In Part 1 of “Turning Standards into Rules,” we began to explore how machine learning algorithms can be used to predict the outcomes of common tax law questions. This is a timely investigation. Algorithms have consistently been shown to outperform human beings in making probabilistic predictions, and recent advances in machine learning have opened up new ways of making algorithmic prediction even more accurate. Researchers now use machine learning algorithms for everything from filtering spam messages to making medical diagnoses. For example, a recent study has shown that algorithms could lead to a significant improvement in outcomes associated with bail determinations. (J. Kleinberg, et al., “Human Decisions and Machine Predictions” 133 (1) Quarterly Journal of Economics 237-293 (2018)). These types of algorithms can also be used to make predictions about tax outcomes.
Like many legal questions, the question of whether a financial instrument is more debt-like or more equity-like can be thought of as a binary classification problem. If researchers can harness enough information from the body of legal decisions, these tools can provide an accurate prediction for how the courts might rule in a new debt vs. equity scenario.
This is not merely a theoretical possibility: my colleagues at the University of Toronto and I have been able to produce accurate predictions of how U.S. courts characterize new financial instruments based on how similar cases have been resolved in the past. We extract information from hundreds of legal opinions and use various machine-learning algorithms to analyze patterns in the data. These patterns highlight the combinations of facts that are most strongly associated with the eventual outcomes of the cases. Following extensive testing and calibration, we can predict the likelihood of outcomes in cases that the system has not previously seen. We report the confidence of our predictions as a percentage based on the probabilistic likelihood of the outcome.
As a result, our system is able to produce accurate predictions on a range of tax questions, including whether courts would characterize new financial instruments as debt or equity. The system also allows us to observe how the probability of a given outcome changes when we alter the fact pattern.
Financial Risk Factors in Focus
In a typical debt vs. equity case, an individual or corporation (the “issuer”) contributes money or property to a corporation. As the Tax Court put it in Litton Business Systems, Inc. v. Commissioner, the legal issue in these cases is: “Was there a genuine intention to create a debt, with a reasonable expectation of repayment, and did that intention comport with the economic reality of creating a debtor-creditor relationship?” In other words, should the funds be considered a loan or an equity contribution?
So what can we learn from a machine learning system trained on a large body of debt vs. equity cases?
One key finding is that the financial risk involved in a transaction has a particularly strong impact on a court’s characterization of the interest as debt or equity. A high degree of financial risk is significant because it indicates little to no expectation of repayment, which weighs in favor of an equity finding. Conversely, a low degree of financial risk indicates a strong expectation of repayment, which is more consistent with bona fide debt. For the purposes of this article, we will explore the impact of three factors that specifically relate to risk, namely:
1. Thinness of capitalization
2. Ability to obtain funding from an arm’s-length source
3. The risk involved in making the advances
Let’s see how altering the financial risk factors in three recent cases affects the algorithm’s predicted results.
The Effect of Risk in a Case with Many Markers of Equity
VHC, Inc. v. Commissioner involved a family-controlled company attempting to claim bad debt deductions for advances it made to its shareholder (RH) and his related companies.
All of our financial risk factors are present in this case. VHC continued making advances despite knowing that they would not be repaid, which indicates thin capitalization. RH’s companies were not profitable in the year leading up to the transactions, and there was insufficient cash flow to service the purported debt. Not surprisingly, RH and his related companies had not been able to obtain loans from third-party sources.
The transaction was subject to scrutiny by the IRS because of the close relationship between the parties, and there were many equity markers in addition to the high-risk profile of RH. Although the obligations had fixed maturity dates, they were effectively meaningless. VHC was subordinated to other creditors and had no right to enforce repayment. VHC had advanced funds to meet initial operating expenses knowing repayment was entirely dependent on the success of RH and his related companies.
If we input the facts of this case into a machine learning algorithm trained on the details of a large number of debt vs. equity cases, it predicts a finding of equity with over 95 percent confidence—which is precisely what the court found.
But how might the outcome have changed if the transaction’s risk profile had been different? If we alter some of the facts to indicate low risk, we receive an outcome of equity with only 84 percent confidence. In other words, if the financial risk factors were not present, the result would be 11 percent less likely to be equity.
The Effect of Risk in a Case with Fewer Markers of Equity
Rogers v. Commissioner discusses whether the taxpayer’s transfer of a property to her wholly-owned corporation was a capital contribution.
Some of our risk factors were present, but not all. The wholly-owned corporation was thinly capitalized: it was newly incorporated, with no other assets except for an initial capital contribution of $50,000. Further, an arm’s-length lender would not have extended credit to the taxpayer.
There were also other equity markers in addition to the risk profile of the corporation: the corporation did not issue a note or debt instrument; there was no repayment schedule or fixed maturity date; the parties were related; the purported debt was subordinated to a third-party creditor; and finally, repayment was expected to be dependent on earnings. Other factors, however, favored a debt classification: the obligation entailed a fixed sum with a specific interest rate, and the issuer made a partial repayment of the principal.
In this case, the machine learning algorithm correctly predicts an outcome of equity with 92 percent confidence. If we alter the facts to indicate low risk, however, the confidence level falls to 54 percent. In other words, if the financial risk factors were not present, this would be a borderline case.
The Effect of Risk in a Case with Few Markers of Equity
Let’s try adding financial risk factors to a scenario with few markers of equity.
Owens v. Commissioner involved an entrepreneur whose business was lending money to distressed companies. Although distressed companies are inherently risky, the court found that only one financial risk factor was present: Owens had made previous advances that were not repaid. Thin capitalization was not an issue because when Owens made his first loans, they were secured. Even though subsequent loans might not have been secured, they were used to protect Owens’ initial advances. The evidence also showed that arm’s-length lenders would have extended credit on similar terms.
There were few other equity markers in Owens. Owens had subordinated his advances to other creditors, but most other facts in the case weighed toward debt: the purported loans were evidenced by promissory notes that had maturity dates; Owens had a definite right to enforce repayment not dependent on earnings; and he did not receive any rights to participate in management.
In this case, the algorithm correctly predicts the debt outcome with 72 percent confidence. But if all our financial risk factors were fully present, we would receive a predicted outcome of equity with 85 percent confidence. This result shows how dramatically financial risk factors can alter the outcome of a case.
Table A summarizes the risk profile present in each of the cases above and the impact of altering the risk factors. VHC, with a high-risk profile, already has so many indications of equity that changing the risk profile of the issuer from high to low makes less of an impact than for Rogers, which has a medium risk profile. The impact of adding financial risk factors to a case with few indications of equity is particularly significant: changing the risk profile from low to high for Owens can flip the outcome of the case completely.
Table A: Sensitivity Analysis of Risk Profiles on Predicted Outcomes
Clearly, financial risk factors can have a significant influence on the characterization of interest in a corporation as debt or equity. The impact is particularly strong where financial risk factors are added to a scenario that has few other indications of equity, as in Owens. By using machine learning algorithms to test different scenarios, we can get a much clearer sense of how the courts actually weigh different factors in complex cases such as these.
Our next instalment in this series on using data analytics and machine learning in law will examine the application of predictive algorithms to worker classification.
Benjamin Alarie holds the Osler Chair in Business Law at the University of Toronto Faculty of Law and is the CEO of Blue J Legal.