Skip to main content

Transaction costs

Until a cost model charges trading, every Sharpe and equity curve is gross — and gross results are the most reliable way to fool yourself, because the strategies that look best in-sample are very often the highest-turnover ones that costs destroy. src/costs/ prices the cost of changing a position so research metrics are net by default.

Research clock only

This models cost for simulation. The live path gets real fills from the broker; it imports no cost model.

The decomposition​

For a trade of q shares at price p in a name with average daily volume ADV and quoted spread s:

cost($) = commission + (s/2)·|q|·p + impact_rate·|q|·p

  • Commission — a flat per-notional broker fee (default 1 bp).

  • Half-spread — (s/2)·|q|·p, the immediate cost of crossing (default 5 bp spread).

  • Market impact — the square-root law (Almgren et al., the robust default):

    impact_rate = η · σ_daily · √(|q| / ADV)

    |q|/ADV is participation (the fraction of a day's volume you demand), σ_daily the name's daily volatility, η a coefficient (~0.3 default). Impact is concave per share (√) but convex in total cost (|q|·impact_rate ∝ |q|^{3/2}) — exactly the property that makes spreading a target across names cheaper than dumping it on one. A linear-participation fallback is offered for a convex-quadratic optimizer.

As an alpha haircut​

Amortizing the round-trip cost over the holding period gives a cost rate per unit time, which is the honest input to portfolio construction:

α_net = α − round_trip_cost_rate / holding_period_years

A +4%/yr alpha with a 2% round-trip held for a month is not a +4% opportunity — it's deeply negative. This is what stops a high-IC, high-turnover signal from looking profitable when it isn't.

Integration​

  • Backtest (engine/backtest.py): every fill is charged on both legs (entry and exit), using a trailing (as-of) ADV and volatility so a past trade's cost never depends on future volume. Metrics become net; total_cost and the gross final capital are reported alongside for the haircut attribution. Net is the default at the service/CLI layer; --gross disables the charge.
  • Portfolio construction (008): the proposed turnover is priced through the cost model and reported as a cost drag on the expected active return (the ex-post accounting; a name-specific cost term in the objective lands with an SOCP solver).
  • Participation cap: a trade demanding more than ~10% of a day's volume is flagged — the seed of a capacity analysis (how much capital before cost eats the edge).

Borrow (short financing)​

Holding a short accrues a borrow cost over time: borrow_rate · notional · holding_years (default 50 bp/yr, --borrow-bps). The backtest charges it per short position by how long it was held, so a short-heavy strategy isn't silently flattered. Long-side margin financing and leverage costs remain out of scope.

Known gaps​

The optimizer uses the linear cost term while the backtest charges the √-law; the gap is small and documented. A name-specific cost term inside the optimizer objective needs an SOCP solver (the ex-post drag report is the v1 stand-in).

Where it runs​

ParametricCostModel in src/costs/. Surfaced as backtest flags (--gross, --commission-bps, --impact-eta) and folded into compute_risk-adjacent flows; the backtest report and the MCP run_backtest tool both return net metrics + total_cost.