
A: Zip code is often highly correlated with race and other protected demographic characteristics. Using it could amount to indirect discrimination, which is illegal under US law.
Q: What is the difference between a classification approach and a logistic regression approach to this problem?
A: Classification (SVM, KNN) assigns each customer directly to a category (will pay, won't pay, can't pay) and gives a hard boundary. Logistic regression outputs a probability of non-payment, which is more nuanced but requires a threshold to convert into a decision.
Q: Name one advantage and one disadvantage of using k-means clustering for customer identification.
A: Advantage: it is quick to implement and does not require labelled training data. Disadvantage: the clusters it finds may not correspond to the categories you actually care about (will pay, won't pay, can't pay), so interpretation is required after the fact.
This material connects to earlier modules on classification (SVM, KNN), clustering (k-means), and logistic regression by applying them to a single integrated problem. The cost estimation and shut-off selection sub-problems (covered in Part 2 of these notes) draw on time-series forecasting (ARIMA, exponential smoothing, GARCH) and optimisation models. The legal constraint on features is a data ethics topic that recurs throughout the course.
Power company case study, analytics framework, given-use-to, classification, SVM, KNN, k-means clustering, logistic regression, probability threshold, customer identification, non-payment prediction, protected characteristics, feature selection, legal constraints on data, credit score model, payment history, supervised vs unsupervised learning, Modules 17–21, discussion cases, UT Austin intro to computer programming