Kyle's lambda (price impact of order flow)
Kyle's lambda measures how much prices move per unit of net buying. It is the slope of price changes regressed on signed order flow, and a low lambda means a deep, liquid market.
01The formula
- the price change over an interval
- signed order flow in that interval: shares bought minus shares sold by aggressive orders
- price impact: how far the price moves per unit of net order flow
- price moves not explained by order flow
- the estimate of λ from data: the regression slope
02See it move
03Where it is used
- Quant researcherEstimates lambda per stock and time of day as a measure of liquidity and of how informed the flow is.
- Execution traderUses lambda to predict how far a planned order will push the price before sizing or splitting it.
- Market makerWidens quotes when lambda rises, because it signals that order flow carries more information.
04Why traders use it
In Kyle's model market makers cannot tell informed from uninformed orders, so they move the price with net flow. Lambda is the price of that uncertainty and the core cost of trading size.
05Worked example
With a true λ of $0.050 per 1,000 shares and 200 intervals, the regression finds λ̂ = $0.0576 with R² = 72.5%. A 10,000-share buy would be expected to move the price by about $0.576.
06How it is built
- For each interval, record the net signed order flow and the price change .
- Fit a straight line by least squares.
- The slope is .
- In Kyle's one-period model the equilibrium value is : more uncertainty about value raises it, more noise trading lowers it.
07The common mistake
Using unsigned volume. Lambda needs the direction of each trade (buyer- or seller-initiated); total volume says nothing about which way the price should move.
08In Python and LaTeX
import numpy as np rng = np.random.default_rng(7) lam, su, se, n = 0.05, 8, 0.3, 200 x = su * rng.standard_normal(n) dp = lam * x + se * rng.standard_normal(n) print(np.cov(x, dp)[0, 1] / np.var(x, ddof=1)) # close to 0.05; the page's sample differs
\Delta p = \lambda\,x + \varepsilon, \qquad \hat\lambda = \frac{\operatorname{Cov}(\Delta p,\ x)}{\operatorname{Var}(x)}09Where it fits
Rather not code the maths yourself? Describe your trading idea and let AI build the bot.
Build a bot with AgenKitThe worked example and the model run the same code, so every number matches. Models simplify markets on purpose. This is education, not financial advice.