Little's Law

You run a matching engine and want to know, on average, how many resting limit orders sit in the book waiting to be filled at any given moment. You know how fast new orders arrive (say, 200 per minute) and, from historical fill data, how long the average order sits in the book before it's filled or cancelled (say, 45 seconds). It seems like it should take a detailed simulation of the whole arrival-and-departure process to figure out the average book size — but it doesn't. There's a relationship connecting these three numbers so simple it looks almost too good to be true, and it holds almost no matter what the arrival pattern or service-time distribution looks like.

An analogy: a busy restaurant

Picture a restaurant that, on average, seats 20 new customers per hour, and each customer stays for an average of 45 minutes (0.75 hours) from being seated to leaving. How many customers are in the restaurant, on average, at any given moment? You don't need to know anything about which customers stay longer or shorter, or whether Friday nights bring more variability than Tuesday lunches — multiply the arrival rate by the average time spent, and you get the average number present: $20 \times 0.75 = 15$ customers in the restaurant at any moment, on average. This works regardless of whether parties arrive steadily or in bursts, and regardless of whether meal times are consistent or wildly variable — as long as the restaurant is in a steady, unchanging pattern over the period you're averaging.

The law, one symbol at a time

Let $\lambda$ be the average arrival rate — how many items enter the system per unit time. Let $W$ be the average time spent in the system per item (from arrival to departure). Let $L$ be the average number of items in the system at any given moment. Little's Law states:

$$L = \lambda W .$$

In plain English: the average number of things present equals how fast things are arriving, multiplied by how long each one sticks around. This holds for essentially any stable system — any arrival process, any service-time distribution, any number of servers, any queueing discipline (first-come-first-served, priority, random) — as long as the system is in steady state, meaning the average arrival rate and average time-in-system aren't drifting over the window you're measuring (items aren't piling up faster than they leave, on average).

The proof, in plain English rather than symbols: track the total customer-minutes accumulated across everyone who passed through the system during a long time window of length $t$. You can compute that total two ways — count how many customers arrived ($\approx \lambda t$) and multiply by their average time each ($W$), giving $\lambda t W$; or, equivalently, average the number of customers present at any instant ($L$) across all $t$ minutes, giving $Lt$. Since both are the same total, $Lt = \lambda t W$, and the $t$ cancels, leaving $L = \lambda W$. This is a close relative of the renewal reward theorem (see the renewal reward theorem) — both work by equating two ways of summing the same total quantity over a long window.

Worked example 1: the order book by hand

Orders arrive at $\lambda = 200$ per minute, and the average time an order sits in the book before fill or cancellation is $W = 45$ seconds $= 0.75$ minutes. By Little's Law:

$$L = \lambda W = 200 \times 0.75 = 150 .$$

On average, 150 orders sit resting in the book at any given moment. No simulation of individual order arrivals or fill times was needed — just the two averages, multiplied. If you wanted to know how the book size would change if the exchange sped up matching so orders sat for only 30 seconds ($W = 0.5$ minutes) on average, the new steady-state book size is $L = 200 \times 0.5 = 100$ — a direct, immediate consequence with no further modeling required.

Worked example 2: solving for the unknown leg

A firm's risk desk tracks, on average, $L = 12$ open hedge positions at any given moment, and knows new hedges get initiated at a rate of $\lambda = 4$ per day. How long does the average hedge stay open? Rearranging Little's Law:

$$W = \frac{L}{\lambda} = \frac{12}{4} = 3 \text{ days} .$$

This is a common practical use of the law: two of the three quantities ($L$, $\lambda$, $W$) are often much easier to measure directly than the third. Here, counting how many hedges are open right now ($L$) and how many new ones get initiated per day ($\lambda$) is straightforward from a position log, while directly measuring "how long does a hedge stay open on average" requires matching up open and close events across the whole log — Little's Law gets you there with a single division instead.

What this means in practice

Little's Law is used throughout market microstructure and operations: estimating average resting order-book depth from arrival and cancellation-time statistics, sizing how many concurrent positions a strategy will typically hold given its trade frequency and average holding period, or capacity-planning a matching engine or execution system. Its power is that it needs no assumption about arrival bursts, fat-tailed holding times, or priority rules — the relationship $L = \lambda W$ holds regardless, as long as the system is in steady state. This also makes it a fast sanity check: if you've independently estimated all three of $L$, $\lambda$, and $W$ and they don't roughly satisfy $L = \lambda W$, something is wrong with one of your measurements or the system isn't actually in steady state over the window you measured.