After Tax Backtesting

After-tax backtesting

An after-tax Sharpe without a rate, a jurisdiction and a lot method is not comparable to anything — including the same strategy next quarter. So report() refuses to print one.

These are modelling assumptions for a backtest, not tax advice. Confirm current law with a qualified professional before relying on any of it.

This is the assembly skill. It owns nothing of its own: it wires ../tax-lot-matching-and-cost-basis/SKILL.md, ../wash-sale-rules/SKILL.md and ../section-1256-and-derivatives-tax/SKILL.md onto a return series and reports both Sharpes. Read those three for the rules and their sources; read this one for what happens when they meet a backtest.

✅ Measured comes from scripts/after_tax.py — numpy + pandas, seed 20260909, one 4-year path, 3.5 s.

1. The guard, and why refusing is the feature

TaxAssumptions(jurisdiction=..., short_rate=..., long_rate=..., lot_method=...)

All four are required positional fields with no defaults, and require_assumptions() raises MissingAssumptions on anything less. ✅ Measured — the demo's two refusals, verbatim:

guard: after-tax results are not comparable without assumptions. Pass a TaxAssumptions
       with ['jurisdiction', 'short_rate', 'long_rate', 'lot_method'].
guard: after-tax result is missing ['long_rate', 'lot_method']. An after-tax Sharpe without
       a rate, a jurisdiction and a lot method is not comparable to anything.

🔑 There is no default rate in this library and there never will be. A rate depends on the jurisdiction, the entity, the bracket and the year. A default turns an after-tax figure into an unlabelled opinion — and the number that comes out is lower than the pre-tax one, which makes it look conservative and therefore trustworthy. It is neither.

report() prints the assumption line above every result, plus the timing convention and any notes, so two after-tax numbers can be compared by comparing their headers first.

2. ✅ Measured — the after-tax turnover penalty

One seeded 4-year path. 1,000 shares held constantly. The only difference between rows is how often the position is sold and bought straight back at the same price with zero transaction cost — so the gross return series is identical by construction and every difference below is tax. Assumptions: US federal individual, short 37%, long 20%, FIFO, wash sales on, losses carried forward, no state tax, no NIIT.

✅ Measured: gross Sharpe has exactly one distinct value across all six rows, and pre-tax P&L is 36,722.53 on every one of them.

🚨 Churning daily costs 36.1% of the after-tax Sharpe of doing nothing, on identical exposure. That is not a commission — commissions are zero here by construction. It is the tax code repricing a strategy whose gross behaviour never changed.

🔑 The penalty saturates. Monthly, weekly and daily churn all pay 16,270–17,108 of tax. Once the basis is reset at every new high there is nothing left to accelerate, so the marginal tax cost of trading more often than monthly is roughly zero here — the first increment of turnover is what costs, and turnover beyond it is free of tax cost (though not of anything else).

3. 🚨 The asymmetry that makes the penalty bigger than it looks

Selling and rebuying immediately does not harvest losses. It only realises gains:

a round trip through a GAIN  ->  realises it, early, mostly at the short-term rate
a round trip through a LOSS  ->  is a wash sale, so the loss is DEFERRED into the new basis

So the basis becomes the running maximum of the price, realised gains are the sum of the new-high increments, and the losses simply never arrive. ✅ Measured:

🚨 A backtest that ignores wash sales reports this strategy as completely tax-free, and its after-tax Sharpe equals its pre-tax Sharpe to three decimals. Not "a little optimistic" — the entire tax bill vanishes, because the disallowed losses net against the gains and produce a paper loss year.

⚠️ Read the loss deferred column as a running total, not as economic loss. At daily churn it reaches −16,736,156 on a book worth about 137,000. That is the ratchet documented in ../wash-sale-rules/SKILL.md §5: each disallowed dollar re-enters the basis and is disallowed again the next day, so the column counts the same dollars hundreds of times. It is the right number for the mechanism and the wrong number to put in a report.

4. ✅ Measured — the same exposure in a §1256 wrapper

Buy and hold, held instead as a §1256 contract (year-end marks, 60/40, 26.8% blended):

Taxable income on a position nobody sold swings from −15,537.93 to +13,430.41 and costs 529.20 in total; the identical exposure held as stock recognises 0.00 every year and pays 0.00. The 2023 loss carries forward and absorbs the 2024 gain, which is why 2024's tax is zero — that is net_and_tax() doing character netting and carryforward, not a rounding artefact.

5. What the engine actually does, so you can disagree with it

⚠️ Every one of those rows is a modelling choice, and the honest way to use this module is to change the one you disagree with and re-run, not to accept the number. In particular:

• Tax is debited at the year end from outside the book. Real tax is paid in quarterly estimates and a filing months later. Moving the debit changes the after-tax path and hence the Sharpe, though not the total.
• The equity curve is not rebalanced to pay the tax. No shares are sold to raise cash, so the gross exposure stays exactly constant, which is what makes the gross Sharpe identical across rows and the comparison clean. A real account would have to fund it.
• Sharpe uses a zero risk-free rate, and sharpe() says so in its docstring. See ../../../fin-libraries/skills/lib-quantstats/scripts/rf_convention.py for how much that one choice moves a Sharpe on its own.
• If the tax bill exceeds the book, after_tax_sharpe is nan, not a number. A Sharpe computed across a zero crossing is meaningless, and the result carries solvent=False.

🚨 Not modelled, and each one is material: state and local tax, the 3.8% net investment income tax, the §475(f) mark-to-market election (which removes wash sales entirely and makes losses ordinary), qualified-dividend versus ordinary treatment, foreign withholding, the §1256 loss carryback election, straddles, constructive sales, and any account wrapper (IRA, 401(k), ISA, SIPP, PEA) — inside which most of this file is irrelevant and the answer is simply the pre-tax number.

6. Reporting it without misleading anyone

An after-tax backtest is hypothetical performance, and in the US that engages the Investment Advisers Act Marketing Rule — see ../../../fin-core/skills/us-market-rules/SKILL.md §6 for what has to accompany it. report() labels the output hypothetical and prints the assumptions; it does not print the trial count, the universe or the period, and ../../../fin-core/skills/research-integrity-guards/scripts/result_manifest.py is where those belong.

🔑 Report the pair, never the after-tax number alone. A tax assumption is one more researcher degree of freedom: a friendly rate and a friendly lot method can move an after-tax Sharpe more than most signal changes, and unlike a signal change it leaves no trace in the equity curve. The same discipline as a cost curve applies — see ../../../fin-core/skills/backtest-validation/scripts/cost_curve.py, and log the tax variant you tried in the trial ledger at ../../../fin-core/skills/backtest-validation/SKILL.md exactly as you would log a parameter.

7. Scripts and where this sits

scripts/after_tax.py — TaxAssumptions (four required fields, no default rates), require_assumptions / MissingAssumptions (the guard), net_and_tax (character netting and carryforward), realised_by_year, churn_blotter, after_tax_backtest, turnover_study, report, sharpe, and demo_prices. It imports the other three skills' modules through the dual-mode idiom, so it works both as python after_tax.py and as fin_skills.tax_accounting.after_tax. numpy + pandas, seed 20260909, 3.5 s.

• The three rule skills it assembles — ../tax-lot-matching-and-cost-basis/SKILL.md, ../wash-sale-rules/SKILL.md, ../section-1256-and-derivatives-tax/SKILL.md.
• The same turnover, priced as execution cost rather than tax — ../../../fin-core/skills/execution-cost-analysis/SKILL.md and ../../../fin-core/skills/backtest-validation/scripts/cost_curve.py. Add them; they are different drags on the same trades.
• Presenting hypothetical performance — ../../../fin-core/skills/us-market-rules/SKILL.md §6.
• Logging the tax variant as a trial — ../../../fin-core/skills/backtest-validation/SKILL.md.
• The A-share equivalent, where the turnover penalty is written into the tax code itself — ../china-ashare-trading-taxes/SKILL.md.