[sijie-Z]

P0. Last of the four research-integrity defects — and the only one that invalidates every backtest number this platform has produced.

What was wrong

CostModel.compute_costs charges commission and slippage on the value it is given, and halves stamp tax on the assumption that half of that value is sells:

commission_cost = abs_turnover * commission_rate
stamp_cost      = abs_turnover * stamp_tax_rate * 0.5

The * 0.5 only holds for the two-way traded value (buys + sells). Both callers passed one-sided turnover — which is the convention this codebase documents at portfolio/constraints.py:16: "Maximum one-sided turnover per rebalance".

So every component was charged at half.

Measured

30% one-sided turnover, commission 3bp / stamp 10bp / slippage 5bp:

	cost
before	3.90 bp	compute_costs(0.30)
after	7.80 bp	compute_costs(0.60)
ratio	2.0000

Hand-check for the same inputs: commission 0.60×3bp + stamp 0.30×10bp + slippage 0.60×5bp = 7.80 bp, matching to 1e-12.

A monthly strategy at 30% one-sided turnover was undercharged by ~47 bp/year.

Fix

Both call sites now pass turnover * 2:

• backtest/engine.py
• backtest/capacity.py

The recorded turnover stays one-sided in both places — it is a reported metric, not the base the cost model is fed. Conflating those two roles is exactly what caused the defect, so the comments now say which is which.

Tests

• test_cost_model.py pins the contract itself: two-way input produces the hand-computed cost, and one-sided input produces exactly half.
• test_engine.py's rebalance test asserted 999_500 — the half-charged figure — and now asserts 999_000.

1318 passed, 1 skipped.

Impact — this is a research-result invalidation event

Affected: every CAGR, Sharpe, Sortino, net-return, drawdown and turnover-adjusted figure produced through BacktestEngine — python main.py run / compare / sweep, the walk-forward aggregate, and everything in README.md's milestone table.

Per NO_LOOKAHEAD_CONTRACT.md's own remediation procedure, the prior results should be marked affected, not deleted — the corrected numbers come from re-running, and the old ones stay traceable.

The result-validity metadata needed to do that does not exist yet (a valid / affected / invalidated / superseded field plus affected_by on the research registry). This PR does not add it; it is the natural next piece of work, and BUG-22 / BUG-23 / BUG-21 have the same requirement. Without it, deciding which results to re-run is a manual exercise in reading commit messages.

To be explicit about the alternative that was rejected: keeping the halved cost so the published numbers stay comparable is not an option. A platform whose stated principle is Truth First cannot knowingly keep a wrong number because it is already written down.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 8103d511-a8ef-4509-a170-42a4987d8f19

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}