How Machine Learning Predicts Which Tax Returns Will Be Audited

The IRS audited about 0.4% of individual tax returns in 2024. That low overall rate masks significant variation. Returns claiming the Earned Income Tax Credit were audited at 1.1%. Returns with income over $1 million were audited at 2.6%. Returns with Schedule C income over $100,000 and high deduction ratios were audited at rates as high as 4-5%.

Understanding these patterns is not just academic. Tax preparers who can assess audit risk for a specific return provide a tangible advisory service. Machine learning models trained on IRS audit data and publicly available audit statistics can estimate the relative audit probability of a particular return based on its characteristics.

What the Models Look At

IRS audit selection uses a scoring system called the Discriminant Information Function (DIF). The exact formula is secret, but decades of analysis by tax professionals and researchers have identified the factors that most influence audit selection:

• Income level and type (self-employment income triggers more audits than W-2 income)
• Deduction ratios relative to income (Schedule A deductions that are high relative to AGI)
• Home office deduction claims (particularly when combined with other red flags)
• Cash-intensive businesses (restaurants, retail, construction, and service businesses)
• Round numbers on schedules that suggest estimation rather than actual recordkeeping
• Large charitable contributions relative to income
• Claimed losses that offset other income, particularly rental real estate losses
• Missing or inconsistent information (1099 income that does not appear on the return)
• Prior audit history

Machine learning models incorporate these known factors plus additional variables that research has shown to correlate with audit selection. The models are trained on datasets that combine IRS Transactional Records Access Clearinghouse (TRAC) data, published audit statistics, and anonymized return data where audit outcomes are known.

How the Prediction Works

A typical ML audit risk model takes the key characteristics of a tax return as inputs and produces a risk score, usually expressed as a multiple of the baseline audit rate. A score of 1.0 means the return has average audit risk. A score of 3.5 means the return is estimated to be 3.5 times more likely to be audited than the average return in the same income bracket.

For example, a return with $180,000 in Schedule C income, $42,000 in deductions (a 23% deduction ratio), a home office deduction, and $15,000 in charitable contributions might receive a risk score of 2.8. The model is saying: based on returns with similar characteristics, this one has roughly 2.8 times the baseline audit probability for its income bracket. In practical terms, that might mean a 1.5-2% audit probability instead of the 0.5% baseline.

The models are not predicting whether a specific return will be audited. No model can do that with the information available. They are estimating relative risk, which is still useful for advisory purposes.

How Tax Preparers Use This Information

The practical application is in tax planning and return preparation. When a preparer knows that a particular return has elevated audit risk, they can take several steps:

Documentation emphasis: Make sure the client has solid documentation for the items driving the risk score. If high charitable contributions are a factor, ensure appraisals are complete and donation receipts are organized. If the home office deduction is contributing, verify that the measurements are accurate and the exclusive-use test is met.

Position evaluation: Some return positions are technically correct but aggressive enough to attract attention. Knowing the overall risk profile of the return helps the preparer decide whether an aggressive position on one item is worth the incremental audit risk when combined with other risk factors.

Client communication: Clients appreciate knowing their audit risk profile. A conversation that says, "Your return has some characteristics that historically correlate with higher audit rates, so I want to make sure your documentation for X, Y, and Z is thorough," provides genuine value and sets appropriate expectations.

Accuracy and Limitations

These models are imperfect, and anyone using them should understand the limitations. The IRS DIF scoring formula is proprietary and changes periodically. The models are trained on historical data that may not reflect current enforcement priorities. The IRS also uses targeted campaigns that focus on specific issues (cryptocurrency reporting, for example) that historical models may not capture.

That said, research published in academic tax journals suggests that well-constructed models can identify high-risk returns with reasonable accuracy. A 2023 study found that returns scoring in the top 10% by the model's risk score were audited at 4.2 times the rate of returns in the bottom 90%. The model was not perfect at predicting individual outcomes, but it successfully stratified returns by relative risk.

Building This Into Firm Workflows

Accounting firms can integrate audit risk scoring into their review workflow. After a return is prepared, the risk model runs automatically and produces a score that the reviewer sees alongside the return. High-risk returns get additional review attention, and the preparer is prompted to document support for the items driving the risk score.

Some firms use the risk score to prioritize quality review during busy season. When you have 200 returns waiting for review and limited reviewer capacity, focusing detailed review on the 30-40 returns with the highest risk scores is a rational allocation of time. Lower-risk returns still get reviewed, but with a lighter touch focused on mathematical accuracy rather than position-by-position evaluation.

The firms getting the most value from audit risk scoring are those that use it as an advisory tool with clients rather than just an internal risk management tool. When you can tell a client, "Here is what makes your return more likely to draw attention, and here is what we have done to prepare for that," you are providing the kind of proactive, personalized advice that differentiates a trusted advisor from a form filler.