Estimating impact

Calibrating an impact model from your own fills and public TCA data, and the traps (selection bias, regime change) that make naive estimates flatter you.

The idea

How do you measure impact at all?

You measure impact as price displacement: the change in a reference price (the mid) from just before your order starts to a chosen point during or after it, signed by your trade direction. Aggregate many such measurements and you can fit how displacement scales with size. But that only works if you first define your windows and your reference price carefully.

Impact is invisible on any single trade, because the price was going to move anyway: noise, plus whatever signal you traded on. It becomes visible only in aggregate, averaged over many trades, where the random moves cancel and the systematic, direction-correlated move (the impact) remains. So the whole procedure is built around three primitives: a reference price, a signed displacement, and a set of measurement windows.

Sample that displacement at two windows. The during-trade window captures peak displacement: temporary plus permanent impact together. A later post-trade decay window (minutes to hours, depending on the instrument) is taken after temporary impact has largely relaxed, leaving the permanent residual. The unit of analysis is the metaorder (a parent decision worked as many child orders), not the individual fill: impact is a property of the whole order's footprint, and the square-root law on market impact is stated in metaorder terms.

Everything here depends on knowing the sign of trades. Yours are known, but to build the volume $V$ and the surrounding flow you usually need to classify other trades' signs (the tick rule, or Lee–Ready 1991), which imports every accuracy caveat from trade-sign inference.

Regressing displacement on signed volume

With displacements collected, you fit a functional form. Regress measured displacement on signed size to estimate the coefficient: a linear fit recovers Kyle's lambda for small orders; a fit against signed $\sqrt{Q/V}$ recovers the square-root-law constant $Y$. The fit's slope is your calibrated impact coefficient, and the post-trade window's slope is your permanent coefficient.

Plot how far the price moved against how much you traded, and fit a line: straight for small orders, square-root for large ones. The linear form (Kyle 1985, small orders or the permanent component) sets displacement $\approx \lambda \cdot (\text{signed order flow})$; the fitted slope $\lambda$ is Kyle's lambda, cleanest for the permanent shift and marginal single-order impact. The square-root form (Gatheral; Almgren et al. 2005, large metaorders) is the one that feeds the impact model and Almgren–Chriss.

Then condition the regression. A single global $Y$ is a starting point; the real edge is conditional impact, letting $Y$ (or $\lambda$) depend on book depth at arrival, recent signed flow, spread, time of day and volatility regime. This is where ML earns its keep: a learned, state-dependent impact function beats a constant, letting you size more aggressively when the book is deep and back off when it is thin.

Why estimating impact is genuinely hard

Impact estimation is hard because impact is small relative to price noise, and because of an identification problem: the price often moves with your trade for reasons your trade did not cause, because you traded on a signal others also acted on. Disentangling impact-you-caused from drift-you-predicted is the central difficulty, and naive regressions over-attribute drift to impact.

Signal/impact confounding, the big one. If you buy because you forecast the price will rise, the price rises partly because you were right, not because you traded. A regression of displacement on your signed volume then over-states impact, charging your cost model for moves your alpha produced. Mitigations: measure on uninformed flow where available (execution-only orders with no alpha), include the signal as a control, or use the post-decay permanent residual, which better isolates the information your order leaked from the broader move.

Tiny signal, huge noise. Impact is a few basis points; the mid wanders tens of basis points over any decay window. The estimator works only in aggregate and converges as $1/\sqrt{N}$, so small desks simply lack the trade count for a tight $Y$, itself a reason the data is valuable. Endogeneity of size and timing. Smart traders trade more when liquidity is deep, so size correlates with the very conditions that reduce impact, biasing a naive fit toward lower impact than you will actually pay when forced to trade in bad conditions.

Trade-sign error from Lee–Ready or the tick rule (a few percent of trades) propagates into the regressors; see trade-sign inference. Regime dependence: $Y$ and $\lambda$ drift with volatility regime, tick-size changes and venue structure, so a coefficient fit on last year's data can be wrong today; date your calibration and refit. And cross-impact: in correlated instruments, trading one moves others; a single-name model misses it, and the cross-impact matrix is far harder to estimate (more parameters, the same noisy data).

What data you need to do this

To estimate impact you need your own fills time-stamped and aligned to the book state at execution, the trade-and-quote stream around each order to compute displacement and the reference mid, and enough metaorders for the $1/\sqrt{N}$ estimator to converge. Clean, reconstructable order-book data with reliable trade signs is the binding constraint, and the resource most people cannot get cleanly.

The minimum viable dataset is: your fills, each with venue timestamp, side, size and price, aligned to the mid at arrival and at each decay window; TAQ / L2 quotes around every order to reconstruct the reference mid and the post-trade decay path (ideally L3/MBO to control for book depth, the best conditioning feature); signed volume for $V$, meaning a trade-sign classifier and its error budget unless you have native sign data; and a clean volatility $\sigma$ per instrument per window, mindful that the bid-ask bounce inflates naive $\sigma$.

This is exactly where the cost guides meet the HF-data guides: how the stream is recorded lives there. The honest 2026 note: this is the step where good data is a durable advantage. The model is public; the calibration is private, and the private calibration is only as good as the aligned data behind it. Most public or retail data cannot recover the book state at your fill, which is why this estimation is so hard to do well outside a desk.

Worked example

A back-of-envelope calibration, illustrative and as of 2026, to show the shape of the procedure, not a real fit. Collect 2,000 of your own buy/sell metaorders in one liquid name with daily volume $V = 5{,}000{,}000$ shares, daily volatility $\sigma = 150$ bps and price \$50. For each, record signed $\sqrt{Q/V}$ and the signed mid displacement from arrival, sampled at +5 minutes (peak, temporary plus permanent) and +2 hours (the permanent residual).

Fit total impact on the +5-minute window. Suppose the slope comes out $\hat{Y} \approx 0.6$. Sanity-check against market impact: a 1%-of-volume order has $\sqrt{0.01} = 0.10$, so it costs about $0.6 \times 150 \times 0.10 = 9$ bps, in the right ballpark.

Now the standard-error reality check. Per-order window noise is about $\sigma$ over 5 minutes, $\sqrt{5/390}\times 150 \approx 17$ bps, against a true signal of about 9 bps: a per-observation signal-to-noise of roughly 0.5. Over $N = 2{,}000$ orders the standard error on $\hat{Y}$ is roughly $17 / (\overline{|x|}\,\sqrt{2000})$, wide enough that you should report a confidence interval on $Y$, not a point estimate, and refit as more orders accrue.

The honest conclusion: your calibrated cost model is $\mathcal{I}(\text{bps}) \approx 0.6\,\sigma\,\sqrt{Q/V}$ with about a third permanent, but with a fat error bar, conditional on this regime, and only as trustworthy as your alignment of fills to the book and your control for the alpha that motivated the trades. A sharper impact estimate is, directly, a higher safe capacity. Numbers are synthetic; the procedure (windows, then signed-$\sqrt{}$-volume regression, then split temporary from permanent, then bound the error, then condition on book state) is the transferable part.