Tax-Loss Harvesting: An Individual Investor’s Perspective

Abstract

In the tax-loss harvesting literature, a typical investor is assumed to have an unlimited supply of offsetting capital gains and can earn annualized tax alpha on the order of 100 bps. Using boosted regression tree analysis and nationally representative investor-level data, we quantified how investor-characteristic and return environment differences yield significant heterogeneity in expected tax-loss harvesting benefits. Overall, investor profiles drive roughly 60% of the variation in tax-loss harvesting outcomes. Our findings demonstrate that investors (and their advisers) can better target who might (and might not) benefit from various tax-loss harvesting strategies based on individual profile differences.

Declaration of Interests

Disclosure: The authors report no conflicts of interest.

Acknowledgments

We are grateful for the comments and suggestions of Andy Clarke, Dhagash Mehta, Alberto Rossi, and the editors of the journal (Stephen Brown, Daniel Giamouridis, and William Goetzmann).

Editor’s Note

This article was externally reviewed using our double-blind peer-review process. When the article was accepted for publication, the authors thanked the reviewers in their acknowledgments. Terence Burnham and one anonymous reviewer were the reviewers for this article.

Submitted 16 February 2021

Accepted 29 July 2021 by William N. Goetzmann

Notes on Risk

Tax-loss harvesting involves certain risks, including, among others, the risk that the new investment could perform worse than the original investment and that transaction costs could offset the tax benefit. There may also be unintended tax implications. We recommend that you consult a tax advisor before taking action.

All investing is subject to risk, including possible loss of principal. Past performance is no guarantee of future results.

Notes

¹ See Figure 1 and related discussions for the range of TLH outcomes.

² See Constantinides (1983), who introduced the idea of tax-loss harvesting to the academic literature.

³ See Appendix A for a derivation of Equation 1(1) $Valu{e}_{TLH}=\text{min}\left(L,LOI\right)\left({\tau }_{t_1}-{\tau }_{t_2}\right)+\text{min}\left(L,LOI\right){\tau }_{t_1}R\left(1-{\tau }_{t_2}\right),$(1) .

⁴ The effective marginal capital gains tax rate at time t₂ may be positive if generating proceeds for consumption purposes or zero if there is no further tax liability (as in the cases of charitable donations and stepped-up basis at death). Note that if $R>0$, then ${\tau }_{t_2}\ge 0$.

⁵ Clearly, investors facing an expected negative return from TLH would likely choose not to harvest, which again demonstrates the importance of heterogeneity in investor characteristics.

⁶ Gain management can be thought of as a levered TLH, where the “leverage” comes at the expense of realizing long-term capital gains and tax liability (albeit at a lower effective tax rate). The amount of short-term capital gains required to derive benefits from gain management is even greater than it is in the case of TLH (Stein et al. 2008). Since our article will show that only a minority of investors may reliably assume the typical level of capital gains implied in the extant literature, it is logical to assume that viable gain management is likely relevant for an even smaller minority of individual investors.

⁷ The United States allows individual investors to offset up to $3,000 of ordinary income in any tax year if realized losses exceed gains. In terms of Equation 1(1) $Valu{e}_{TLH}=\text{min}\left(L,LOI\right)\left({\tau }_{t_1}-{\tau }_{t_2}\right)+\text{min}\left(L,LOI\right){\tau }_{t_1}R\left(1-{\tau }_{t_2}\right),$(1) , this provision increases the amount of loss-offsetting income for which tax savings can be generated (i.e., higher $LOI$) and also affects the effective marginal tax rate used by potentially combining different income, short-term capital gains, and long-term capital gains tax rates that the losses may offset. The United Kingdom allows an annual capital gains exemption (currently £12,300 for the 2020–21 tax year) before realized gains become taxable (i.e., $LOI$ for which tax savings are generated in Equation 1(1) $Valu{e}_{TLH}=\text{min}\left(L,LOI\right)\left({\tau }_{t_1}-{\tau }_{t_2}\right)+\text{min}\left(L,LOI\right){\tau }_{t_1}R\left(1-{\tau }_{t_2}\right),$(1) is zero until the annual exemption amount has been achieved).

⁸ Berkin and Ye (2003) showed that whether this higher TLH alpha is attainable may depend on the liquidation strategy. This is an example that highlights the importance of understanding the determinants of TLH alpha in an integrated framework, a task we undertake later in the article.

⁹ See Campbell, Lettau, Malkiel, and Xu (2001) and Bessembinder (2018) for the magnitudes of idiosyncratic volatility in the US stock market.

¹⁰ Similar to Chaudhuri et al. (2020), in this study, we focus on tax-loss harvesting in the context of a long-only broad-based market investment; this is the most common form of participating in the equity market that is relatable to the broadest set of individual investors. We acknowledge that certain sophisticated investors may stand a chance of achieving greater TLH alpha than our analysis indicates, provided that these investors apply more advanced tax-aware investment techniques (e.g., gain management or short-selling) and have the requisite investor characteristics—to have ample short-term capital gains to offset.

¹¹ Some readers may question why we used this synthetic index instead of the S&P 500 or Russell 1000. This decision was driven by the ease of accessibility to daily information on returns and corporate actions at the individual security level. Axioma 400 behaves very similarly to the S&P 500, Russell 1000, and Russell 3000, with $R^2$ of 95%–96%, $\beta$ of 0.99–1.01, and $\alpha$ of –0.01–0.01 when we regressed our constructed Axioma 400 monthly returns on these other indexes.

¹² In unreported analysis, we considered other thresholds and found little economic difference with lower values than 10%. Threshold values above 10% generally resulted in a lower TLH alpha because of significantly reduced loss realization opportunities.

¹³ As discussed in the prior section, the primary goal of this article is to understand the impact of various investor profiles on TLH alpha. Tax-loss harvesting—harvesting losses immediately and deferring gains realizations as long as possible—is our focal point. How investors might optimize their gain/loss harvesting in response to different tax rates is beyond the scope of this study.

¹⁴ We are assuming current tax treatment over the entire period, which may not align with actual tax treatment (since the tax code may change in the future). Implicitly, we are assuming independence of asset pricing effects and tax rates in our computation of TLH alphas.

¹⁵ When replacing the position rather than eliminating it, replacement securities are necessary to comply with the wash-sale rule. This rule does not allow a tax loss if a substantially identical security is purchased 30 days before or after the loss transaction—the “wash-sale period.” See IRS Revenue Ruling 2008-5 for additional information.

¹⁶ Our research design results in realistic transaction costs and new cost bases that would commonly arise from pursuing TLH strategy; both are material considerations in quantifying TLH alpha. At the same time, the usage of an identical replacement security also ensures no tracking error in our analysis; we expect the tracking error from pursuing TLH in a real-world setting to have an average of zero and, as such, have negligible impact on TLH alpha.

¹⁷ We do not assume a fixed trading commission, because it is rare to find in the current marketplace.

¹⁸ Sponsored by the Federal Reserve Board and administered by NORC at the University of Chicago, the SCF is a repeated triennial cross-sectional study that is generally thought to have the best information on US households’ balance sheets. See Campbell (2006), who compared the SCF with other publicly available sources.

¹⁹ See Appendix B for details on how we defined net worth groupings.

²⁰ For all investor profiles, we assumed the short-term capital gains tax rates for harvesting and long-term capital gains tax rates for liquidation. Our results based on these archetypes may be overstating the benefit of TLH for a real-life investor who is otherwise identical to our archetype but with short- and long-term capital gains.

²¹ In this section, we assume counterfactually that the relevant tax rates for the archetypes remain constant throughout our historical analysis. In the next section, we explore the impact of a significantly more diverse tax rate arrangement on TLH alpha.

²² Taxable equity holdings include all publicly traded equities held by the household in their taxable financial accounts, excluding those held in tax-advantaged retirement savings accounts, such as 401(k), 403(b), and IRA accounts. Capital gains include capital gains realized from all sources, as reported in the household’s tax document. See Appendix B for detailed definitions of SCF variables.

²³ In unreported analysis, we determined 10% of loss per year to be a reasonable upper bound of average annual loss harvest achievable with TLH, absent a more sophisticated channel, such as a long–short vehicle (see Liberman et al. 2020).

²⁴ Compared with the case of using capital gains only, LOI leads to a noticeable increase for the 50th–95th groups, because the addition of up to $3,000 tends to double or triple the existing capital gains for these investors.

²⁵ This unlimited LOI/EQ assumption is modeled after the baseline setup in Chaudhuri et al. (2020).

²⁶ At the end of each calendar year, we divided the loss carryforward amount by the portfolio balance, which generated 15 carryforward-to-balance ratios for a given investor profile. Each point in each line in Figure 2 represents the average of these 15 ratios.

²⁷ Specifically, date t’s loss may be defined as $L_t=L_{t-1}-\text{min}(L_{t-1},LO{I}_{t-1})+\Delta L_t$, where $\Delta L_t$ is the new loss harvested between date t – 1 and date t. For Type 1 and 2 investors, this is almost always $L_t=L_{t-1}-LO{I}_{t-1}+\Delta L_t$ because $L_{t-1}>LO{I}_{t-1}$ by a wide margin, and it continues to grow over the course of TLH.

²⁸ See Southall (2020).

²⁹ We assumed that the investor holds the index constituents in exact proportion to their capitalization weights and rebalances at the same time as the index.

³⁰ In unreported analysis, we found that over the 38-year period, a TLH investor with direct indexing would have generated total losses 80% greater than with commingled vehicles only. Much of these additional losses are harvested in bull markets when time-series volatility is low and the average market return is high. While cross-sectional volatility is generally also low in these environments, such volatility represents “the only game in town” for TLH investors looking for losses to harvest during periods of rising equity market returns.

³¹ As discussed before, TLH may also be useful in making an active—higher-turnover—investment strategy more tax efficient by reducing tax drag of excess capital gain realizations from active management.

³² Ivković, Poterba, and Weisbenner (2005) found that individual investors tend to generate loss harvests in response to rising capital gains in an attempt to reduce the tax liability toward the end of the calendar year. This behavior is also consistent with our conjecture on how Type 1 and 2 investors will approach the topics of TLH and direct indexing.

³³ For example, we would anticipate a lower TLH alpha for those with high and recurring LOI but with a smaller tax spread than Type 3 or 4 investors. Similarly, some investors may relate to investor Type 1 or 2 on cash flows and liquidations but have a greater LOI; for these investors, a higher TLH alpha than what we have shown for Type 1 or 2 would be more likely.

³⁴ Creating a robust mapping of the determinants to the TLH alpha requires a sizable sample of investor profiles as described in this subsection. With 150,000 distinct cases, we chose the fund-based tax-loss harvesting for our BRT-based analysis because it allowed us to compute the TLH alpha in a reasonable time frame. The main objective of our BRT-based analysis was to shed light on how the machinery of TLH alpha generation works, rather than to pin down the level of the TLH alpha for a particular individual investor profile (the subject of the prior section). Given this context, our conjecture is that the mode of TLH implementation will not be of central importance to achieving our objective.

³⁵ See Goldberg et al. (2019), who applied a similar overlapping approach to the 1995–2018 period.

³⁶ See also Elith, Leathwick, and Hastie (2008) for an overview of BRT for practitioners.

³⁷ All hyper-parameters on our short list feature the following: (1) tree complexity that allows splitting three to six times, (2) a number of fitted trees ranging from 160 to 800, and (3) a learning rate ranging from 0.09 to 0.19. Key results represented in relative importance charts and partial dependence plots are all very stable across these hyper-parameters. The five-fold cross-validation MSE is plateaued and fully stabilized around all of our hyper-parameters.

³⁸ Despite this relatively low average impact, LIQ is a steady driver of TLH outcomes, as shown later in Figures 6 and 7.

³⁹ The average return plot is also the least smooth (with notable kinks around 4%) because, unlike all other variables that have more even representation, it is fit on 24 distinct values.

⁴⁰ As we reminded readers earlier, the focus here is the curvature and slope of these plots. Since the plot is generated by inputting the driver’s local value and all other variables’ average values, the level of the resulting TLH alpha may not appear particularly intuitive at times.

⁴¹ This number is calculated as 120 bps (from Figure 7) – 66 bps (31 bps + 35 bps from Figure 6) = 54 bps.

⁴² This is also reminiscent of the results in Table 5.

⁴³ As mentioned before, all partial dependence plots—whether univariate or bivariate—assume average values for all other variables. In particular, in this case 30% is assumed for HTR throughout the surface.

⁴⁴ For details of the variables we do not define in this appendix, see “Appendix: Survey Procedures and Statistical Measures” in Bhutta et al. (2020).

⁴⁵ The exposition and notation closely follow those in Chapter 8 of James, Witten, Hastie, and Tibshirani (2017).