Like what you see? Click here to sign up for Accounting Today's daily newsletter to get the latest news and behind the scenes commentary you won't find anywhere else.
Audit sampling is selecting a group of items such as invoices for investigation to draw inferences about an account balance. Ratio analysis involves comparisons between two financial statement accounts such as current ratios and gross profit percentage. Reasonable tests involve using financial and nonfinancial data to estimate an account balance. An example would be multiplying items sold by price to determine expected revenue.
However, there are audit engagement risks with current auditing techniques. A sampling risk is the risk that an auditor’s conclusion might differ if the auditor had been able to examine an entire population rather than inferring from an audit sample. Relying on audit sampling increases the risk that fraudulent activities might not be detected in a sample.
Fraudsters can often predict the analytical procedures the auditor will perform, such as ratio analysis and reasonable tests. In response, the fraudster will design the fraud to be prevented from detection. These risks increase the probability of litigation from client shareholders, fines and sanctions by governing bodies and negative publicity.
In June 2012, the accounting firm Mayer Hoffman McCann P.C. settled with the State of California for its audit of the City of Bell. In a December 2010 report by the California State Controller Office, MHM was criticized for not following audit procedures, which included neglecting to increase its sample size to test transactions. MHM relied on sample sizes that were not representative of account balances to issue a clean audit opinion. Had MHM increased its sample size and followed audit procedures, MHM might have identified the deficiencies to discover the multimillion-dollar fraud and embezzlement of public funds by city officials. MHM settled with the state by paying a $300,000 fine and its license was placed on probation for a period of two years. The firm originally faced losing its license to practice in California and a million-dollar fine.
As business entities grow and use more complex computerized accounting information systems, auditors have to rely on sampling procedures for practicality. However, relying on sampling reduces audit quality and increases audit business risk. This is because of the labor cost of manually auditing an account balance and the time pressure for completing the audit before the deadline.
Artificial neural networks and case-based reasoning systems can improve audit quality and reduce audit engagement risk by analyzing an entire account population. An ANN is an interconnected network of artificial neurons to model relationships between inputs and outputs and to find patterns in data. An ANN works by having input neurons at a layer that sends the information received to internal layers called the hidden layer of neurons via synapse connections. Using historical data to train the ANN, it learns patterns and strengthens or weakens the synapse connections between neurons. The hidden layers of neurons send information to output neurons.
An ANN could continuously audit and monitor account balances. Using a large municipality in a study, researchers Eija Koskivaara and Barbro Back used ANNs to predict account balances of 10 accounting units. The ANNs were given between five to seven years of monthly account balances to train and model. The researchers defined the ANN to “forget” learning, required the ANN to predict account balances from values of other account balances, and required the ANN to predict account balances every month for up to one year into the future.
On average, the ANN models in the study were able to predict within 12 percent of the actual account balances compared to 19 percent by the budgeted amount. The ANN was a more accurate predictor compared to other analytical methods such as using previous monthly ending balances or an average of ending account balances. The ANN predicted with greater accuracy accounts which had trends, were more stable and were related to other accounts. Researchers also demonstrated that with more historical data to train the ANN, it could predict the account balances more accurately.
In another study, conducted by Hsueh-Ju Cheng, Shaio-Yan Huang and Chung-Long Ku, an ANN turned out to be a better predictor and detector of fraud litigation than regression analysis and auditor judgment. Researchers used 50 fraud and 100 non-fraud cases of public companies in the same industry and time period to train an ANN, while 25 fraud cases and 50 non fraud cases were used for testing. A group of 30 Big Four CPA audit professionals with an average of 11 years of experience were then asked to use their professional judgment to determine whether fraud litigation would occur in a test case which was either fraudulent or not fraudulent. A regression model was also used, employing fraud risk factors identified in test cases.
The ANN was able to predict 81 percent of the test cases of public companies as fraudulent compared to 73 percent of the regression model and only 60 percent by auditors using their professional judgment. Perhaps more alarming, the auditors incorrectly identified 50 percent of the test companies as false negatives. A false negative is when a test case company is considered to have a low fraud litigation risk when it was fraudulent. The regression model fared worse, with an error rate of 68 percent. However, the ANN had a 26 percent misjudgment rate, which demonstrates its ability as a better fraud predictor than regression analysis and auditor judgment.
Case-based reasoning systems can also be used by auditors. CBR is a form of artificial intelligence in which current problems are solved based on the solutions of similar problems in the past. CBR solves problems by first retrieving cases similar to the current problem. It reuses information and solutions from past cases to solve current problems by simulating the past solution for the current problem. If there are differences, a CBR revises the solution. Once the solution is adapted to the current problem, it is then stored in memory.
CBR systems can be used to identify suspicious transactions. Using transaction data from a large national bank, 39 risky transactions were tested for research, in a study by Gun Ho Lee. The CBR system was given 150 example transactions to make conclusions on the 39 risky transactions. For 31 out of the 39 problems, the CBR system was able to identify a conclusion from a similar example transaction. The CBR system was able to revise the conclusion for the remaining eight and store the solutions as an example.
The research demonstrated that a CBR system can accurately review and effectively identify suspicious transactions within a database in a short amount of time. The CBR system is more consistent with assigning risk, since auditors reached different conclusions on similar cases. Once the prototype CBR system had 1,000 example cases, the CBR system was able to reuse solutions for 97 percent of the test cases. The more example cases the CBR system has, the better the auditor could use the system to identify and assess risks.
Auditors can start using and implementing ANN and CBR techniques today at a low cost. There is off-the-shelf ANN software (some for free and open source) that can forecast account balances. These software programs also work with Microsoft Excel, which makes it easy for auditors and other users to learn and implement. Several ANN software programs on the market are compatible with Microsoft Excel, including NeuroXL, NeuroTools and Alyuda Forecaster XL. Neural network software has already been implemented as tools for stock and financial forecasting. An example of CBR software that works with Excel is CasePower by Inductive Solutions. However, custom solutions can also be developed for specific auditor requirements.
Those tools are a small investment but can provide significant time savings and improved quality for auditors, since a computer program automates the audit process, performs stronger data analytics and provides better predictability. An ANN can analyze an entire account population continuously, and the auditor can use professional judgment to determine whether the ANN predicted balance and the reported balance is material enough to need further testing. Or, an ANN can predict with greater accuracy the likelihood a public company is committing fraud compared to human auditors.
CBR systems have demonstrated how they identify risky transactions that an auditor should investigate further. This reduces the wasted time for auditors on testing transactions that do not have issues. ANNs and CBR systems have proven they offer better audit effectiveness, better audit quality and reduce audit business risk at a low cost for public accounting firms. It’s time these tools are used by auditors.