| Course | BU 351 Introduction to Business Analytics |
|---|---|
| Unit | Unit 5 |
| Paper type | Predictive exercise |
| Length | About 1,010 words, 6 pages |
| Format | APA 7 student paper |
| School | Herzing University |
| Program | BS in Business Analytics |
| Updated | September 2026 |
Free sample paper for BU 351 Unit 5
Less Accurate Than Guessing "Nobody Leaves," Yet Worth Using: A Predictive Exercise Ranking Climbing Gym Members by Their Risk of Leaving in the Next Two Months
Student Name
Business, Herzing University
BU 351: Introduction to Business Analytics
Instructor Name
March 18, 2026
Less Accurate Than Guessing "Nobody Leaves," Yet Worth Using: A Predictive Exercise Ranking Climbing Gym Members by Their Risk of Leaving in the Next Two Months
What Is Being Predicted
This exercise builds a simple model for Basecamp Climbing's retention team to answer the third question from the Unit 1 framing, which asks how well members active at the end of June 2025 can be sorted by their risk of leaving over the following two months, using only what was known on that date. The prediction is made for individual members, not for the chain as a whole, and the horizon is 60 days, which is long enough for a retention action to be tried and short enough for staff to act on the list. On June 30, 2025, 1,589 members were active, and 176 of them, 11.1 percent, canceled by August 29.
The Split
Before any model was fitted, the 1,589 members were randomly divided: 1,112, about 70 percent, were used to fit the model, and 477 were held back to test it. The test group contained 61 members who canceled, 12.8 percent. The model never saw the test group while it was being fitted, so its performance there estimates how it would do on members it has not met. Shmueli and Koppius (2011) stress that predictive power should be judged on data not used to build the model, since a model always looks better on the records it was fitted to.
A random split was chosen over a split by time, such as fitting on members who joined earlier and testing on those who joined later, because the prediction date is fixed and every member in the snapshot shares it. A time-based check will still be needed later: the real test of the model is whether a version fitted in June 2025 ranks members well in, say, January 2026, after conditions have changed.
The Model, Described in Words
The model is a logistic regression, which estimates the probability of an outcome that is either yes or no, here whether a member cancels within 60 days, from a set of inputs (Hosmer et al., 2013). It uses seven inputs, each chosen because the exploration in Unit 4 suggested it mattered and because it would be known on the prediction date. Whether the member joined within the previous three months is included because early months carry the most cancellations. Class attendance and referral status go in because both were associated with lower early cancellation. The number of autopay failures is included because failures were strongly associated with leaving. Student and family plans are included because early cancellation differed by plan. And an indicator for the Aurora gym is included because Aurora's rate stood apart. Recent visits were deliberately excluded because, as Unit 2 found, that field is recorded at each member's last date and would leak the outcome into the inputs.
What the Model Found
The fitted model's coefficients are easiest to read as changes in the odds of canceling, holding the other inputs constant. Being in the first three months of membership multiplied the odds by about 2.65. Each autopay failure multiplied the odds by about 1.70. A student plan multiplied them by about 1.78. Taking the introductory class cut the odds to about 0.56 of what they would otherwise be, and a referral to about 0.82. The family plan and the Aurora indicator made little difference once the other inputs were included, with odds multipliers of 0.98 and 1.08. These are associations in historical data, not proof that changing any input would change behavior.
The two strongest inputs, early membership and payment failures, match the exploration in Unit 4, which is a useful check that the model is picking up real patterns rather than noise in the training data.
How Wrong It Was
Accuracy is the wrong measure here. In the test group, 416 of 477 members stayed, so a rule that predicted no one would leave would be right 87.2 percent of the time. The model, using a cutoff that flags members with a predicted probability of 15 percent or more, was right 80.7 percent of the time, which is worse. But the rule that predicts no one will leave is useless, since it identifies no one to help.
The measures that matter for a retention team are how many flagged members actually leave and how many leavers are caught. At the 15 percent cutoff, the model flagged 81 members in the test group. Of those, 25 canceled, a hit rate of 31 percent, compared with 12.8 percent if members were chosen at random. The model caught 25 of the 61 members who canceled, 41 percent, and missed 36. Ranked from highest risk to lowest, the top fifth of the test group, 95 members, canceled at 30.5 percent and included 47.5 percent of all cancellations. For a staff with time to contact only some members, the model more than doubles the value of each call compared with calling at random.
| At a 15% cutoff, test group of 477 | Actually canceled | Actually stayed |
|---|---|---|
| Flagged by the model | 25 | 56 |
| Not flagged | 36 | 360 |
What Would Break It
Four conditions would weaken the model. First, a change in prices or plans, like the March 2025 increase, alters who leaves and why, so a model fitted on one period may not hold in the next. Second, the autopay failure count in the file covers the twelve months before each member's last date rather than a fixed window ending June 30, so it may carry a little information from after the prediction date; a production version should rebuild that field as of the prediction date. Third, a new gym or a change in staff would add conditions the model has never seen. Fourth, if the chain acts on the list and keeps some flagged members, the next year's data will show fewer cancellations among high-risk members, and a model refitted on that data may learn that the risk factors matter less than they do. The model should be refitted every six months and checked against the held-back approach each time. The model, the gyms and the members are all fictional, built for teaching (Provost & Fawcett, 2013).
References
Hosmer, D. W., Jr., Lemeshow, S., & Sturdivant, R. X. (2013). Applied logistic regression (3rd ed.). Wiley.
Provost, F., & Fawcett, T. (2013). Data science for business: What you need to know about data mining and data-analytic thinking. O'Reilly Media.
Shmueli, G., & Koppius, O. R. (2011). Predictive analytics in information systems research. MIS Quarterly, 35(3), 553-572. https://doi.org/10.2307/23042796
The Herzing BU 351 Unit 5 instructions
For the predictive unit, students build a simple predictive model and evaluate it honestly. Most versions want the target defined with its level and time horizon, a training and testing split described, the model and its inputs explained in words, results reported on the test data and a discussion of how accurate the model is in terms the business can use. Many rubrics also ask what would make the model fail. Common techniques include linear or logistic regression and simple decision trees. Instructors care more about honest evaluation than high accuracy. Expect a few pages with a results table or confusion matrix, citing modeling sources in APA 7.
How this BU 351 Unit 5 example is built
The sample defines the prediction, which members active on June 30, 2025 would cancel by late August, and gives the base rate of 11.1 percent. The split section explains the random 70 to 30 division and why testing on held-back members matters. The model section names each input with a reason and explains the excluded visits field. Results are reported as odds multipliers in plain language. The evaluation section explains why accuracy misleads with a rare outcome, then reports hit rate, catch rate and lift, with a small confusion matrix. A final section lists four conditions that would weaken the model and recommends refitting every six months.
Where the points sit in the BU 351 Unit 5 rubric
Herzing's rubric for the predictive exercise commonly includes definition of the target, the data split, explanation of the model, honest evaluation and discussion of limitations. The target earns full points when its level and horizon are clear. The split is judged by a stated method and a test set kept separate from fitting. Model explanation rewards inputs described with reasons before output appears. Evaluation carries heavy weight and rewards metrics suited to the problem, such as precision and recall for a rare outcome, rather than accuracy alone. Instructors reward papers that identify leakage and conditions that would break the model, as the sample does.
BU 351 Unit 5 help before you submit
The most common Unit 5 mistake is reporting accuracy on the training data, which always looks better than performance on new records. Another is celebrating high accuracy when the outcome is rare and a naive rule would score higher. Students often include inputs that would not be known at the time of prediction, which leaks the answer into the model. Some papers present coefficients without translating them into plain language. Others skip limitations, implying the model will work forever. Finally, compare your model with a simple baseline; showing how much better it does than random selection or a naive rule is what makes the result meaningful. State the base rate at the start, since every metric is easier to judge against it.
Get a BU 351 Unit 5 example written to your instructions
Hand us the Unit 5 instructions, your BU 351 rubric, the dataset and whichever technique the assignment requires. We write a custom example that fixes target, grain and horizon first, splits the records before fitting, justifies every input, reports error on the held back portion and names what would break the model. First custom sample free, back in 24 to 48 hours.
More BU 351 papers and business analytics samples
- BU 351 Unit 1: Analytics types and framing
- BU 351 Unit 2: Data quality assessment
- BU 351 Unit 3: Descriptive analysis
- BU 351 Unit 4: Exploratory analysis
- BU 351 Unit 6: Prescriptive or optimization exercise
- BU 351 Unit 7: Reporting to a business audience
- BU 351 Unit 8: End-to-end analytics case
- BU 453 Unit 5: Single-table queries
- BU 454 Unit 5: Distribution and relationship charts
- BU 455 Unit 5: Distributed processing exercise
- BU 407 Unit 5: Sensitivity table analysis
BU 351 Unit 5 questions, answered
Which technique should the exercise use?
Whatever the instructions name, and where the choice is yours, the simplest one that fits the target. A straight line through a trend, a regression on two or three inputs, or a basic classification rule are all appropriate at this level and all easier to explain than something elaborate. The criterion rewards being able to say why the technique suits the question.
How should the data be split?
By whatever keeps the held back records genuinely unseen. For anything with a time dimension, split by date and hold back the later period, because a random split lets the model learn from the future. For records with no ordering, a random portion set aside works. Say which you used and why, since the reasoning is part of what is graded here.
How accurate does the model have to be?
Accuracy is not what the criterion measures at this level. A model that performs modestly and is honestly evaluated scores better than one claiming near perfect results with no holdout behind them. What the marker wants is a defensible setup, error reported against unseen records, and a clear statement of what the figure means for the decision the case describes.
Where can I find a free BU 351 Unit 5 sample paper?
The whole BU 351 Unit 5 exercise, from the split to the confusion matrix, is open above to read at no cost. If your data or target differ, send your prompt and a description of the variables, and a free custom predictive exercise will be prepared, generally in 24 to 48 hours.
How accurate does the BU 351 Unit 5 model have to be?
Useful matters more than accurate. The sample's model scores below a rule that predicts nobody leaves, yet its top-ranked members cancel at more than twice the average rate. Explain how the model would help a decision, and compare it with a baseline.