Table 5 Comparative advantages of...

Table 5

Comparative advantages of Bayesian vs Frequentist approaches for fraud auditing

Objective of fraud audit	Bayesian	Frequentist
Controls
Preventive controls over fraud (passive)	No	No
Detective control (group or individual transactions)	Yes	Yes
Corrective control allowing lost cost, accurate recovery from fraud	No	No
Economics
Able to identify transaction sets that are fraudulent	Yes	Yes
Objective is maximizing net savings from fraud	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to apply firm’s actual loss function	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to calculate loss under competing decisions	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to compute fraud cost	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to calculate loss under competing decisions	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Operations
Generalist algorithm	Yes	No, requires a hypothesis testing framework
Empirical fraud detection methodology	Yes	Yes
Supports labeling of individual transactions as potentially fraudulent	Yes	Yes, with limitations
Can be scaled up for large transaction volumes	Yes	Yes, though p−values for a decision are less and less reliable as the application is scaled up to larger transaction numbers
Simple and low cost to implement	Yes	No, requires a hypothesis testing framework
Highly efficient, low cost transaction processing	Yes	Yes
Many tools available for implementation	No	Yes
Comparison with competitive methods
Autoencoders	Competitive	Not Competitive
Benford tests	Competitive	Not Competitive
Sarbanes-Oxley tests	Competitive	Competitive for Section 302 tests, but not for Section 404 tests
Supervised Rule-based methods	Competitive	Competitive
Supervised Tree-based algorithms	Competitive	Competitive
Supervised Methods with misclassification	Competitive	Competitive
Unsupervised classification methods	Not competitive, requires labeling	Not competitive, requires labeling

Objective of fraud audit	Bayesian	Frequentist
Controls
Preventive controls over fraud (passive)	No	No
Detective control (group or individual transactions)	Yes	Yes
Corrective control allowing lost cost, accurate recovery from fraud	No	No
Economics
Able to identify transaction sets that are fraudulent	Yes	Yes
Objective is maximizing net savings from fraud	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to apply firm’s actual loss function	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to calculate loss under competing decisions	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to compute fraud cost	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Able to calculate loss under competing decisions	Yes	No, p−values only estimate the probability that our decision is correct for a group of transactions being fraudulent
Operations
Generalist algorithm	Yes	No, requires a hypothesis testing framework
Empirical fraud detection methodology	Yes	Yes
Supports labeling of individual transactions as potentially fraudulent	Yes	Yes, with limitations
Can be scaled up for large transaction volumes	Yes	Yes, though p−values for a decision are less and less reliable as the application is scaled up to larger transaction numbers
Simple and low cost to implement	Yes	No, requires a hypothesis testing framework
Highly efficient, low cost transaction processing	Yes	Yes
Many tools available for implementation	No	Yes
Comparison with competitive methods
Autoencoders	Competitive	Not Competitive
Benford tests	Competitive	Not Competitive
Sarbanes-Oxley tests	Competitive	Competitive for Section 302 tests, but not for Section 404 tests
Supervised Rule-based methods	Competitive	Competitive
Supervised Tree-based algorithms	Competitive	Competitive
Supervised Methods with misclassification	Competitive	Competitive
Unsupervised classification methods	Not competitive, requires labeling	Not competitive, requires labeling

[ViewLarge]

Sharing Unavailable