The Alpha With a Half-Life

Direct indexing and tax-loss harvesting: what the benefit is in year seven.

We built forty-two ten-year accounts out of the schedules of investments that index funds file with the regulator — 166 documents, three separate filers, 2006 to 2026. Each account holds every common stock in the filer's schedule at the filer's own weights, sells any lot standing below its own cost at every quarter-end, buys back within the wash-sale bar, and does nothing else. Every price in the study is one division of two filed numbers. No vendor, no aggregator, no survey.

The finding

The median account harvested 7.72% of itself in year one and 0.58% in year seven — about a fourteenth. For the median start date, half of everything the account will harvest in ten years is harvested by the end of year two.

The decay is not gentle at the top end either: across all forty-two start dates, no account harvested more than 3.33% in year seven, against a year-one maximum of 41.30%.

The half that is not usually published

Harvesting lowers cost basis. It does not remove gain — it moves it forward. After ten years:

• The median account sits on an unrealised gain of 65.01% of its own market value.
• At a single assumed rate of 23.8%, unwinding that in one day costs 15.47% of the account.
• Over the same decade it harvested 30.62% in losses and was forced by index departures to realise 15.76% in gains — because a rule that tracks an index must sell what leaves it.
• Net realised loss, taken start date by start date: 16.09%, worth 3.83% in deferred tax at the same rate.

So the lock-in at year ten is about four times the net tax deferred over the decade that produced it. That ratio is the finding. What anyone should conclude from it is not this report's business.

And it is not uniform. On seven of the forty-two start dates the decade realised more gain than loss. The worst, a March 2009 start, finished 20.40% net against itself.

And the rule is doing tax work, not stock selection

We ran every account twice: once harvesting, once with the harvest test switched off and everything else identical. The two end within a median 0.16% of each other over ten years (range −1.10% to +3.41%).

That control is the most important thing in the study, because two earlier versions of our own replacement rule failed it and are printed as errors. The first concentrated the account from 502 positions to 74 and beat its own control by +48.7% — a momentum portfolio wearing a tax-loss-harvesting label. The second still manufactured +30.3%. Only the third, which buys back toward the filer's own filed weights, produced drifts near zero.

A backtest that does not publish a no-harvest comparison is not measuring tax-loss harvesting.

Every harvest figure here is a floor, and we measured the floor

The regulator publishes only the third month of each quarter, so the study tests four times a year rather than 250. A daily harvester sees dips a quarter-end grid never sees.

We did not assume the size of that gap; we measured it. Running the identical account on a filer that published roughly half as many observations over the same decade, on the same market, harvested 60.7% to 86.3% of what the quarterly series harvested — a penalty of 18% to 39%. The direction is one-way. Every harvest figure in this report understates what a daily programme could have taken.

The lock-in half does not have that problem: unrealised gain is a stock, not a flow, and it reads the same on any grid. It is also where the two independent filers agree most closely — 65.01% against 65.65%.

The three numbers

What is in the full report

The eight quarterly markers and their current readings. The age effect measured with the market held fixed — the same quarter, the same prices, accounts of different ages, which is the only way to separate "the account got old" from "the market went up", and which shows a young account taking ten to twenty-five times an old one in a shallow fall, and an eleven-quarter-old account still taking 13.04% in December 2008. The full lock-in table for five cohorts. Six sensitivity runs, including what reinstating dividends would plausibly do to the late years. The split detection and the share issuance it cannot tell apart. A page on each of five start dates and one on the no-harvest control. Every error the research made, printed. A claim-level citation register of 38 rows, including the figures we could not verify — and one place where the source pack contradicts itself, printed rather than tidied.

Disclosures. Finnacl Research publishes impersonal investment research, identical for every purchaser and tailored to no one's circumstances. It is not individualised investment advice, and Finnacl Research is not an investment adviser (publisher's exclusion, Investment Advisers Act §202(a)(11)(D)). No statement in this report says what any reader should do, and none refers to any reader's tax position. The 23.8% rate is an assumption of the research. Nothing here is a forecast. The author holds no position in any fund or security named. © Finnacl Research 2026.