Timing Is Everything: The Step Transaction Doctrine in GSS Holdings

I. Introduction

Machine learning is powerful. It can rapidly reveal and quantify deep insights while analyzing the merits of tax positions. It can help tax experts optimize tax planning strategies, identify potential weak points in tax plans, and quantify risks for clients.¹ While seasoned tax professionals have traditionally performed these functions without relying on machine learning and data science, the speed and versatility of artificial intelligence-powered tools to assess variations on particular facts and circumstances are starting to change the game.

In this article, we explore how tax experts can use machine learning tools to safely test and assess potential litigation strategies before deploying them at trial or on appeal. This can be especially useful for cases involving questions of law that turn on interrelated factors, such as the step transaction doctrine. By way of illustration, we put the recent GSS Holdings Court of Federal Claims decision under our machine learning microscope to see what we can see.²

The case is being appealed by the taxpayer, GSS Holdings, to the Federal Circuit. The trial court reached its decision on cross-motions for summary judgment, favoring the government. It applied the step transaction doctrine to step together two transactions, finding them to constitute one asset sale for tax purposes. The court concentrated its analysis on a two-day period leading up to and including the sale, rejecting the taxpayer’s argument that the analysis should begin many years earlier, when the relevant agreements were first negotiated. When following the trial court’s characterization of the relevant time period, Blue J predicts a government win with 65 percent confidence.

On appeal, GSS argues that the trial court did not apply the step transaction doctrine correctly. More specifically, it contends that the court used an inappropriate hybrid test in reaching its decision and did not consider the appropriate time frame.

Machine learning can help us assess how important timing is to the outcome of the case. Without passing judgment on the likelihood of the appellate court adopting GSS’s preferred view of timing, we can explore how successful GSS would be if the longer time period is used as the basis for the step transaction doctrine analysis on appeal.

Our analysis using Blue J’s step transaction doctrine machine learning model suggests that if GSS can convince the Federal Circuit to extend its analysis beyond December 2011, to include the times at which the earlier agreements were entered into, the step transaction doctrine would probably not apply to step together the transactions at issue and the taxpayer would prevail with 94 percent confidence.

II. Background

A. Step Transaction Doctrine

The step transaction doctrine is a judicial substance-over-form doctrine.³ The doctrine has been summarized as giving tax effect to the substance as opposed to the form of a transaction by ignoring (or collapsing) steps of an integrated transaction for tax purposes that separately are without substance or independent significance.⁴ As the trial court in this case acknowledged, there is no single, universal test. Courts have identified and used three different tests to determine if the step transaction doctrine ought to apply in any given set of circumstances.⁵ If the requirements of any of the three tests are met, then the step transaction doctrine will likely apply to combine multiple transactions into one for tax purposes.⁶ The three tests are the end result test, the interdependence test, and the binding commitment test:

1. End Result Test — This involves determining whether separate transactions are really parts of a single transaction that was intended from the outset to be taken for the purpose of reaching the ultimate result.⁷ It is the parties’ intended end result of the series of interrelated steps that controls the tax consequences of the whole.⁸

2. Interdependence Test — This focuses on the relationship between different steps instead of the end result. The test involves determining whether, on a reasonable interpretation of objective facts, the steps were so interdependent that the legal relations created by one transaction would have been fruitless without completion of all the steps.⁹

3. Binding Commitment Test — This involves determining whether there was a binding commitment to undertake a later step in a series of transactions.¹⁰

The government and the trial court focused on the end result test in GSS Holdings. GSS now argues that the court did not properly apply the appropriate principles.

B. Factual Background

GSS Holdings involves the activity of Liberty Street Funding LLC, a wholly owned subsidiary of GSS. For tax purposes, Liberty is a flow-through partnership in which GSS is a partner. Liberty is a commercial paper conduit. Commercial paper conduits like Liberty make longer-term investments funded by the issuance of short-term commercial paper.¹¹ To mitigate liquidity risks, commercial paper conduits frequently enter into liquidity asset purchase agreements (LAPAs) for the longer-term investments they purchase.¹² In exchange for a fee, the counterparty to a LAPA agrees to pay the commercial paper conduit a preset price for the investment that is the subject of the LAPA no matter the market value of the investment at the time the LAPA is exercised.¹³

In this case, Liberty invested in the Aaardvark IV note in 2006. Liberty then entered into a LAPA with the Bank of Nova Scotia (BNS) as counterparty. BNS also acted as Liberty’s administrator, controlling Liberty’s operations and absorbing much of the benefit and risk of its activities.¹⁴ Liberty was originally consolidated on the BNS balance sheet, but in 2007 adoption of new accounting rules by the Canadian regulator resulted in BNS seeking to remove Liberty from its balance sheet.

To achieve this result, Liberty entered into a first loss note (FLN) agreement with a third party, Reconnaissance. Reconnaissance loaned money to Liberty to be used to compensate Liberty’s LAPA counterparties if the counterparties incurred a loss as a result of paying a preset price for an investment that was greater than the market value at the time the LAPA was exercised.¹⁵ The interest on the loan was high (at one point reaching 32 percent) and for tax purposes Reconnaissance became Liberty’s partner as creditor on the note.¹⁶ The effect of the FLN arrangement was to offload some of the exposure associated with the LAPAs from BNS to Reconnaissance.

In 2011 BNS adopted international financial reporting standards that required BNS to again consolidate Liberty on its balance sheet. Therefore, there was no longer a need to have a third party (that is, Reconnaissance) bear the risks associated with the LAPAs. Accordingly, a BNS subsidiary, Scotiabank (Ireland), began the process of stepping into the position of Reconnaissance for the FLN. This process was finalized on December 29, 2011, with Scotiabank (Ireland) acquiring all the rights and obligations of the FLN from Reconnaissance. As a result, Scotiabank (Ireland) itself became Liberty’s partner for tax purposes. The next day, Liberty exercised the Aaardvark IV LAPA, thereby triggering the sale of Aaardvark IV to BNS (see Figure 1). Because BNS incurred a substantial loss when it paid Liberty the preset price of $220.6 million for the Aaardvark IV investment, Liberty’s FLN balance of $24 million was transferred to BNS (see Figure 2).

In 2011 Liberty reported a loss of $22.5 million as a result of the sale of Aaardvark IV. This loss flowed through to GSS. To generate this loss figure, Liberty had netted the basis of the Aaardvark IV investment that it listed as $244.6 million (that is, the $220.6 million preset value of the asset plus the $24 million FLN balance transferred to BNS) against the sales proceeds of $222.1 million (that is, the preset value of the asset under the LAPA plus about $1.5 million in insurance proceeds).

But in a 2013 amended return, GSS characterized the transaction differently, reporting the $22.5 million as an ordinary loss under section 165 (for the transfer of $24 million less the $1.5 million in insurance proceeds). Thus, GSS separated the payment out of the FLN balance from the Aaardvark IV sale. GSS claimed that the 2011 Liberty return erroneously combined the transfer payment with the sale of the asset. The IRS applied the step transaction doctrine to characterize the events as one single sale transaction as Liberty had done in 2011, so that the $22.5 million loss was a loss on the sale of business assets. Further, because the sale was to a related party, the IRS applied section 707(b)(1) to disallow the deduction of the loss.¹⁷

1. Trial court decision.

The tax refund case was decided by the Court of Federal Claims on cross-motions for summary judgment. GSS argued that the applicability of the step transaction doctrine should be determined based on the entirety of the facts surrounding the agreements, commencing with the purchase of Aaardvark IV and the corresponding LAPA, which was entered into in 2006. GSS argued that the doctrine did not apply because the LAPA and FLN were entered into for different and independent nontax purposes, and the FLN payment was not an intended or anticipated result at the outset of the series of transactions. Consequently, according to GSS, the FLN payment and the sale should be seen as distinct transactions for tax purposes and the loss from the FLN payment should be characterized as a deductible ordinary loss.

The government’s position was that the relevant transactions generating the loss were those that occurred at the end of December 2011. Therefore, it was appropriate to apply the step transaction doctrine and characterize the loss as derived from one transaction — the sale of the Aaardvark IV investment under the LAPA. Further, because it was a sale between related parties, section 707(b)(1) prohibited its deduction.

The court denied GSS’s motion for summary judgment but granted the government’s cross-motion. In doing so, the court focused on the taxpayer’s intention when the payment was made out of the FLN balance in 2011. According to the court, the payment of $24 million from the FLN balance to BNS had “no purpose other than to offset . . . some of the loss built into the Aaardvark purchase price.”¹⁸ The FLN was intended to be made in conjunction with a capital sale as a result of exercising the Aaardvark IV LAPA, and therefore the payment was part of the sale when it occurred. The court concluded that whether the two transactions were stepped together or simply analyzed based on their substance as a unified whole, the result would be the same.

2. Appeal.

GSS has appealed the trial court decision and at the time of this writing only the appellant’s opening brief has been filed. GSS argues that the trial court erred by not applying the step transaction doctrine test put forth by either party. Rather, the court allegedly employed its own hybrid test that borrowed elements from the economic substance test and the end result step transaction doctrine test to come to its conclusion. GSS argues that the trial court replaced the “timing rule for the step transaction doctrine with the timing rule for the economic substance doctrine.”¹⁹ GSS asks the Federal Circuit to apply the timing rule for the step transaction doctrine, with the outset of the series of transactions being the purchase of Aaardvark IV and negotiation of the LAPA in 2006, or at the latest 2007 when the FLN structure was put into place with Reconnaissance.²⁰ GSS framed the relevant question as “whether, in 2006 and 2007, the parties intended the ultimate result here,” and maintains it did not.²¹

III. Machine Learning Analysis

A. Blue J’s Prediction

Blue J’s step transaction doctrine predictor uses a machine learning model to predict whether a court will respect a series of steps or transactions as being separate for tax purposes with independent tax consequences, or as steps of a single taxable transaction. The predictor requires a user to respond to a series of questions about the context and the facts and circumstances of the scenario of interest. Each of the questions represents a factor that has been found to materially inform courts’ decisions about the application of the step transaction doctrine. The machine learning model draws on information from approximately 200 past decisions and has a prediction accuracy of approximately 90 percent.

When the predictor questions were answered based on the facts accepted by the trial court in GSS Holdings, Blue J predicted with 65 percent confidence that the step transaction doctrine would apply to step the Aaardvark IV sale and FLN balance payment together into one single sale transaction for tax purposes (the baseline prediction). Thus, Blue J correctly predicted the trial court outcome with 65 percent confidence. The prediction largely resulted from focusing the analysis on December 29 and 30, 2011, as the trial court did, rather than beginning the analysis in 2006, when the Aaardvark IV LAPA was entered into, which GSS argued was more appropriate.

B. Time-Frame-Dependent Factors

Notably, GSS now appeals on the basis that the trial court considered the wrong time frame and conflated the economic substance and step transaction doctrine analyses. GSS argues that the trial court should have assessed the whole series of transactions from the beginning, starting with negotiations between Liberty and BNS for the Aaardvark IV LAPA in 2006, or at the latest when the FLN was issued in 2007, rather than focusing on the final steps leading to the loss in isolation. To determine whether the taxpayer’s preferred approach would have affected the outcome if it were accepted, we focus on five particular questions within Blue J’s predictor that could be answered differently depending on the time frame considered for the analysis. These five factors are whether:

1. the parties contemplated the timing of all transactions before the first transaction;

2. there was anything in writing before completion of the first transaction that contemplated subsequent transactions;

3. there was a change in valuation of Aaardvark IV between transactions;

4. there were intervening events between transactions; and

5. the parties were certain of the cumulative outcome of the transactions.

We will address these time-frame-dependent factors in order.

1. Parties contemplated timing of transactions.

According to our data, if the parties contemplated the timing of all transactions at the outset, then the court is more likely to find that the step transaction doctrine applies to step multiple transactions into one for tax purposes. In fact, our data show that the doctrine was applied in 80 percent of cases in which parties were found to have contemplated the timing of all transactions at the start.

Because the trial court found that the first relevant transaction was the exercise of the Aaardvark IV LAPA in 2011, the parties likely would have contemplated the timing of the second transaction, the FLN balance payment that took place immediately after. But if GSS’s argument is accepted and the first transaction to be considered in the analysis was entering into the Aaardvark IV LAPA in 2006, then it is arguable that the parties did not contemplate the timing of the resulting transactions at the outset given that the FLN was not issued, or even necessary, at the time. Even if 2007 is taken as the starting point of the analysis (when the FLN was issued), GSS argues that the parties would not have contemplated when or even if the Aaardvark IV LAPA would be exercised in the future, if there would be a loss on the sale, or if there would be any FLN balance at the time to compensate for that loss, resulting in the FLN balance payment.²²

2. Documentary evidence contemplating transactions.

According to our data, having documentary evidence indicating that subsequent transactions were contemplated prior to the first transaction tends to support a court’s application of the doctrine.

Based on the trial court’s analysis, there were various agreements in writing before the exercise of the Aaardvark IV LAPA that arguably contemplated the sale itself and the payment out of the FLN balance at some point in time (for example, the original Aaardvark IV LAPA entered into in 2006, various LAPA renewals, the issuance of the FLN in 2007, and the agreement transferring the FLN from Reconnaissance to Scotiabank (Ireland) in 2011). But if the analysis begins with the Aaardvark IV LAPA in 2006, then there was nothing in writing that contemplated the subsequent FLN payment because the FLN was not yet issued. Notably, if 2007 is taken as the starting point, it is possible to argue that the already existing Aaardvark IV LAPA was a written document that contemplated the Aaardvark IV sale and the subsequent FLN agreement contemplated the potential loss payment out of the FLN balance at some point in the future. But this factor has negligible significance on the prediction in this particular fact pattern, as will be seen shortly.

3. Change in valuation of Aaardvark IV.

If the valuation of a transferred interest changes between transactions, then the court might be less likely to apply the doctrine, according to our data. Using the trial court’s analysis, the exercise of the Aaardvark IV LAPA and the FLN balance payment happened almost simultaneously and therefore the valuation of the Aaardvark IV investment would likely be materially unchanged. However, if entering into the Aaardvark IV LAPA in 2006 is taken as the first transaction, then the value of Aaardvark IV unequivocally decreased by 2011, which caused the loss triggering the payment from the FLN balance. We also presume there would be some change in valuation between the time that the FLN structure was put into place and when the Aaardvark IV LAPA was exercised in 2011. In fact, GSS claims that Liberty still expected to make a profit on Aaardvark IV at the time of the FLN in 2007.²³

4. Intervening events.

Our data suggest that if there are intervening events that alter the parties’ intended course of action between the transactions in a series, those might make it less likely that a court would apply the step transaction doctrine. For example, in Penrod the Tax Court held that the step transaction doctrine should not be applied to step together the transactions at issue, finding that the end result of the series was not intended by the parties at the outset but came about because of the subsequent souring of relationships between them.²⁴

Once again, given the short time frame considered by the trial court, there were no intervening events between the Aaardvark IV sale and FLN payment. However, based on GSS’s 2006 time frame the regulatory changes that prompted Liberty to adopt the FLN structure in 2007 could be viewed as an intervening event that affected the way the series of transactions unfolded. Even if 2007 is taken as the starting point for the analysis, there could be further intervening events to consider, including the accounting method changes in 2011 that led to Scotiabank (Ireland) taking over the FLN from Reconnaissance.

5. Certainty of outcome.

The parties’ certainty about the cumulative outcome of the transactions is an understandably strong factor in a court’s decision to apply the step transaction doctrine, according to our data. In GSS Holdings, the outcome of the sale (that is, the resulting loss and payment out of the FLN balance) and the tax consequences were likely certain to the parties at the end of December 2011, but not obvious in 2006 before the FLN agreement was entered into, or even in 2007 after the FLN was issued.

C. Using GSS’s Time Frame for Analysis

If the Federal Circuit agrees with GSS’s approach, beginning the analysis in 2006 when the Aaardvark IV LAPA was entered into, then each of the five factors discussed earlier likely weighs against the court’s application of the step transaction doctrine. Specifically, Blue J predicts with 94 percent confidence that by conducting the analysis using GSS’s time frame (beginning in 2006) and altering the answers to the time-frame-dependent questions, the step transaction doctrine would not be applied and the payment from the FLN balance and sale would be viewed as separately taxable transactions. This prediction remains the same if the analysis begins with the FLN in 2007, even though debatably there would be written documentation contemplating the subsequent transactions in general terms. The first three data rows in the table summarize the above factors when analyzed using the trial court’s versus GSS’s time frames (beginning in 2006 or 2007).

Scenario Testing Effect of Each Time-Frame-Dependent Factor on Prediction

	Parties Contemplated Timing of Transactions	Documents Contemplating Transactions	Change in Valuation of Aaardvark IV	Intervening Events	Certainty of Outcome	Blue J Prediction (confidence)
Baseline prediction	Yes	Yes	No	No	Yes	Single taxable transaction (65%)
GSS approach 2006	No	No	Yes	Yes	No	Multiple taxable transactions (94%)
GSS approach 2007	No	Yes	Yes	Yes	No	Multiple taxable transactions (94%)
Scenario 1	No	Yes	No	No	Yes	Multiple taxable transactions (84%)
Scenario 2	Yes	No	No	No	Yes	Multiple taxable transactions (55%)
Scenario 3	Yes	Yes	Yes	No	Yes	Single taxable transaction (54%)
Scenario 4	Yes	Yes	No	Yes	Yes	Multiple taxable transactions (61%)
Scenario 5	Yes	Yes	No	No	No	Multiple taxable transactions (81%)

To understand the relative importance of each of the five factors in Blue J’s prediction, we systematically tested the effect of changing each one from the baseline prediction in isolation to reflect the situation as it would exist if the analysis were conducted using 2006 as the starting point (Scenarios 1-5 in the table). Our predictions illustrate that any of the following four factors alone would have been sufficient to alter Blue J’s baseline prediction, and result in the court’s finding that the series of transactions represent multiple taxable transactions with varying levels of confidence:

1. the parties did not contemplate the timing of transactions at the outset (Scenario 1);

2. there was no documentary evidence indicating that before the first transaction subsequent transactions were contemplated (Scenario 2);

3. there were intervening events that altered the parties’ intended course of action between the transactions (Scenario 4); or

4. the parties were uncertain that the cumulative outcome of the transactions would occur (Scenario 5).

The strongest factors in this case appear to be whether the parties contemplated the timing of the transactions at the outset and whether they were certain that the cumulative outcome of the transactions would occur. Although a change in the valuation of Aaardvark IV between transactions also weighed against application of the doctrine, the factor was not strong enough alone to change the baseline prediction. Rather, it simply reduced the confidence in the prediction from 65 percent to 54 percent (Scenario 3).

In reality, those changes to the time-frame-dependent factors would likely not occur in isolation. If GSS’s 2006 starting point were adopted by the court, all five factors would likely change together, making it highly likely that a court would not apply the step transaction doctrine to the facts of the case, a result Blue J identifies with 94 percent confidence.²⁵ Substantially the same prediction results if 2007 is taken as the starting point for the analysis.

IV. Conclusion

Using GSS Holdings, we demonstrate how Blue J’s machine learning technology can be used to test the efficacy of a litigation strategy before it’s employed in court. In this case we tested GSS’s strategy of beginning the step transaction doctrine analysis at the time that the Aaardvark IV LAPA was established, rather than focusing on the last two steps that led to the loss, as originally argued by the government. If the court took the approach proposed by GSS, Blue J predicts with 94 percent confidence that the court would not have applied the doctrine to step together the sale and FLN balance payment for tax purposes. But because the trial court rejected GSS’s approach, it held that the doctrine should apply — a result that Blue J predicted with 65 percent confidence. The Federal Circuit is likely to have the final say on the issue, but if it agrees that GSS’s approach is correct in law then Blue J predicts that there is a high likelihood that the trial court’s decision will be reversed.