HomeJournalCareers

Our Approach to Direct Indexing and Tax-Loss Harvesting

Direct indexing gives investors control over the individual stocks that make up their index exposure, allowing a portfolio to accommodate existing holdings, investment restrictions, and changing cash needs. It also enables tax-loss harvesting at the level of individual purchases. However, managing this means coordinating decisions across the account: a sale that creates a useful tax loss also changes what the investor owns.

At Autonomous, portfolio construction begins with the client's goals, investable wealth, and views about the market. Our system leverages portfolio management methodologies, goal-driven optimization and simulation to develop allocations that balance long-term investment exposure, funding for specific goals, and the client's strategic convictions. The Autonomous Index translates the eligible index exposures within those allocations into one client-specific benchmark. We manage holdings and tax lots against that target, evaluating trades for their tax consequences, implementation costs, and effect on portfolio risk.

Building on established direct indexing and tax-loss harvesting strategies, our approach brings AI systems together with deterministic, quantitative portfolio construction to implement them across the client’s allocation. This article explains that investment process: how we construct the Autonomous Index, use factor models to measure risk, and evaluate tax-aware trades. Our forthcoming technical papers will present our architectural and methodological contributions to AI and investment management in greater detail.

What direct ownership can mean for your taxes

When you invest in an index fund, you own shares of a fund that holds the underlying securities. With direct indexing, you own a portfolio of individual stocks selected and weighted to track a chosen index or mutual fund, such as SPY. The objective is to maintain similar market exposure while gaining more control over individual holdings, including which shares to sell and when.

Tax-loss harvesting puts that control to work. When a holding falls below its purchase cost, selling it can realize a capital loss that may offset taxable gains elsewhere. The proceeds can then be reinvested in eligible replacements chosen to maintain the portfolio’s intended exposure. For an investor with gains from selling appreciated investments or reducing a concentrated position, usable losses can reduce the current tax bill and leave more capital available to invest.

These opportunities can arise even when the overall market is rising. Individual stocks do not move together, and some positions may trade below their purchase prices while the portfolio appreciates. An investor holding a fund cannot separately sell one of its underlying stocks to recognize that loss. Direct ownership makes those individual positions available for tax management.

The purchase history provides further flexibility. Older shares in a company may carry a substantial gain while a recent purchase is below cost. Each purchase creates a tax lot with its own cost basis and holding period. Selecting the recent lot for sale can realize the loss while leaving the appreciated shares invested. The eventual benefit depends on the investor’s ability to use the loss, trading costs, wash-sale rules, and future tax consequences. For US individual taxpayers, unused capital losses generally carry forward indefinitely to offset future capital gains. The timing of their use still affects the economic benefit. 

At Autonomous, we apply these decisions across the eligible exposures in the client’s allocation. That combined perspective matters when different funds own the same companies—and when the client’s allocation changes.

The Autonomous Index: from plan allocations to a single benchmark

A client’s allocation may include funds providing broad U.S. equity exposure, smaller companies, international markets, or strategies focused on particular industries and investment styles. Their underlying indexes often own many of the same companies. Managing each exposure independently can create unnecessary trading when the allocation changes, even where much of the intended ownership remains the same.

The Autonomous Index combines eligible index exposures before we construct the direct portfolio. Each source index contributes according to its allocation in the plan, and overlapping companies accumulate into a single target weight. The benchmark record retains each source index’s contribution, while the shares and their tax lots are managed together.

Consider a client moving from a direct portfolio tracking SPY to one tracking QQQ. Many companies appear in both portfolios, although at different weights. If the transition is implemented by liquidating one portfolio and establishing the other, the account may sell shares it will shortly need to buy again. Those transactions incur costs and can realize taxable gains. For shares sold at a loss, repurchasing within the wash-sale window can defer the potential tax loss we harvests.

Our composite approach evaluates the change in each company’s combined target weight. As a simplified illustration, assume a $100,000 transition with 54% matched overlap: the smaller of each shared company’s starting and target weights sums to 54%. Approximately $54,000 could remain in existing positions, while $46,000 of holdings above their new targets would be sold. If shared companies represent 93% of the target portfolio, the proceeds would fund approximately $39,000 of additions to those companies and $7,000 of purchases in companies unique to the new target.

This becomes useful when the plan changes. Reducing one allocation and increasing another may leave a company’s combined target weight nearly unchanged. When an adjustment is needed, we can evaluate the available lots across the position, without treating the change as separate transactions in each allocation.

This combined views also provides us with more choices in rebalancing. When weights drifts in the portfolio, proceeds from a loss sale can fund another underweight index allocation in the composite. New funding contributions can move the portfolio toward its combined target allocation. We evaluate these purchases against the full benchmark, taking account of how each trade changes the client’s investment exposure.

Preserving the intended exposure

When we select an equity allocation in a financial plan, its benchmark defines the exposure the client expects to receive. Direct management introduces differences as we accommodate restrictions or replace stocks sold for tax purposes. Some differences are small and broadly diversified; others leave the account more sensitive to a particular industry or style of investing. We need to measure their combined effect before deciding how much tax-management flexibility the portfolio can support.

We call the difference between portfolio and benchmark returns the active return. Tracking risk estimates how much that difference is likely to vary. A portfolio with higher tracking risk has more room to depart from the benchmark in either direction, including periods of underperformance. Unexpected outperformance also warrants examination, since it can reflect an exposure the client did not intend to take.

Position weights give an incomplete account of this risk. A one-percentage-point underweight in a volatile semiconductor company can have a very different effect from the same underweight in a regulated utility. Several small underweights in regional banks can accumulate into meaningful sensitivity to credit conditions. We use a tracking-risk limit to control these departures across the portfolio, allowing individual holdings to differ where the resulting exposure remains acceptable.

The appropriate limit depends on the mandate. A client bringing a concentrated, low-basis holding may accept a gradual transition toward the benchmark to avoid a large immediate tax bill. A newly funded account has fewer inherited constraints. We consider the tax cost of adjustment alongside the risk of retaining the existing holdings.

Measuring benchmark risk with factors

To track a benchmark, we need to understand how market movement changes the account’s exposure to the forces affecting stock returns. Company names are a poor map of portfolio risk. Banks react to many of the same changes in credit conditions, and energy producers share a sensitivity to commodity prices.

A factor risk model gives these relationships a quantitative structure. We use industry standard factor risk models in our portfolio construction. For each security, the model estimates common sources of variation, including industry and investment characteristics such as size, value, momentum and quality, together with risk specific to each company. It also estimates how those factors move together and how much of each company’s risk is specific to it. Applied to the differences between the portfolio and its benchmark, these estimates give us the tracking risk and show where it comes from. It gives us a way to evaluate a proposed trade in the context of the holdings already in the account. Several small differences can reinforce the same exposure, while others offset one another.

Picture two portfolios whose stock weights differ from an index’s by the same total amount. One spreads its underweights across unrelated companies. The other concentrates them in a single industry. A holdings comparison would show them as equally close to the benchmark, yet their economic exposures are quite different. The same distinction matters after a tax sale, because a replacement from the same industry can still differ from the original stock in size, valuation, or momentum.

The benchmark sets the intended exposure, and Axioma’s estimates tell us how far the portfolio departs from it. Our optimizer combines the risk estimates with holdings, cash, eligible purchases, tax lots, and restrictions, and those inputs can pull against each other. The lowest-risk basket might require selling a large embedded gain. The largest available loss might remove an exposure that is costly to replace. Settling those conflicts means deciding what the loss is actually worth to the investor.

From a loss lot to a portfolio trade

Return to the client from earlier, who bought shares in a company several years ago and added to the position recently. The older shares have appreciated, while the recent purchase is below cost. Selecting the newer lot for sale allows us to realize its loss and retain the appreciated shares. Selling proportionately across both purchases would realize some of the older gain as well, reducing the net loss available.

We evaluate that sale together with the purchases needed to reinvest the proceeds. An existing underweight elsewhere in the composite may provide a suitable destination. Another replacement might preserve the industry allocation but increase a factor exposure the account already holds. Our optimizer compares these combinations using the estimated tax effects, trading costs and resulting tracking risk. A smaller loss can be preferable when capturing it requires fewer costly adjustments.

Purchase eligibility also depends on trading history. Acquiring substantially identical securities within 30 days before or after a loss sale can affect the deduction, including certain purchases by a spouse or in an IRA. We screen the activity recorded in the managed account; final tax treatment also depends on relevant activity outside it.

The account carries the consequences of these decisions forward. Replacement purchases establish new tax lots, contributions provide additional cash to invest, and withdrawals require us to select holdings for sale. Each subsequent rebalance uses that accumulated history.

Historical simulation: opportunity and tracking

To make sense of how these decisions accumulate, We evaluated a hypothetical $1 million Autonomous Index portfolio from January 2020 through December 2025, comparing it with a fund-based implementation of the same allocation. The account began with cash and received no subsequent contributions or withdrawals. We carried forward its holding and tax histories throughout the study, allowing us to examine how harvesting opportunities changed as the account matured.

The directly managed account harvested $262,223 of losses after wash-sale adjustments. The largest opportunities arose during the market decline in 2020, when many recently purchased shares fell below cost. During the subsequent recovery, appreciated holdings remained invested, and further harvesting depended on losses available among the remaining tax lots. The annual results show how unevenly those opportunities occurred across the period.

To translate the losses into an illustrative tax benefit, we assumed the highest 2025 marginal rates for a single New York City resident and sufficient eligible gains to use the losses immediately. After accounting for tax on gains realized within the strategy, the cumulative estimated benefit was $97,491 before trading costs. The comparison table also reports this amount net of modeled trading costs. Tax savings were not reinvested.

We also measured the investment exposure maintained while making these trades. Annualized tracking error for the tax-aware direct index portfolio was 1.42%. International equities, municipal bonds and Bitcoin exposure followed identical paths in the comparison accounts, isolating the investment differences to the US equity implementation. The accompanying graph and table report portfolio returns separately from the estimated tax benefits.

These tax estimates cover realized capital gains and losses before liquidation. They exclude taxes on distributions and on a final sale of the remaining holdings, whose cost basis generally falls as losses are harvested. The technical notes provide the annual tax effects and calculation assumptions.

Where direct indexing is most useful

We expect the economics to be more favorable when a client can use losses promptly and hold the replacement investments long enough to benefit from the deferral. A business sale or the gradual reduction of a concentrated stock position can create gains against which those losses have immediate value. Regular contributions also give us more scope to harvest and rebalance without selling appreciated holdings.

An established portfolio with substantial embedded gains presents a different problem: changing its exposures can itself generate a tax bill, while available losses may be limited. We weigh that cost against the benefit of moving closer to the client’s intended allocation, including the additional benchmark risk of leaving existing positions in place.

Technical notes

Portfolio state and factor risk

The implemented decision works directly with proposed trades. Let denote current security weights, the benchmark weights, proposed purchases, and proposed sales across bounded tax-lot disposal segments. A membership matrix maps those lot segments back to securities. The post-trade active weights are

The risk model represents each security through common-factor exposures and security-specific risk. Let be the exposure matrix, the annual factor covariance, the vector of annual specific volatilities, and a square root satisfying . Active variance is evaluated directly in factor form:

Forecast annual tracking error follows from the same quantity:

The risk inputs are annual and expressed in decimal units, so no further frequency adjustment enters the construction calculation. Realized tracking error is calculated separately from the observed portfolio and benchmark return series.

The tax-aware portfolio decision

Each sale segment has a maximum size and a marginal tax liability per dollar sold. A creditable loss gives a negative value, a gain gives it a positive value, and a loss affected by the account’s current wash-sale state receives no tax credit in the objective. With denoting the modeled trading-cost rate, the implemented program has the structural form

subject to the principal portfolio constraints

Here is the tracking-risk limit and is available cash as a share of portfolio value. Purchase eligibility, client restrictions, exclusions, and the account’s wash-sale state add further linear restrictions to the feasible trades. The bounded disposal segments retain the amount, basis, holding period, and tax character of the lots available for sale, allowing tax cost to enter the same convex decision as benchmark risk and turnover. Moehle, Kochenderfer, Boyd, and Ang develop the broader convex-optimization framework in Tax-Aware Portfolio Construction via Convex Optimization.

The composite benchmark

If the financial plan assigns weights to eligible index benchmarks , the composite target is

Weights for the same security add across source indexes. The benchmark recipe retains the individual terms , preserving attribution while the account holds one combined position. Exposures that remain in approved fund wrappers can be represented alongside the directly held component when measuring the whole portfolio.

Tax lots and wash-sale state

For lot with quantity , adjusted basis , and sale price , the realized gain or loss is

The tax value depends on character, applicable rates, available gains or income, and wash-sale treatment. IRS Publication 550 describes a wash sale when substantially identical stock or securities are acquired within 30 days before or after a loss sale; 26 U.S.C. §1091 supplies the statutory rule. The taxpayer’s complete circumstances and professional advice determine final tax treatment.

Return transformation and evidence conventions

Portfolio and benchmark returns are compared over identical dates. Active return for period is . Realized tracking error over observations is the sample standard deviation of , annualized with the frequency-appropriate factor. Wealth compounds from simple returns as

For an index level , the simple return is . A continuously compounded return converts back to a simple return through . Portfolio and benchmark return differences are calculated after both series use the same convention. Tax cash flows, contributions, withdrawals, fees, and trading costs are recorded separately from investment returns.

Illustrative New York tax scenario

We assume the highest 2025 statutory marginal rates for a single NYC resident: 37% federal for short-term gains, 20% federal for long-term gains on stock, 3.8% net investment income tax, 10.9% New York State tax, and 3.876% NYC tax. The combined rates are 55.576% and 38.576%, respectively, before deduction effects. We assume the affected gains remain in these brackets throughout each year. New York's 10.9% top rate applies above $25 million of taxable income. We hold the rates constant across the 2020–2025 simulation as an illustrative high-tax scenario. Sources: IRS income-tax brackets, capital-gains rates, net investment income tax, and New York's 2025 schedules and worksheets.

Let and be allowed short- and long-term losses in year , and and be gains realized by the strategy. Assuming sufficient external gains of each character remain available for immediate offset, the estimated tax benefit is

We sum annual benefits without reinvestment or discounting. Negative values represent a net tax cost. Disallowed losses receive no current credit; any later recognition is included only when recorded as allowed by the simulation. The fixed-rate calculation excludes bracket crossings, deductions and their interactions, credits, and alternative minimum tax. It assumes the full 3.8% NIIT applies to the affected investment gains. An investor's actual marginal tax effect depends on taxable income, loss utilization, and the character of gains available for offset.

Scope of the measured evidence

Each study follows three separately funded $1 million accounts over 1,507 trading dates: tax-aware direct indexing, the same construction method with its tax preference set to zero, and a fund comparison. The US stock portfolios use the same Axioma AXUS51-MH factor-risk framework. Decisions use preceding-session holdings and risk vintages, with execution modeled at opening prices and performance measured at closing prices. Two basis points of executed notional are deducted as trading costs. Fund expenses are embedded in ETF prices; advisory fees are excluded.

The composite begins with the following allocation.

The US benchmark maintains fixed relative weights in the four source indexes. Its fund comparison rebalances SPY, IJH, SPHQ and SOXX to those weights daily. The remaining positions follow the same buy-and-hold path in all three composite accounts, with distributions reinvested and no reallocation between sleeves. Bitcoin uses an opening-minute purchase quote and hourly candles ending at the NYSE close, including shortened sessions.

The input histories incorporate the vendor revisions available when the study was assembled. Holdings snapshots may be carried forward for up to seven calendar days. Equity targets exclude fund cash and currency balances; incomplete identifier coverage is handled by rescaling the remaining constituents. Invalid weight units are reconstructed from complete same-date market values where available. For an unquoted position or a loss of risk coverage, we model a cash exit at the current quote or the most recent observed price within ten calendar days. The proceeds remain in the account.

Disclaimer

Investment advisory services are provided by Autonomous Wealth Management LLC (“AWM”), an investment adviser registered with the SEC. Registration does not imply a certain level of skill or training. Autonomous Technologies Group is an affiliate of AWM. AWM’s advisory fees and conflicts of interest are described in AWM’s Form ADV Part 2A and Form CRS.

This article is research and education. It is not investment, tax or legal advice, and it is not a recommendation of any security or strategy. Tax-loss harvesting generally defers tax rather than eliminating it: a harvested loss lowers the cost basis of the replacement shares, so a later sale can realize a larger gain. Whether any loss can be used depends on the investor’s complete circumstances, including activity in accounts AWM does not manage. Investing involves risk, including loss of principal, and a direct indexing portfolio can underperform its benchmark.