The IRS Audit Credit-Card-to-Cash Estimation Method for Cash Businesses – Houston Tax Attorneys


When it comes to income taxes, cash businesses have always been a challenge for the IRS. Cash is hard to track. Businesses, whether large or small, often fail to keep records of cash transactions. In other cases, businesses keep the records lose the records by the time the IRS audits the business years later. And there are businesses that simply underreport cash, knowing that the IRS is unable or unlikely to notice or be able to do anything about it.

This is one aspect of tax where the widespread adoption of crypto currency could make the IRS audit much easier. In theory, it could eliminate the need for IRS audits. But absent something like that, the IRS will continue to audit cash heavy businesses and use estimation methods to identify what it seems to be unreported income. The amount of tax at issue is significant. This is not a minor issue for taxpayers or for the IRS.

We have addressed a few of these types of income-reconstruction cases on this site before. This time, we are going to consider the credit-card-to-cash method the IRS uses for cash-heavy businesses. The recent Clinco v. Commissioner, T.C. Memo. 2026-16, case provides an opportunity to consider this method. It involves a restaurant and the IRS audit that applied this method to estimate–and increase–the business’ income and tax.

Facts & Procedural History

The taxpayer ran a restaurant and bar near UCLA. The cafe was a family operation. The taxpayer owned 66.6%, a brother owned the remaining interest through 2014, and a third brother served as the manager and bookkeeper. In 2015, the taxpayer converted the entity that owned the cafe into a single-member LLC.

The cafe had about 60 employees and most of them only worked three to four hours per shift.

The taxpayer prepared his own 2015 tax returns. On his Schedule C for for the cafe, the taxpayer reported gross receipts of more than $1.6 million. After claiming $1.4 million in cost of goods sold and $600,000 in total expenses, he reported a net loss of about $400,000.

The IRS pulled the tax returns for audit in 2019. During a personal meeting, the IRS noted that the taxpayer had estimated that 10% of restaurant revenues came from cash.

The IRS agent believed there was a discrepancy between re…

The IRS agent believed there was a discrepancy between reported gross receipts and what the audited gross receipts should have been. This hunch was based on a credit-card-sales-to-cash ratio. This hunch prompted the IRS agent to summons the taxpayer’s bank records and conduct a full bank deposits analysis. She also pulled third-party information reporting returns data, including four Forms 1099: a Form 1099-MISC from UCLA and three Forms 1099-K from First Data Reporting, American Express, and Grubhub. The IRS agent purportedly reconciled the bank deposits analysis with this third-party data and the taxpayer’s statement about cash receipts. The result was reconstructed gross receipts of approximately $2.29 million—well above what the taxpayer had reported.

The IRS issued an IRS Notice of Deficiency for 2015 asser…

The IRS issued an IRS Notice of Deficiency for 2015 asserting underreported gross receipts. The taxpayer then died shortly after the notice was sent. His wife timely filed a petition with the U.S. Tax Court to challenge the IRS determination.

When Can the IRS Reconstruct a Taxpayer’s Income?

This case starts with the fundamental question as to whether the IRS can reconstruct a taxpayer’s income and, if so, then when can to do so? We have touched on these topics several times on this site.

Section 446(b) of the tax code gives the IRS authority to compute a taxpayer’s income. To do so, the IRS has to use a method that clearly reflects income. The IRS is generally only able to do this if the taxpayer’s own method does not clearly reflect income.

Courts have interpreted these concepts broadly. The rules can be summarized by the idea that when a taxpayer’s books and records are incomplete, unreliable, or do not match third-party information, the IRS can step in and calculate income using a reasonable method. The reconstruction does not have to be perfect. It has to be reasonable in light of all surrounding facts and circumstances.

The IRS has several recognized indirect methods. For example, the net worth method compares assets and liabilities at the beginning and end of a year. The markup method applies industry-standard percentages to cost of goods sold. These two methods are not all that common.

The IRS’s go-to method is the bank deposits method

The IRS’s go-to method is the bank deposits method. This method adds up all deposits, subtracts identifiable nontaxable items, and treats the remainder as income. Thus, an IRS audit where the bank deposits method is used requires the IRS to review every deposit and make judgment calls about what is income and what is not. In a cash-heavy business like a restaurant, these judgment calls can get complicated fast and they can lead to trade-offs that make the results incorrect and, in most cases, unreliable. This brings us to the credit-card-to-cash ratio.

About the Credit-Card-to-Cash Ratio

Theh credit-card-to-cash ratio is also not a method that the IRS uses all that often. It is usually used in cash heavy businesses given the limitations of a pure bank deposit method.

For this credit-card method, as in this case, the IRS agent will ask the taxpayer what percentage of sales comes from cash versus credit cards. That answer sets the stage for everything that follows. If the known credit card receipts represent 90% of gross income, and the examiner can verify them through Forms 1099-K, then simple math yields the estimated cash component.

The IRS Retail Industry Audit Technique Guide specifically instructs examiners to ask about this ratio during the initial interview and verify it against bank deposits.

This is what happened in this case. The IRS agent verified credit card receipts through Forms 1099-K totaling over $2 million from First Data Reporting, American Express, and Grubhub. She added the Form 1099-MISC from UCLA. She then applied the taxpayer’s own 10% cash estimate to arrive at $228,929 in estimated cash receipts. The total reconstructed gross receipts came to approximately $2.29 million–which was in excess of the $1.6 million the taxpayer reported.

The IRS agent could have also used industry data to provi…

The IRS agent could have also used industry data to provide the percentage. The IRS will use industry data if they are working with a tax savvy taxpayer who does not volunteer an estimate of cash payments or who offers an unreasonably low estimate. The case does not address it, but the IRS agent didn’t have to do that here as it accepted the taxpayer’s own estimate. It may be that the estimate provided was in line with industry standards and the IRS agent checked, or it may be that the IRS agent just accepted the estimate to move on with the audit.

Defending Against the Credit-Card-to-Cash Ratio

The credit-card-to-cash ratio is clearly just an estimate. This is why it is not used all that often. But with that said, the courts have sanctioned the use of this method in other cases.

For example, the Supreme Court allowed this method to be used in United States v. Fior D’Italia, Inc., 536 U.S. 238 (2002), in the context of FICA taxes on unreported tips. There, the IRS examined a restaurant’s credit card slips, calculated the average tip percentage, assumed cash-paying customers tipped at the same rate, and multiplied the resulting rate by total receipts to estimate aggregate unreported tips. The Court held that this aggregate estimation method was authorized by law, so long as the method was reasonable.

Back to the present case, in this case, the court noted that the taxpayer had the burden to disprove the IRS agent’s conclusions. The taxpayer raised several arguments against the IRS adjustment to income. Each one fell short. They were that the IRS failed to label the 1099 income correctly (even though it was reported by third parties to the IRS), that the IRS failed to account for non-taxable capital contributions to the business, and that the IRS had confused this cafe with a similar one in the area that was not owned by the taxpayer. The court noted that mere arguments as to these items were not enough to overcome the IRS’s determination.

The court opinion does not address this

The court opinion does not address this, but the way to overcome the taxpayer’s burden in these cases is to put a better reconstruction in front of the court. The taxpayer’s own records—even partial ones—can serve as a starting point. A forensic accountant or other expert can take those records and build an alternative analysis that accounts for variables the IRS overlooked, such as non-cash tips paid out to employees, comped meals, vending or catering revenue that does not follow the same cash-to-card split, or seasonal fluctuations that make a single annual ratio misleading.

Industry data from comparable restaurants can also be used to challenge the IRS’s assumed percentages. The goal is not necessarily to prove the IRS’s number is wrong down to the dollar. The goal is to get competing evidence into the record so the court has something to weigh against the IRS’s analysis. Without that, the court is left with only the IRS’s method and the taxpayer’s unsupported objections—and as this case shows, unsupported objections do not carry the day.

The Takeaway

The IRS’s has broad authority to reconstruct income. It frequently does this as part of its “income probe” when it starts an IRS audit. While the bank deposits method is most effective and common methods the IRS uses for identifying unreported income, it has other methods that it uses in cash-heavy businesses. This case provides an example of the credit-card-to-cash ratio where the IRS combines third-party data, bank records, and a taxpayer’s own statements to build a reconstruction. This case also shows that the courts will uphold the IRS’s method as reasonable absent evidence in the record showing that the method is not reasonable and/or that some other method is more reasonable.

Watch Our Free On-Demand Webinar

In 40 minutes, we’ll teach you how to survive an IRS audit.

We’ll explain how the IRS conducts audits and how to manage and close the audit.  



Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews


The Classification algorithm is a supervised learning method that trains data to determine the category of future observations. This is why firstly, let us understand what is supervised learning.

What is Supervised Learning

Understanding Supervised Learning

Supervised learning develops a function to predict a defined label based on the input data.

The model in Supervised Learning learns by action. During training, the model examines which label is related to the given data and, as a result, can identify patterns between the data and particular labels.

Let us understand supervised learning with an example of Speech Recognition. It is an application where you train an algorithm with your voice. Virtual assistants such as Google Assistant and Siri, which recognize and respond to your voice, are the most well-known real-world supervised learning applications.

Supervised Learning might sort data into categories (a classification challenge) or predict a result (regression algorithms). This article will specifically address everything we need to know about classification in Machine Learning.

What is Classification in Machine Learning?

The process of recognizing, interpreting, and classifying objects or thoughts into various groups is known as classification. Machine learning models use a variety of algorithms to classify future datasets into appropriate and relevant categories with the help of already-categorized training datasets.

Classification in Machine Learning

In other words, classification is a type of “pattern recognition.” In this case, classification algorithms applied to training data detect the same pattern (same number sequences, words, etc.) in consecutive data sets.

Want to Become a Master in Machine Learning? Then visit here to Learn Machine Learning Training

Types of Classification Models

There are four primary classification tasks you could come across:

  • Binary Classification
  • Multi-Class Classification
  • Multi-Label Classification
  • Imbalanced Classification

Binary Classification

The term “binary classification” refers to tasks that can provide one of two class labels as an output. In general, one is regarded as the normal state, while the other is abnormal. The following examples can assist you in better comprehending them.

For example, for email spam detection, the normal condition is “not spam,” whereas the abnormal state is “spam.” “Likewise, Cancer not found” is the normal condition of an activity involving a medical test, whereas “cancer identified” is the abnormal state.

The normal state class is usually allocated the class label 0, whereas the abnormal state class is assigned the class label 1.

Some of the popular algorithms used for binary classification are:

  • Decision Trees
  • Logistic Regression
  • Support Vector Machine
  • k-Nearest Neighbors
  • Naive Bayes

Machine Learning Training

  • Master Your Craft
  • Lifetime LMS & Faculty Access
  • 24/7 online expert support
  • Real-world & Project Based Learning

Multi-Label Classification

We refer to multi-label classification tasks as those in which we need to assign two or more distinct class labels that can be predicted for each case. A simple example is photo classification, in which a single shot may contain many items, such as a puppy or an apple, and so on.  In this type of classification, you can predict many labels rather than just one.

The most common algorithms are:

  • Multi-label Random Forests
  • Multi-label Decision trees
  • Multi-label Gradient Boosting

Multi-Class Classification

Tasks that have two or more class labels are called multi-class classification.

The multi-class classification does not differentiate between normal and abnormal results. 

In some situations, the number of class labels might be rather big. For instance, a model may predict that a photograph belongs to one of thousands or tens of thousands of faces in a facial recognition system. Examples are classified into one of several known classes.

Some of the popular algorithms used for multi-class classification are:

  • Naive Bayes
  • k-Nearest Neighbors
  • Random Forest
  • Gradient Boosting
  • Decision Trees

Imbalanced Classification

Imbalanced Classification refers to tasks in which the number of items in each class is distributed unequally. In general, unbalanced classification problems are binary classification tasks in which most of the training dataset belongs to the normal class and just a small percentage to the abnormal class.

Learners in Classification Problem

There are two types of learners in a classification problem, namely:

  • Eager Learners
  • Lazy Learners

Eager Learners

Eager learning occurs when a machine learning algorithm constructs a model shortly after obtaining training data. It’s named eager because the first thing it does when it obtains the data set is, it creates the model. The training data is then forgotten. When new input data arrives, the model is used to evaluate it. The vast majority of machine learning algorithms are eager to learn.

Lazy Learners

Lazy learning, on the other hand, occurs when a machine learning algorithm does not develop a model immediately after receiving training data but instead waits until it is given input data to analyze. It’s named lazy because it waits until it’s absolutely essential to construct a model if it builds any at all. It only saves training data when it receives it. When the input data arrives, it uses the previously stored data to evaluate the output. Instead of learning a discriminative function from the training data, the lazy learning algorithm “memorizes” the training dataset. The eager learning algorithm, on the other hand, learns its model weights (parameters) during training.

Types of Machine learning Classification Algorithms

Classification algorithms use input training data in machine learning to predict the likelihood or probability that the following data will fall into one specified category. One of the most popular classifications used to sort emails into “spam” and “non-spam” categories, as employed by today’s leading email service providers.

They are two types of classification models, namely:

  • Linear Models
  • Non-linear Models

1. Linear Models

Support Vector Machine

Support Vector Machine

The support vector machine (SVM) is a frequently used machine learning technique for classification and regression problems. It is, however, mostly employed to tackle categorization difficulties. SVM’s main goal is to determine the best decision boundaries in an N-dimensional space that can classify data points, and the optimal decision boundary is known as the Hyperplane. The extreme vector is chosen by SVM to locate the hyperplane, and these vectors are referred to as support vectors.

Logistic Regression
Logistic Regression

In logistic regression, the sigmoid function returns the probability of a label. It is used widely when the classification problem is binary, for example, true or false, win or lose, positive or negative.

Logistic regression is used to determine the right fit between a dependent variable and a set of independent variables. Because it quantifies the factors that lead to categorization, it beats alternative binary classification algorithms like KNN.

Enroll in our Machine Learning With Python Training program today and elevate your skills!

HKR Trainings Logo

Subscribe to our YouTube channel to get new updates..!

2. Non-Linear Models

Decision Tree

Decision Tree

The classification model is developed using the decision tree algorithm as a tree structure. The data is then divided down into smaller structures and connected to an incremental decision tree to complete the process. The final output looks like a tree, complete with nodes and leaves. Using the training data, the rules are learned one by one, one by one. Every time a rule is learned, the tuples that cover the rules are removed. The technique is repeated on the training set until the termination point is reached.

The tree is built using a recursive top-down divide and conquer method. A leaf symbolizes a classification or decision, and a decision node will contain two or more branches. The root node of a decision tree is the highest node that corresponds to the best predictor, and the best thing about a decision tree is that it can handle both category and numerical data.

Kernel SVM

A kernel in SVM is a function that assists in problem resolution. They provide you shortcuts so you don’t have to complete hard calculations. Kernel is remarkable since it allows us to go to higher dimensions and do smooth calculations. It is possible to work with an infinite number of dimensions with kernels.

K-Nearest Neighbor

The K-Nearest Neighbor technique divides data into groups based on the distance between data points and is used for classification and prediction. The K-Nearest Neighbor algorithm implies that data points near together must be similar. Hence, the data point to be classed will be grouped with the closest cluster.

Naive Bayes

The classification algorithm Naive Bayes is based on the assumption that predictors in a dataset are independent. This implies that the features are independent of one another. For example, when given a banana, the classifier will notice that the fruit is yellow in color, rectangular in shape, and long and tapered. These characteristics will add to the likelihood of it becoming a banana in its own right and are not reliant on one another. Naive Bayes is based on the Bayes theorem, which is represented as:

P(A|B) = (P(A) P(B|A)) / P(B)

 Here:
         P(A | B) = how likely B happens
         P(A) = how likely A happens
         P(B) = how likely B happens
         P(B | A) = how likely B happens given that A happen

Stochastic Gradient Descent

It is an extremely effective and simple method for fitting linear models. If the sample data is vast, Stochastic Gradient Descent is beneficial. For classification, it provides a variety of loss functions and penalties.

The only benefit is the ease of implementation and efficiency. Still, stochastic gradient descent has several drawbacks, including the need for many hyper-parameters and sensitivity to feature scaling.

Random Forest

Random Forest

Random decision trees, also known as random forest, may be used for classification, regression, and other tasks. It works by building many decision trees during training and then outputs the class that is the individual trees’ mode, mean, or classification prediction.

A random forest (meta-estimator) fits several trees to different subsamples of data sets and averages the results to increase the model’s predicted accuracy. The sub-sample size is similar to the original input size; however, replacements are frequently used in the samples.

Artificial Neural Networks

Artificial Neural Networks

A neural network uses a model inspired by neurons and their connections in the brain to convert an input vector to an output vector. The model comprises layers of neurons coupled by weights that change the relative relevance of different inputs. Each neuron has an activation function that controls the cell’s output (as a function of its input vector multiplied by its weight vector). The output is calculated by applying the input vector to the network’s input layer, then computing each neuron’s outputs via the network (in a feed-forward fashion).

Machine Learning Training

Weekday / Weekend Batches

Conclusion

In this blog, we looked at what Supervised Learning is and its sub-branch Classification, some of the most widely used classification models, and how to predict their accuracy and see whether they are trained correctly. 

Related Articles:

Feature Selection Techniques In Machine Learning



Source link