RESEARCH

What is tax-loss harvesting worth? About 2,000 backtests across four benchmarks

2026-11-13 · 7 min read

Direct indexing means owning the stocks of an index yourself instead of through a fund. The main appeal is tax alpha: in any year, some stocks fall even when the index rises, and their losses can be sold to offset gains elsewhere. We ran about 2,000 five-year backtests of Abacus Labs' harvesting engine against four benchmarks, with accounts from $10,000 to $1 million, to see what an investor can reasonably expect.

  • Large caps: a median of about 0.7 percent a year over five years against the S&P 500 and Russell 1000. That is in line with published research from AQR, Vanguard and others.
  • Mid and small caps: more volatile stocks mean more losses to harvest, about 0.9 percent a year for the S&P MidCap 400 and 1.1 to 1.5 percent for the S&P SmallCap 600.
  • Timing dominates: depending on the start date, the same strategy earned between 0.1 and 2.1 percent a year.
  • The benefit is front-loaded: about 8 percent of the account is harvested in year one, about 1 percent a year by year five.
  • Accounts of $50,000 and up can track the S&P 500 closely with about 100 stocks. Smaller accounts give up more tracking.

What "tax alpha" means here

Tax alpha is the extra after-tax return of the harvested portfolio over an index fund with the same benchmark, after removing any difference in pre-tax return. It measures what the harvesting adds, not whether the stock picks happened to beat the index.

Two assumptions shape the number, and published studies differ on both. First, we assume the harvested losses offset capital gains the investor realizes elsewhere, taxed at top federal rates: 40.8 percent short-term and 23.8 percent long-term, including the 3.8 percent net investment income tax. Second, our headline figure assumes the whole portfolio is sold at the end of five years and the deferred gains are taxed. If the portfolio is instead held until death, when heirs receive a stepped-up cost basis, or donated, the median roughly doubles, to about 1.5 percent a year for the S&P 500.

How we tested

Each backtest starts from cash, buys the benchmark's stocks, and then makes one decision on the first trading day of each month. It sells any tax lot that has fallen at least 5 percent below its cost and reinvests the proceeds in a mix of stocks chosen to keep tracking error low. Every trade is in whole shares and pays 5 basis points in costs. The wash-sale rule is enforced 30 days in both directions. The universe is point in time: index membership as it stood on each date, including companies that later failed. The comparison is an index fund in the same benchmark, with the same dividends and dividend taxes.

We ran two sets. The first starts every quarter from 2002 to 2021 on a $1 million account, against the S&P 500 and the Russell 1000: 80 start dates, so the result does not depend on picking good years. The second uses 15 start dates chosen to include market peaks, crashes and calm years, and crosses four benchmarks with four account sizes.

Smaller companies leave more losses on the table

Harvesting feeds on dispersion: the more individual stocks move apart from their index, the more of them sit below cost at any moment. Mid- and small-cap stocks are more volatile, so they harvest more.

Portfolio sold after five yearsNever sold (step-up or donation)
0%1%2%3%S&P 500: sold after five years 0.67%0.67S&P 500: never sold 1.45%1.45S&P 500Russell 1000: sold after five years 0.61%0.61Russell 1000: never sold 1.60%1.60Russell 1000S&P MidCap 400: sold after five years 0.89%0.89S&P MidCap 400: never sold 2.21%2.21S&P MidCap 400S&P SmallCap 600: sold after five years 1.14%1.14S&P SmallCap 600: never sold 3.04%3.04S&P SmallCap 600
Median annual tax alpha over five years, 15 start dates from 2000 to 2023 and four account sizes ($10k to $1M), threshold harvesting with the fewest stocks that track the benchmark within about 1 percent predicted. Losses offset gains elsewhere at top federal rates. Hypothetical backtest; see disclosures.

The ordering holds at every account size. The trade-off is tracking error. A small-cap portfolio holds about 300 stocks and still drifts about 3 percent a year from its index, against about 1.5 percent for a 100-stock large-cap portfolio.

When you start matters more than anything you control

An account opened just before a market fall buys every stock near its high, and the decline that follows creates losses across the portfolio. An account opened after a long rally may wait years for its first meaningful loss.

0%0.5%1%1.5%2%2.5%200220062010201420182021S&P 500starts before the 2008 crash2009 bottom
Annual tax alpha over five years by start date: 80 quarterly starts, January 2002 to October 2021, $1 million S&P 500 account holding the whole index. The Russell 1000 line is nearly identical (range 0.16 to 2.11 percent). Hypothetical backtest.

Across the 80 start dates, annual tax alpha ranged from 0.09 to 2.11 percent for the S&P 500, and was positive for every one. Accounts opened in 2007 and 2008 earned about 2 percent a year. Accounts opened in the spring and summer of 2009, near the bottom and at the start of a long bull market, earned 0.2 to 0.5 percent. Our second set shows the same pattern: accounts opened at the 2000 peak, the 2007 peak and the month of the Lehman failure earned about 2 to 3 percent a year.

The benefit is front-loaded

Harvesting resets a lot's cost to today's price. After a few years most lots sit well above cost, and there is little left to harvest. This is the "decay" every study reports.

0%2%4%6%8%Year 1: losses harvested 7.92%7.92Year 1Year 2: losses harvested 2.99%2.99Year 2Year 3: losses harvested 1.91%1.91Year 3Year 4: losses harvested 1.45%1.45Year 4Year 5: losses harvested 1.20%1.20Year 5
Median losses harvested in each year of an account's life, as a percent of the account's value at the start of that year. 80 quarterly starts, $1 million S&P 500 account, no new deposits.

New money reverses the decay, because each deposit buys new lots at current prices. These tests assume no deposits, so they show the conservative case. An investor who adds regularly, or who harvests over a longer horizon with periodic reinvestment, should expect the later years to look more like the early ones.

How our results compare with published research

Different studies use different tax rates, horizons and assumptions about selling, so the fair comparison is to studies that, like ours, tax the portfolio at liquidation.

StudySetupAnnual tax alpha
Abacus Labs (this study)S&P 500 and Russell 1000, 80 starts 2002–2021, 5 years0.68% and 0.78% (median)
AQR, Liberman et al. (2023)Russell 1000, years 1–5 of a ten-year path0.69%
Vanguard, Paradise et al. (2024)Rolling windows 1982–2023, range of investor profiles0.47% to 1.27% (median)
Chaudhuri, Burnham and Lo (2020)Top 500 US stocks, 1926–20181.08%, or 0.82% with the wash-sale rule
J.P. Morgan Asset ManagementS&P 500, ten-year vintages since 1995about 1% on average, 0.4% to 2.0% by vintage

Our results sit in the middle of the published range and close to AQR's. One difference is worth noting: we harvest about 8 percent of the account in year one, where AQR and others report about 13 percent. We only sell lots down 5 percent or more and check once a month. Daily monitoring and a lower threshold harvest more, and Vanguard's research finds that daily monitoring accounts for most of the value. Abacus Labs can do both; these tests use the conservative settings.

Account size and the number of stocks

With whole shares, a small account cannot hold 500 stocks in the right proportions. The answer is to hold fewer stocks, chosen by an optimizer to behave like the index. More stocks lower tracking error but add little tax alpha: in our tests, 30 stocks and the whole index harvested about the same.

S&P 500 accountStocks heldRealized tracking errorAnnual tax alpha
$10,000503.0%0.47%
$50,0001001.8%0.82%
$100,0001001.5%0.72%
$1,000,0001001.5%0.68%
$1,000,000, whole indexabout 3700.9%0.85%

Medians over the 15 start dates. Platforms that buy fractional shares can track closely at lower balances.

What to keep in mind

  • Losses need gains to offset. Without other capital gains, only $3,000 of net losses a year can offset ordinary income; the rest carries forward. In that case the benefit over five years is much smaller, about 0.05 percent a year in our tests, though the carried-forward losses keep their value for later gains.
  • Harvesting defers tax; it does not erase it. Each harvested loss lowers the cost basis of what you hold. The value comes from paying later, from turning short-term rates into long-term rates, and from never paying if the assets pass to heirs or charity.
  • Fees and tracking error count. A 0.7 percent tax alpha is a gross figure. Management fees come out of it, and a portfolio that tracks its index within 1.5 percent can trail it in any year.
  • Small-cap portfolios drift. Our tests never sell a stock at a gain, so winners that leave the index stay in the portfolio. A real program would allow a small gains budget to keep tracking in line.

Tax-loss harvesting with Abacus Labs   How accurate is our risk model?

Hypothetical, backtested results. They do not represent actual trading or the results of any client account, and they were prepared with the benefit of hindsight in choosing the method. Results are gross of advisory fees and assume top 2026 federal tax rates, no state tax, losses fully usable against other capital gains, and liquidation at the end of five years unless stated otherwise. Tax treatment depends on each investor's circumstances. Past performance, actual or simulated, does not guarantee future results. Abacus Labs is software; it does not give investment or tax advice. "S&P 500", "S&P MidCap 400", "S&P SmallCap 600" and "Russell 1000" are trademarks of their respective owners, who do not sponsor or endorse this research.