| Course | ACC 657 Advanced Data Analytics |
|---|---|
| Topic | Topic 3 |
| Paper type | Rule baseline benchmark |
| Length | About 990 words, 2 pages |
| Format | APA 7 student paper |
| School | Grand Canyon University (GCU) |
| Program | MS in Accounting |
| Updated | September 2026 |
Free sample paper for ACC 657 Topic 3
The Bar a Refund Fraud Model Must Clear: Three Store Rules, 11,400 Flags and 95 Confirmed Cases in 1,000 Reviews
[Author Name]
Colangelo College of Business, Grand Canyon University
ACC-657: Advanced Data Analytics
Topic 3 Assignment
[Instructor Name]
[Date]
The organization, setting and figures below are a composite written as a model document. No real employer, client, colleague or patient is described.
The Bar a Refund Fraud Model Must Clear: Three Store Rules, 11,400 Flags and 95 Confirmed Cases in 1,000 Reviews
The Question and the Limit
Callan Home Goods, a composite chain of 140 stores, loses money to cashiers who process fraudulent refunds: refunds for goods never returned, voids of sales that were in fact completed, and refunds paid to cards the cashier controls. Refund and register fraud of this kind is one of the most common schemes in retail cash handling (Association of Certified Fraud Examiners, 2022). The question is which refund and void transactions warrant investigation.
Any answer is bounded by capacity. The loss prevention team of six can investigate about 1,000 flagged transactions a year, each requiring video review and a store interview. A method that flags more than that forces investigators to choose among its flags, and the choice then matters as much as the method.
Before anyone proposes a predictive model, the store's existing rules need a measured result, stated in a form a model can be compared with. This benchmark provides it.
The Three Rules
Each rule is stated precisely enough to code, and each targets a specific loss.
Table 1
Rules as Coded
| Rule | Definition | Loss it targets |
|---|---|---|
| R1: Unmatched refund | A refund at a store with no sale of the same item at that store in the previous 30 days | Refunds for goods never bought or returned |
| R2: Void after count | A void entered on a register after the drawer for that shift was counted | Voiding completed cash sales and keeping the cash |
| R3: Own-card refund | A refund paid to a card enrolled under the processing cashier's own loyalty number | Refunds diverted to the cashier |
Flags Across the Year
The rules were run against all 2.3 million refund, void and sale transactions for the illustrative year. R1 flagged 7,900 transactions, R2 flagged 2,600 and R3 flagged 1,400. Some transactions trip more than one rule: 460 tripped two and 20 tripped all three. Counting each transaction once, the rules flagged 11,400.
That is more than eleven times what investigators can open. Whatever the rules catch, they cannot be used as they stand.
Flags Against Confirmed Cases
Loss prevention confirmed 380 refund frauds during the year, through investigations, cashier admissions and case closures. The flags contain 250 of them, about two in three. The other 130 were found through tips and stock counts and tripped no rule.
Table 2
Flags and Confirmed Cases by Rule
| Rule | Flags | Confirmed cases in flags | Flags per confirmed case |
|---|---|---|---|
| R1 | 7,900 | 150 | 53 |
| R2 | 2,600 | 70 | 37 |
| R3 | 1,400 | 90 | 16 |
| Any rule, each transaction once | 11,400 | 250 | 46 |
Flags that were real are much scarcer. Only one flag in 46 is a confirmed case. The rules differ sharply: R3 is right about once in 16 flags, R2 about once in 37 and R1 about once in 53. A single accuracy figure would hide all of this, since a rule that flagged nothing would be more than 99.9 percent accurate on 2.3 million transactions. For rare events, the useful measures are the share of cases caught and the share of flags that are real (Saito & Rehmsmeier, 2015).
Two Rankings at 1,000 Reviews
Because investigators can open 1,000 flags, the benchmark ranks all 11,400 and counts the confirmed cases in the top 1,000 under two orderings.
Ranked by refund amount, largest first, the top 1,000 contain 40 confirmed cases. Large refunds are mostly legitimate returns of furniture and appliances, and fraudulent refunds are usually kept small to avoid attention.
Ranked by the number of rules each transaction trips, with amount breaking ties, the top 1,000 contain 95. The 480 transactions that trip two or more rules come first, and they are rich in real cases, since a refund that is both unmatched and paid to the cashier's own card has few innocent explanations.
The second ranking finds more than twice as many cases for the same investigative effort. It becomes the benchmark: 95 confirmed cases in 1,000 reviews.
What the Benchmark Does Not Settle
The 95 is a count of cases found, not of money recovered or losses prevented. Refund frauds confirmed during the year averaged about $1,900 each, but the range is wide, and a method that found fewer, larger cases could be worth more. The benchmark also says nothing about the 130 confirmed frauds no rule caught. Those were found through tips and inventory counts and involved mostly gift card schemes, a pattern the three rules were not designed to see. A proposer who claims a model reaches them should show it on that group separately, since adding them to the overall count would mix two different questions: whether the model ranks the rules' territory better, and whether it covers ground the rules leave open.
What the Labels Hide
The confirmed cases are not a neutral record of all fraud. Many investigations were opened because one of these same rules flagged a transaction, so the rules helped create the labels they are now graded against. A fraud pattern the rules miss is less likely to have been investigated and therefore less likely to appear among the 380 (Bolton & Hand, 2002). This flatters the rules: the benchmark overstates how well they catch fraud in general and understates what a method that finds different patterns could achieve. A model that finds cases outside the rules' reach may look worse on these labels than it is, and its proposers should be allowed to show that with a sample of investigations drawn from its own top flags.
The Bar and How It Must Be Cleared
Any model proposed to replace the rules must find more than 95 confirmed cases in its top 1,000 transactions. It must do so on the following year, held out from its development entirely, with the rules ranked by rules tripped and scored on the same year for comparison (Provost & Fawcett, 2013). A model that beats 95 only on the year it was trained on has not cleared the bar.
References
Association of Certified Fraud Examiners. (2022). Occupational fraud 2022: A report to the nations. ACFE.
Bolton, R. J., & Hand, D. J. (2002). Statistical fraud detection: A review. Statistical Science, 17(3), 235-255. https://doi.org/10.1214/ss/1042727940
Provost, F., & Fawcett, T. (2013). Data science for business: What you need to know about data mining and data-analytic thinking. O'Reilly Media.
Saito, T., & Rehmsmeier, M. (2015). The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS ONE, 10(3), Article e0118432. https://doi.org/10.1371/journal.pone.0118432
ACC 657 Topic 3 instructions, in plain terms
ACC-657 Topic 3 asks students to benchmark rule-based detection before any model is proposed. Many papers report how many flags rules produce. The assignment rewards benchmarks that state each rule precisely enough to code, compare flags with confirmed cases and report both the share of cases caught and the share of flags that are real. Students should rank flags against investigative capacity, explain how labels created by the same rules bias the result and state the figure and held-out period a model must beat.
Inside the ACC 657 Topic 3 example
The finished benchmark states its rules and then grades them against known outcomes. A refund with no matching sale in the previous 30 days is flagged, as is a void entered after the drawer was counted and a refund paid onto a card enrolled under the cashier's own loyalty number. Across the illustrative year's 2.3 million transactions the three rules flag 11,400, and those flags include 250 of the 380 refund frauds loss prevention confirmed, about two in three. Only one flag in 46 is a confirmed case. Because investigators can open about 1,000 flags a year, the benchmark ranks the flags two ways. Ranked by amount, the top 1,000 contain 40 confirmed cases. Ranked by how many rules each transaction trips, they contain 95. That figure, 95 cases at 1,000 reviews, becomes the bar.
Reading the ACC 657 Topic 3 rubric
Faculty grading the benchmark reward a figure a later proposal can quote. Excellent benchmarks show that accuracy is meaningless when a rule flagging nothing scores above 99.9 percent. Credit goes to measures by rule, since precision differs sharply among them. The rubric values ranking flags within capacity and choosing the stronger ordering as the bar. The highest level names the label bias that favors the rules and the frauds they never reach. Benchmarks that stop at total flags usually lose marks with graders.
ACC 657 Topic 3 help: mistakes that cost marks
Stopping at the flag count is where this benchmark most often falls short, because 11,400 hits say nothing about whether the rules find fraud. A paper that reports how many confirmed cases the rules caught but never how many flags were false leaves half the performance unmeasured. Ranking by amount is the common default, and within the same 1,000 reviews it finds fewer than half as many cases as ranking by rules tripped. Ignoring review capacity makes every rule set look usable. The subtle loss is treating confirmed cases as a neutral answer key, when investigators found many of them by following these very rules, which inflates the rules' catch rate. A benchmark that never states the figure a model must beat, and the data it must beat it on, leaves the next topic with no test to run.
Get an ACC-657 Topic 3 example written to your instructions
Send the ACC-657 Topic 3 instructions and your classroom rubric, with the transaction data and any outcome labels your section provides. A custom example gets written to those criteria, with rules tied to their target losses, results graded against known cases, rankings compared at fixed capacity and a bar set for later models, returned within 24 to 48 hours. The first request is free.
More ACC 657 papers and MS in Accounting samples
- ACC 657 Topic 1: Sampling Versus Scoring Memo
- ACC 657 Topic 2: Feature Leakage Review
- ACC 657 Topic 4: Logistic Model Assumption Check
- ACC 657 Topic 5: Overfit Accuracy Dq Post
- ACC 657 Topic 6: Model Drift Monitoring Plan
- ACC 657 Topic 7: Model Rerun Specification
- ACC 657 Topic 8: Champion Model Defense Memo
- ACC 658 Topic 3: Modified Accrual Recognition Worksheet
- ACC 656 Topic 3: Goodwill and Bargain Purchase Problem
- ACC 660 Topic 3: Functional Currency Translation Worksheet
- ACC 653 Topic 3: Quality Cost Trade-off Report
ACC 657 Topic 3 questions, answered
Why grade rules if a model is coming anyway?
Because a model is only worth adopting if it does better than something simpler, and nobody can say whether it does without a measured baseline. Rules are cheap, transparent and already understood by investigators. If a model finds 97 cases in 1,000 reviews against the rules' 95, the added complexity may not pay for itself. Measuring the rules first gives that judgment a number to stand on.
What is wrong with using confirmed cases as labels?
Nothing, provided their origin is stated. Confirmed cases are the best outcome data most organizations have, but they exist only where somebody looked, and people looked where earlier rules pointed. Frauds that no rule catches rarely appear among them at all. The benchmark therefore understates what the rules miss, and a careful paper names that bias and suggests a random review of unflagged refunds to estimate it.
Can the rules be tuned to improve the benchmark?
Not on the same data used to grade them. Adjusting a threshold until the confirmed-case count rises and then reporting that count measures how well the rules fit this year, not how well they will work next year. If tuning is part of your assignment, hold back a later period, tune on the earlier one and report the result on the period the tuning never touched.
Where can I find a free ACC 657 Topic 3 sample paper?
The full benchmark sits above, with the question and limit, the three rules, flags across the year, flags against confirmed cases, two rankings, what it does not settle, what the labels hide and the bar. Share your rules for a free worked benchmark.
Why rank flags by rules tripped in ACC 657 Topic 3?
Transactions tripping two or more rules have few innocent explanations, so the top 1,000 hold 95 confirmed cases against 40 when ranked by amount.