Nonparametric Impact Curve Estimation

The standard textbook story is that market impact scales with the square root of order size relative to volume — a tidy formula that's easy to write down and easy to fit with a single coefficient. But the true relationship for any particular stock, venue and time period might bend differently: flatter for small orders, steeper for large ones, or shaped in a way no simple formula captures. Nonparametric impact estimation drops the assumption that impact follows any pre-chosen functional form and instead lets the data trace out the curve.

Why not just fit the square-root law

Fitting $\text{impact} = c \times \sqrt{Q/V}$ (order size $Q$ over volume $V$, scaled by a constant $c$) is fast and interpretable, but it forces every stock and regime into the same shape. If the true curve is actually closer to linear for small trades and only curves upward near the top of the observed size range, a square-root fit will systematically overstate impact for small orders and understate it for large ones — exactly backwards from what a desk needs, since small-order estimates set day-to-day trading costs and large-order estimates set risk limits for block trades.

The nonparametric approach

Instead of picking a formula, bucket historical trades by relative size (say, into 20 bins by $Q/V$ percentile) and compute the average realized impact within each bin — this is binning, the simplest nonparametric estimator. A smoother version replaces hard bins with a kernel-weighted average: each point's estimate is a weighted average of nearby observations, with weight declining smoothly with distance in size, rather than a discontinuous jump at a bin boundary. Both let the curve bend however the data says it bends, at the cost of needing considerably more data than a two-parameter formula, since each local region of the curve is estimated mostly from observations near it rather than borrowing strength from the whole sample.

Worked example: bins reveal a kink

A desk buckets six months of parent orders into deciles by $Q/V$ and computes mean implementation shortfall (in basis points) per decile:

Decile ($Q/V$)	1 (smallest)	3	5	7	9	10 (largest)
Mean impact (bp)	2	5	9	15	26	45

Plotting these ten points shows impact rising roughly linearly through decile 7, then steepening sharply for deciles 9 and 10 — a kink a single square-root fit across the whole range would smooth away, understating the true cost of the largest orders by conflating them with the more gently-sloped middle of the distribution. A nonparametric fit — say, a kernel smoother or a shallow regression tree that's allowed to split near decile 8 — would capture that kink and give a much more honest large-order cost estimate, which matters directly for how the desk should schedule and price its biggest trades.

What this means in practice

Nonparametric impact curves are best used as a diagnostic and a sanity check on any parametric model a desk relies on for scheduling and cost estimation — if the binned data disagrees with the fitted formula in a particular size range, that's a sign the formula is misspecified there, not that the data is noisy. The tradeoff is real: nonparametric estimates need enough trades per bin to be stable, so they work best pooled across many similar names, and they don't extrapolate — a bin with no historical data near the largest sizes a desk might ever need to trade gives no answer at all, whereas a parametric formula (rightly or wrongly) always produces one.