Kyle's lambda: liquidity as a slope, not a spread

August 14, 2026

The bid-ask spread tells you the cost of a small trade. Kyle's lambda tells you what happens when the trade is large: how far price moves per unit of order flow. It is the cleanest single measure of market depth, and it is estimated with a regression anyone can run.

Ask how liquid a market is and most answers reach for the bid-ask spread. The spread is a real cost and it answers a narrow question: what does it cost to trade a small amount right now?

It says nothing about the question that matters when size is involved: if I need to buy a meaningful quantity, how far will I push the price?

That is what Kyle's lambda measures.

Where it comes from

Albert Kyle's 1985 model describes a market with an informed trader, uninformed noise traders, and a market maker who cannot tell them apart. The market maker sees only total order flow and must set prices from it.

Their rational response is to move price in proportion to net order flow. Heavy net buying might be an informed participant, so quoting higher protects against being run over. The constant of proportionality is lambda:

`` price change = lambda x net order flow ``

Lambda is therefore the price impact per unit of signed volume, and it is the inverse of market depth. High lambda means a thin book where modest size moves price a long way. Low lambda means a deep book that absorbs size.

The elegance is that a single number captures what an order book snapshot only gestures at, and unlike a depth snapshot it reflects what actually happened when size arrived rather than what was displayed.

Estimating it

The estimator follows directly from the equation. Over a window of trades:

1. Sign each trade's volume by aggressor, positive for buyer-initiated and negative for seller-initiated, exactly as for CVD. 2. Compute the price change across each interval. 3. Regress price change on signed volume.

Lambda is the slope of that regression. In practice that is the covariance of price change and signed flow divided by the variance of signed flow.

Two implementation notes that matter more than they look.

Intervals with zero signed flow should be excluded, not treated as zeros. They carry no information about impact and including them biases the slope toward zero.

Be consistent about degrees of freedom. Standard covariance functions typically use the sample convention while standard variance functions may default to the population convention. Mixing them scales lambda by a factor of (n-1)/n. On a few hundred observations that is under one percent, which is small enough to survive review indefinitely and still be wrong.

Reading it

Lambda's units are price per unit of volume, which means its absolute value is instrument-specific and not comparable across symbols. An index future and a thinly traded commodity have lambdas that differ by orders of magnitude purely because of contract size, tick value and typical volume.

So the same rule as VPIN applies: read it as a percentile against the instrument's own history. "Lambda is 0.0004" is not a statement anyone can act on. "Lambda is at the 92nd percentile of its own three-month distribution" is.

What rising lambda tells you is that the book has thinned relative to normal for this instrument. The same order that was absorbed yesterday moves price further today. That is useful regardless of direction, and it is often the most actionable microstructure reading available, because it speaks directly to execution.

Where it is genuinely useful

Execution and sizing. If impact per unit is elevated, the same intended position costs more to establish. That argues for working the order more patiently, or for reducing size, and it is a decision lambda can inform directly.

Regime awareness. Liquidity withdrawal frequently precedes and accompanies volatile conditions. A thinning book is not a directional forecast, but a market that has become easy to move is a market where stops are more likely to be reached and where a given catalyst produces a larger response.

Interpreting a level break. A break through a level on elevated lambda took less flow to achieve than the same break on low lambda. That is real context about how much conviction the move required.

The limits

It is a window estimate, so it depends on window length and it is noisy on short samples. It describes recent impact, not a structural constant.

It inherits the trade signing. If aggressor classification is wrong, the regression is fitted to a mislabelled variable, and a systematically inverted sign convention would produce a negative lambda, which is economically meaningless and a useful red flag.

It is not directional. Lambda is a magnitude of responsiveness. It says nothing about which way price goes, only how far a given imbalance pushes it.

The summary

The spread prices immediacy for small size. Lambda prices impact for real

size. They are different costs, and only one of them scales with your

position.

Two markets can show identical spreads while one absorbs a thousand contracts without flinching and the other gaps on two hundred. The spread cannot tell them apart. Lambda can, which is why it belongs in any serious read of liquidity.