Benjamin Alarie is the Osler Chair in Business Law at the University of Toronto and the CEO of Blue J Legal Inc. Bettina Xue Griffin and Christopher Yan are senior legal research associates at Blue J Legal.
In this inaugural installment of Blue J Predicts, the authors use machine-learning models to anticipate and analyze the possible outcome of the IRS’s appeal in Cross Refined Coal. Future installments will focus on other pending tax cases.
Copyright 2021 Benjamin Alarie, Bettina Xue Griffin, and
All rights reserved.
Technology now makes it possible for tax practitioners to accurately anticipate the outcome of tax disputes. This article demonstrates how using machine-learning algorithms to study legal source data, such as past court decisions and revenue rulings, can provide novel and valuable insights into how tax disputes may be resolved.
The dispute we focus on in this inaugural installment of Blue J Predicts is Cross Refined Coal.1 The D.C. Circuit heard oral arguments April 12 in an appeal of the Tax Court’s decision, which found in favor of the taxpayer. The IRS appealed on the basis that the Tax Court incorrectly found the existence of a partnership for federal tax purposes.
Blue J predicts with 90 percent confidence that the appeal will be dismissed on the issue of whether a partnership exists. This article examines how a machine-learning model can generate this prediction based on the facts of the case, and it identifies the strengths and weaknesses of the parties’ positions using machine learning.
The dispute involves the denial of credits claimed by the taxpayer, Cross Refined Coal LLC, a producer and seller of refined coal. The IRS made adjustments to partnership items of the taxpayer for the 2011 and 2012 tax years on the basis that Cross lacked a nontax business purpose and thus was not operating as a bona fide partnership.
Cross was originally formed by AJG Coal Inc. (AJGC). AJGC invests in and operates coal refineries through limited liability companies such as Cross. AJGC successfully sought two investors to invest in Cross: (1) USA Refined Coal LLC (USARC), a subsidiary of FMR LLC (Fidelity), and (2) Schneider Electric Investments 2 Inc. Through a series of purchase agreements, Cross was eventually owned by the three entities: Fidelity (51 percent indirectly through USARC), AJGC (24 percent), and Schneider (25 percent).
The Cross venture, like many of AJGC’s ventures, operated at a pretax loss. Its business model contemplated purchasing raw coal from a utility plant, refining the coal, and then selling the refined coal back to the utility plant at a discounted price. The discounted rate would give the utility plant an incentive to purchase refined coal rather than use raw coal. This process entitled Cross to section 45(e)(8) tax credits for the production and sale of refined coal, which allowed the taxpayer to generate a post-tax profit. Without the tax credits, there was no economic reason for the buy-high-and-sell-low model.
USARC and Schneider made initial contributions of $4 million and $1.18 million, respectively, for the purchase of their interests in Cross. They made additional contributions for operating expenses throughout 2010 to 2013 (USARC’s $22 million and Schneider’s $10.5 million). Fidelity’s agreement with AJGC contained a liquidated damages clause that limited Fidelity’s risk of loss. When specified trigger events occurred, Fidelity had the right to exit from Cross and would receive from AJGC a pro rata portion of its initial $4 million contribution. Fidelity was not entitled to a return of its due diligence costs or its monthly contribution to the operating costs.
III. Blue J Predicts
Blue J predicts with 90 percent confidence that the D.C. Circuit will find that Cross operated as a partnership for federal tax purposes.
Our prediction is based on a data-driven analysis using a machine-learning model trained on the facts and results of 169 tax decisions between 1949 and 2021 that directly address whether a de facto partnership exists. Although we were able to locate a total of 203 decisions that addressed that issue in the 1949-2021 period, 34 of them were excluded from the data set principally because there was inadequate information about the underlying facts of the case or the dispute was decided by the court on another basis (that is, the court didn’t resolve whether there was a de facto partnership).
Blue J’s machine-learning models are constructed through a human-driven research and development process.2 The model used in this analysis has been generated from a machine-learning process known as supervised learning, in which machine-learning models are trained using labeled data. The result is a model that will generate predictions in hypothesized new scenarios based on analysis of the outcomes of previously decided cases. Each factor in a hypothesized scenario is weighed differently depending on the other facts. The model can detect when a given factor should be given more significance based on the facts at hand, and it provides an overall prediction coupled with a confidence level.
Our experience suggests that this supervised machine-learning modeling approach can be quite accurate. Our diagnostics reveal that the model for de facto partnership determinations is 91.4 percent accurate at predicting the outcomes in federal de facto partnership disputes since 1949.
Let’s turn to what this machine-learning model can reveal about the pending appeal in Cross Refined Coal.
IV. The Tax Court Decision
The Tax Court found that Cross was a bona fide partnership for federal tax purposes. It further found that each of the three partners had contributed to the partnership and shared in the profits and agreed to bear the risk of loss. The court applied the eight-factor test set out in Luna3 to determine whether the parties intended to, and did in fact, join together in the conduct of an undertaking or enterprise.
The IRS’s core argument before the Tax Court was that an unprofitable pretax venture cannot be found to have a real business purpose such that it can be classified as a partnership. It argued that Cross was a venture that existed solely to monetize tax credits rather than pursue a viable business purpose. However, the IRS raised this point within the framework of the eight Luna factors. The Luna factors do not squarely address the issue of the bona fides of a partnership arrangement; rather, the decision in Luna assessed whether the conduct of the parties amounted to a de facto partnership. The Luna decision does not assess the business purpose of a venture; it looks to the conduct of the parties to determine if, in reality, the parties intended to share in the profits and risk of loss as a partnership.
The Luna factors are:
the agreement of the parties and their conduct in executing its terms;
the contributions, if any, that each party has made to the venture;
the parties’ control over income and capital and the right of each to make withdrawals;
whether each party was a principal and co-proprietor, sharing a mutual proprietary interest in the net profits and having an obligation to share losses, or whether one party was the agent or employee of the other, receiving for his services contingent compensation in the form of a percentage of income;
whether business was conducted in the joint names of the parties;
whether the parties filed federal partnership returns or otherwise represented to the IRS or to persons with whom they dealt that they were joint venturers;
whether separate books of account were maintained for the venture; and
whether the parties exercised mutual control over and assumed mutual responsibilities for the enterprise.
The IRS asserted that two key Luna factors were insufficiently met for Cross to be a true partnership: (1) the capital contributions of each partner were deficient; and (2) there was no meaningful sharing in the profits and losses. The IRS argued that the capital contributions provided by Fidelity and Schneider were de minimis relative to the anticipated 10-year profits from the tax credits, and that the liquidated damages clause allowed Fidelity to claim back $2.5 million of its initial $4 million contribution such that no meaningful contributions were actually at risk. The Tax Court rejected those arguments on the basis that the liquidated damages clause did not eliminate the economic reality of the contribution and that the partners continued to contribute $1.9 million to the operating expenses, even when no credits were being generated during a nine-month facility shutdown.
In terms of profits, the IRS maintained that there was no sharing of profits because the business model contemplated that the sale of refined coal always results in pretax losses and that tax credits should not be included in the assessment of profits. The Tax Court rejected that argument. It concluded that because Congress has offered a tax incentive to encourage business activity that would otherwise not be viable, it would be inconsistent with the purpose of the credit to require the parties to share only in pretax profits in order to be considered a partnership.
On the risk of loss, the IRS’s position was twofold: First, the liquidated damages clause insulated Fidelity from loss; and second, there was no risk of loss for Fidelity and Schneider because “everything essential to the refined coal operation was already in place” from AJGC. The court rejected this argument as well. It reasoned that because the section 45 credit is a production credit, which is not earned until the refined coal is produced and sold, parties are exposed to the risk inherent in the production and sale of refined coal (including facility closures and environmental and regulatory risk), many of which did in fact manifest in this case.
V. The IRS’s Position on Appeal
On appeal, the IRS argued that the lion’s share of the capital contributions were made by AJGC and that the Tax Court grossly overstated the actual risk borne by Fidelity and Schneider. The IRS pointed to the fact that the Cross facilities were already operational before Fidelity and Schneider invested, and that the months in which the facility could not operate were ones with fewer expenses such that there was minimal risk of loss for Fidelity and Schneider. Ultimately, the IRS again argued that the risk of loss and initial capital contributions were minor and disproportionate to the anticipated high returns associated with the tax credits. The IRS maintained that a nontax upside for the partners is required for a real partnership to exist.4
VI. The Taxpayer’s Position on Appeal
First, the taxpayer argued that the risk of loss did in fact materialize as demonstrated by the fact that Fidelity lost $1.7 million of its initial capital contribution to Cross. Second, the taxpayer asserted that when the parties entered into the agreement, there was uncertainty about whether tax credits would ever be generated, given the demand as well as the regulatory and environmental risks. Third, the taxpayer contended that the partners’ contributions were not de minimis because the capital contributions corresponded exactly to the cost of producing the refined coal. Finally, the taxpayer sought to distinguish the IRS’s precedents on the basis that Cross was not a facade or a sham business. Cross had employees, conducted transactions, and produced and sold refined coal.
VII. Assessing the Parties’ Preferred Facts
As with any multifaceted test, parties will disagree about how specific facts should be characterized, which factors are actually satisfied, and the significance of each factor. Disagreements about the characterization of facts highlight the value of using machine-learning tools to discern the substantive implications of different potential and conflicting characterizations. With Blue J’s machine-learning algorithm, we can assess the full range of plausible scenarios by changing the facts and factors that are in dispute to test the likely outcome for each scenario.
Based on the version of facts accepted by the Tax Court, Blue J predicts with 90 percent confidence that a partnership will be found to exist. This indicates that if there is no change to the characterization of the facts accepted by the Tax Court, there is a high probability that the D.C. Circuit will dismiss the IRS’s appeal.
But what would happen if the appellate panel were to disagree with some of the Tax Court’s characterizations of the factors? Although the Tax Court’s findings of fact are unchallenged, that doesn’t mean that the D.C. Circuit will arrive at the same conclusions on the characterization of the Luna factors based on those underlying findings. We can test the likelihood of Blue J’s prediction by adjusting the scenario to see what would happen if the appellate court decided differently on each of the factual characterizations in dispute.
Each of the eight Luna factors can be further divided into many smaller objective factors. After canvassing the case law, we find that at least 30 different factors can materially influence a court’s analysis of whether a partnership exists. Each of those 30 factors is given different weight by the machine-learning model, depending on the particular facts of a given scenario.
Let’s begin our work with the model by stipulating that the judges on the appellate panel accept the strongest plausible position for the IRS — that is, they completely agree with the agency’s characterization of the facts. In the IRS’s best-case scenario, would the machine-learning model predict success for the agency? Our analysis suggests that yes, the result would likely switch to “no partnership,” with 69 percent confidence.
We generated that prediction by modifying the facts to better align with the IRS’s interpretation. To do this, we assumed, arguendo, that:
the parties didn’t divide the profits of the partnership because there were no pretax profits to divide (based on the IRS’s position that tax credits should not be included in the analysis of whether there was an intent to share profits);
the parties didn’t divide the losses proportionally (based on the IRS’s position that Fidelity and Schneider had very little risk of loss compared with AJGC);
the contributions of some of the partners were neither necessary nor valuable to the partnership (based on the IRS’s position that the capital contributions were de minimis); and
the capital contributions of some of the partners resulted in a debt obligation rather than equity (based on the IRS’s position that the contributions were fixed debt).
With those modifications to the characterization of the facts by the Tax Court, Blue J’s machine-learning algorithm predicts with 69 percent confidence that the IRS would prevail. Although this shows that the IRS’s most favorable view of the case would likely lead to the agency’s preferred outcome, it is reasonable to question how likely it is that those characterizations would be adopted by the D.C. Circuit and whether the IRS must succeed on all those characterizations to prevail.
VIII. Stress-Testing the IRS’s View of the Facts
To flip the predicted result from 90 percent in favor of the taxpayer to 69 percent in favor of the IRS, we gave the IRS the benefit of having all the disputed factors characterized in the most beneficial way possible. But what would happen if the IRS doesn’t succeed on those four disputed factors? Is any one of them alone determinative of the outcome?
To answer those questions, we assess the effect of changing each of the four key factors individually, beginning with the premise that the IRS will succeed on the other three factors.
A. Stress Test 1: Pretax Profits
If the IRS succeeds on all of its characterizations except for whether profits must be assessed pretax, the prediction switches to a finding that a partnership does exist, with 80 percent confidence. This represents a change of -49 percent from the IRS’s best-case scenario. It thus indicates that the IRS losing on this issue alone would result in a likely win for the taxpayer.
The importance of this factor aligns with (1) how extensively this issue was discussed in the Tax Court’s decision and (2) the significant amount of time the D.C. Circuit spent questioning the parties on this issue. The IRS’s position hinges on whether the section 45 tax credits should be included or excluded from the analysis of whether the parties agreed to share in the profits and whether the parties did in fact share in the profits. If the appellate court decides in favor of including the tax credits, that would necessitate a finding that the parties agreed to share in the profits, because the tax credits were the only profits to share. By running a Blue J analysis on the scenario, we are able to demonstrate that the threshold issue of whether section 45 tax credits should be included is alone likely to be determinative of the outcome of the appeal. In other words, Blue J’s algorithm predicts that this is not an issue that the IRS can afford to lose and that it is the most important of the four contentious issues.
B. Stress Test 2: Risk of Loss
If the IRS succeeds on all of its characterizations except whether the parties bore the risk of loss proportionally, the model still predicts there is no partnership, but with 59 percent confidence (-10 percent). If the D.C. Circuit finds that the parties proportionally shared in the losses of the venture, the IRS is still likely to succeed on appeal if it also succeeds on the characterization of the other three factors. In other words, Blue J’s algorithm predicts that losing on this issue would make this a borderline case for the IRS.
C. Stress Test 3: Contributions
If the IRS succeeds on all of its characterizations except whether the partners’ contributions were de minimis, the model predicts that there is a partnership, with 57 percent confidence (-26 percent). If the D.C. Circuit finds that the parties did meaningfully contribute to the venture beyond a de minimis contribution, the taxpayer will likely succeed despite the IRS prevailing on the characterization of the other three factors. In other words, Blue J’s algorithm predicts that losing on this issue alone would put the IRS in a risky position of being more likely than not to lose on appeal.
D. Stress Test 4: Debt or Equity
If the IRS succeeds on all of its characterizations except whether the contributions should be characterized as debt or equity, the model still predicts there is no partnership, with 59 percent confidence (-10 percent). If the D.C. Circuit finds that some parties contributed capital in the form of equity rather than debt, the IRS is still more likely than not to succeed on appeal if it also succeeds on the characterization of the other three factors. In other words, Blue J’s algorithm predicts that losing on this issue would make this a borderline case for the IRS.
IX. Machine-Learning Model Takeaways
The four stress tests indicate that there are two issues the IRS cannot afford to lose: (1) whether the profits must be assessed pretax and (2) whether the contributions of the partners were de minimis. Losing on either one will likely result in the taxpayer succeeding on appeal. Those two issues are not equally important; the first issue affects the probability of the outcome by almost twice as much as the second (a change of 49 percent as opposed to 26 percent).
Using a machine-learning model provides quantifiable insights into situations like the dispute in Cross Refined Coal. Several factors in the case favor a partnership characterization; for example, the fact that the parties shared management and operational responsibilities. However, there are also contentious characterization questions involving several elements of the eight-part Luna test for de facto partnerships. Using a machine-learning model, we are able to identify, rank, and quantify the importance of the disputed characterizations and to demonstrate, based on the previously decided case law, that it is more likely than not that the IRS will be unsuccessful on appeal.