Tax Lot Matching And Cost Basis

Tax lot matching and cost basis

A backtest reports the P&L of a position. A tax return reports the P&L of a set of lots, and which lot you sold is a choice. Until you say which choice, "the strategy made $X" is not a statement anyone can tie to a tax return.

These are modelling assumptions for a backtest, not tax advice. Every rule below carries the publication and section it came from and the date this library read it. Tax rules change; a dated claim that is wrong next year is fixable, an undated one is not. Confirm current law with a qualified professional before relying on any of it.

Everything marked ✅ Measured comes from scripts/lot_matching.py — numpy + pandas, seed 20260909, one four-year single-name blotter, under 1 s. Everything marked ✅ source-verified was read in IRS Publication 550 (2025), Investment Income and Expenses, Cat. No. 15093R, revised 2026-03-05, read 2026-09-09; page numbers are the printed page numbers of that edition.

1. ✅ Measured — one blotter, four rules, four answers

Seeded blotter: 72 trades (43 buys, 29 sells) in one name, 2022-01-03 to 2025-10-20, marked at 107.2423. The strategy's total P&L, realised plus unrealised, is +32,595.80 under every rule.

🚨 The realised gains differ by 20,093.73 — 61.6% of the entire strategy's P&L. HIFO reports a realised loss on a strategy that made money. LIFO reports zero long-term gain on a four-year book. Every one of these is arithmetically correct.

✅ Measured — the same four rules at an assumed 37% short / 20% long rate (an input the script prints, never a default it supplies):

Spread 7,378.78 in tax, on one blotter, in one name, in one year. This is the number a pre-tax backtest silently sets to zero.

2. ✅ source-verified — what Publication 550 actually says

FIFO is the default, and it is a default, not a preference. Pub 550 (2025) p. 62, under Stocks and Bonds → Identification not possible: "If you buy and sell securities at various times in varying quantities and you cannot adequately identify the shares you sell, the basis of the securities you sell is the basis of the securities you acquired first." The same paragraph closes the average-price door for ordinary stock: "Except for certain mutual fund shares, discussed later, you cannot use the average price per share to figure gain or loss on the sale of the shares."

Specific identification is an act, not a setting. Pub 550 (2025) p. 62, Adequate identification → Broker holds stock, verbatim — you make an adequate identification if you:

• "Tell your broker or other agent the particular stock to be sold or transferred at the time of the sale or transfer, and"
• "Receive a written confirmation of this from your broker or other agent within a reasonable time."

Two conditions, both contemporaneous with the sale. 🚨 A spreadsheet that re-picks the lots in April is not a specific identification of a trade made last August, and Pub 550 says the identified stock is what was sold "even if stock certificates from a different lot are delivered" — the identification is the operative act, which is exactly why it has to have happened.

"LIFO" and "HIFO" are not statutory methods for securities. There is no LIFO election for stock in Pub 550's list. What a broker's "LIFO", "HIFO" or "tax lot optimizer" setting does is stand in for the identification — it instructs the broker at the time of the sale and generates the confirmation. That is legitimate as a specific identification, and it is legally a different thing from a method you may choose after the fact. Pub 550's only named methods, and only for mutual funds (p. 67), are "Specific share identification" and "First-in first-out (FIFO)".

🔑 The practical test: if the broker confirmations say FIFO and your P&L report says HIFO, the report is not defensible. Model whatever your account is actually configured to do, and record which it was.

3. The invariant that makes the whole subject tractable

realised(rule) + unrealised(rule)  is IDENTICAL for every rule
realised(rule)                     is not

✅ Measured — the script computes realised + unrealised under all four rules and reports the number of distinct values: 1.

🔑 Lot selection moves P&L between this year and a later year, and between the short-term and the long-term bucket. It does not create or destroy a dollar. A vendor claim that a lot method "added N bps" is quoting the realised column and hiding the unrealised one; the gain a HIFO sale did not book is still sitting in the basis of the lots you still own.

What it is worth is real and comes from three places, none of which is free money:

1. Deferral — tax paid later is worth less. The size of this is the discount rate, not the gain.
2. Character — long-term versus short-term. In the table above LIFO reports 0.00 long-term on a four-year book; FIFO reports 10,336.65. That is a rate difference on real dollars.
3. Offset — a realised loss this year can meet a realised gain this year.

And one place it is worth less than it looks: ../wash-sale-rules/SKILL.md, because a harvested loss followed by a repurchase inside 30 days is deferred rather than banked.

4. Holding period: one calendar year, not 365 days

Long-term means held more than one year. is_long_term() uses pd.DateOffset(years=1), not a day count: a lot bought 2024-02-29 and sold 2025-02-28 is short; sold 2025-03-01 it is long. 🚨 A 365-day implementation gets leap years and month-ends wrong, and a year-end harvesting rule lives precisely on that boundary — so the error is not random, it lands on the trades the strategy cares about most.

5. Average basis is narrower than every tutorial implies

✅ source-verified, Pub 550 (2025) p. 67, Average Basis. You may use it only if the shares are identical, were acquired at different times and prices, are left with a custodian or agent, and:

• "They are shares in a mutual fund (or other regulated investment company);"
• "They are shares you hold in connection with a DRIP, and all the shares you hold in connection with the DRIP are treated as covered securities"; or
• "You acquired them after 2011 in connection with a DRIP."

Ordinary stock is not on that list. average_basis(blotter, share_class="stock") raises rather than returning a number, and quotes the restriction in the message. ✅ Measured: the demo blotter under average basis, if it were a fund, realises 5,942.94 — a fifth number, and one that is simply unavailable for the single-name equity book that produced it.

⚠️ Two details that get lost: the election must be made in writing to the custodian and is revocable only within a narrow window (Pub 550 p. 68, Revoking the average basis method election), and the holding period still runs FIFO — Pub 550 p. 68: "To determine your holding period, the shares disposed of are considered to be those acquired first." So average basis changes the size of the gain and not its short/long character. The script implements it that way.

6. What this skill does not model, and where it goes wrong quietly

🚨 Short sales are refused, not approximated. match_lots raises if a sale exceeds inventory. Short positions are IRC §1233, a different regime with its own holding-period rules, and silently netting them into a long lot stack is how a plausible wrong number gets produced.

⚠️ One symbol per blotter. Lots are per security per account. Running two symbols through one blotter would match a sale of A against a lot of B; the module does not check, so the caller must split.

⚠️ The 1099-B is not the answer either. Brokers report basis only for covered securities and their wash-sale adjustment (box 1g) is computed within one account and by CUSIP, so the same-security purchase in your other account, or in your IRA, is not in it. The identity between your model and the form is a thing to check, not a thing to assume.

7. Scripts and where this sits

scripts/lot_matching.py — match_lots (fifo / lifo / hifo / lofo), is_long_term, term_split, unrealised, compare_rules, spread_fraction, tax_due (rates are arguments, there is no default), average_basis (refuses ineligible share classes), and make_blotter for the seeded demo. numpy + pandas, seed 20260909, under 1 s.

• The loss that gets disallowed and re-based — ../wash-sale-rules/SKILL.md.
• Contracts with no lots: 60/40 and the December 31 mark — ../section-1256-and-derivatives-tax/SKILL.md.
• Attaching all of this to a backtest, with the stated-assumptions guard — ../after-tax-backtesting/SKILL.md.
• The US rulebook this sits inside — settlement, margin, Reg SHO, and the one-paragraph wash sale summary this plugin expands — ../../../fin-core/skills/us-market-rules/SKILL.md.
• Turning a tax drag into a breakeven and a capacity limit, the same way a cost assumption is handled — ../../../fin-core/skills/backtest-validation/scripts/cost_curve.py.
• Where the turnover that drives all of this comes from — ../../../fin-core/skills/execution-cost-analysis/SKILL.md.