After-tax performance lens: model annual taxation of trading gains (report-layer first)

[dayfine]

Motivation

2026-07-19 ad-hoc analysis (session note): applying a naive 35%/yr tax on net MTM gain to the Run D record path compresses terminal wealth $80.1M → $18.8M (no loss carryforward) / $21.8M (with carryforward) — CAGR 18.0% → 11.7-12.3%. The effect is structural (tax paid annually out of compounding capital, ~0.65× log-growth), and it hits short-holding-period strategies (avg hold ~46d ⇒ nearly all gains short-term ordinary-rate) far harder than buy-and-hold, which defers essentially everything and pays LTCG once. Any honest-tradeable comparison of this strategy vs an index arguably needs an after-tax lens: rough after-tax edge estimate is still ~+4.5pp/yr vs after-tax SPY-TR B&H, but the headline multiple changes dramatically.

Why it isn't simple

• Basis: realized-gains taxation ≠ yearly-MTM (§475 trader election). Needs per-trade realization dates from trades.csv, not the equity curve.
• ST/LT split: holding period per trade decides ordinary vs LTCG rate; open positions at year-end defer.
• Loss carryforward across years (capital-loss netting rules).
• Wash sales: the strategy re-enters names within 30 days routinely (recoverer re-buys) — disallowed losses could be material.
• Cash-flow realism: tax paid in April from the account is a forced outflow the simulator doesn't model; at full deployment it forces sales (interacts with cash floor / sizing).
• Comparator fairness: after-tax strategy must be compared to after-tax benchmark (deferral + LTCG + dividend taxation on the TR series), per the total-return-comparator rule.

Proposed shape (incremental)

1. Phase 1 — report-layer lens (pure, post-run): an exe over an existing run dir (trades.csv + equity_curve.csv) computing after-tax wealth under configurable regimes: mtm_flat (the 35% naive model, reproduces the ad-hoc numbers), realized_st_lt (per-trade holding periods, ST/LT rates, carryforward), optional wash-sale adjustment. All rates/toggles in config (sexp), no simulator changes. Emits a summary table per run for release reports.
2. Phase 2 (only if Phase 1 shows decisions would change): in-sim April tax outflows (forced-liquidation realism), behind a default-off config flag per experiment-flag-discipline.

Non-goals for now

Not urgent; not blocking the resistance-v2 promotion or margin milestones. No in-sim changes in Phase 1. OCaml only (no-python rule).

Acceptance (Phase 1)

• Reproduces the 2026-07-19 naive numbers on the Run D dir ($18.8M / $21.8M) under mtm_flat 0.35.
• realized_st_lt mode agrees with a hand-computed small fixture (unit-tested).
• Runs against any scenario output dir; wired into the release-report flow as an optional section.

🤖 Generated with Claude Code

[dayfine]

Spec refinements from user review (2026-07-19) — these pin the Phase-1 model:

1. Annual taxation at year-end. April payment deferral is explicitly NOT modeled.
2. Losses are never deducted in-year — carryforward only (offset future gains; applied against ST gains first as the taxpayer-favorable order).
3. Rates: ST 35%; LT (holding ≥ 365d) 23.8% (LTCG 20% + NIIT 3.8%).

Reference run (awk prototype over Run D scenarios-2026-07-16-131756, realization basis by exit-year from trades.csv, cross-netting + carryforward): $80.14M pre-tax → $26.84M after-tax (CAGR 18.0% → ~13.2%). Compare naive MTM-flat 35%: $18.8M (no carry) / $21.8M (carry).

Findings the tool should surface per-run:

• The strategy is structurally tax-efficient: stops realize losses fast while winners ride unrealized (2019/2022/2023/2025 all realized-net-loss years despite rising MTM equity) — carryforwards offset ~$8M of the 2026 AXTI realization year.
• LT treatment is nearly immaterial at avg hold 46d (material LT only 2021/2024; AXTI at 336d missed ≥365 by 4 weeks). Worth a per-run 'days-to-LT for top winners' diagnostic column — a tax-aware exit-deferral lever would touch the exit spine, so it is out of scope unless separately proposed.
• Acceptance target updated: reproduce $26.84M on Run D under realized_st_lt (st 0.35) (lt 0.238) (lt_days 365) + carryforward, in addition to the earlier mtm_flat numbers.