Calibrating A Market Impact Model

A market impact model — something like "expected impact scales with the square root of the fraction of daily volume traded" — is a useful shape, but a shape alone doesn't tell a trader how many basis points a specific order in a specific stock will actually cost. Two stocks that trade the same average daily volume can have very different impact, because one has a tighter spread, deeper resting size, or more resilient liquidity than the other. Calibration is the process of fitting a chosen model's free parameters to a firm's own historical execution data, so the model's predictions match what actually happened when real orders were worked in real markets — turning a plausible functional form into a usable, firm-specific tool.

What gets fit, and against what

A common impact model takes the form

$$I = c \left(\frac{Q}{V}\right)^{\alpha} \sigma ,$$

where $I$ is the price impact as a fraction of price, $Q$ is order size, $V$ is average daily volume, $\sigma$ is the stock's daily volatility, and $c$ and $\alpha$ are free parameters to be estimated from data — $\alpha$ around 0.5 (a square-root law) is a common empirical finding, but the exact value and the scale factor $c$ vary by asset class, venue, and time period, and have to be estimated rather than assumed. Calibration means collecting a large sample of the firm's own past executions — order size, participation rate, realized impact measured against a clean pre-trade benchmark — and fitting $c$ and $\alpha$ (via regression, typically on log-transformed variables since the relationship is a power law) to minimize the gap between predicted and realized impact across that sample.

The practical difficulty is separating a trade's own impact from everything else moving the price at the same time — market-wide moves, news, other participants' unrelated flow. A calibration exercise needs enough trades, spread across enough different market conditions, that this noise averages out and the systematic relationship between order size and impact becomes estimable; too small a sample and the fitted parameters are just noise dressed up as a model.

Worked example: fitting a simple power law

Historical data (simplified) shows three participation-rate buckets and their average realized impact in basis points: at $Q/V = 1\%$, impact is 8 bps; at $Q/V = 4\%$, impact is 16 bps; at $Q/V = 9\%$, impact is 24 bps. Taking logs and fitting $\log I = \log c + \alpha \log(Q/V)$ (holding volatility roughly constant across the sample for simplicity), the ratio of impacts, $16/8 = 2$ against a size ratio of $4/1=4$, gives $\alpha = \log(2)/\log(4) = 0.5$ — consistent with a square-root law. Checking the third point: predicted impact at $Q/V=9\%$ is $8 \times (9/1)^{0.5} = 8 \times 3 = 24$ bps, matching the observed 24 bps exactly, confirming $\alpha \approx 0.5$ and $c$ calibrated so that 1% participation predicts 8 bps of impact for this stock.

What this means in practice

An uncalibrated impact model, even with a theoretically sound functional form, can be badly wrong for a specific firm's flow, order types, or venues — and a badly wrong impact model leads directly to badly sized and badly scheduled orders, either trading too aggressively (unexpected cost) or too passively (unnecessary opportunity cost). Firms with serious execution operations re-calibrate periodically, since liquidity conditions, spreads, and volatility regimes shift over time, and a model fit to last year's market can drift out of step with this year's.