Queue Imbalance As A Predictor

At any instant, the best bid and best offer each have a queue of resting orders behind them — some number of shares waiting to buy at the bid, some number waiting to sell at the offer. Queue imbalance simply compares the two sizes. It turns out to be a remarkably effective, remarkably cheap predictor of which way the price ticks next, and it's one of the first signals every market-making and high-frequency desk builds.

The intuition: a tug of war at the front line

Picture two queues at a single door, one queue of people trying to push their way in, one trying to push their way out, and the door only opens when one queue's pressure exceeds the other's. If 900 people are pushing in and only 100 are pushing out, it's a reasonable bet the door swings inward next. Queue imbalance applies exactly this logic to the limit order book: a much bigger queue on the bid than the offer suggests more resting buying interest is stacked up, making an uptick — the price at the offer getting taken and a new, higher price becoming the best bid — statistically more likely than a downtick.

The formula, one symbol at a time

Let $Q^{bid}$ be the total resting size at the best bid and $Q^{ask}$ the total resting size at the best offer. Queue imbalance is:

$$I = \frac{Q^{bid} - Q^{ask}}{Q^{bid} + Q^{ask}} .$$

In plain English: this is the same normalized-imbalance construction used for trade flow, but applied to the resting book instead of executed trades — it ranges from $-1$ (offer queue completely dominates) to $+1$ (bid queue completely dominates), and empirically, higher values of $I$ are followed by a higher probability that the next price move is up rather than down.

Worked example: converting imbalance into a probability estimate

Suppose historical data on a liquid stock shows: when $I$ is near $+0.8$ (bid queue roughly nine times the offer queue), the next price move is up about 65% of the time; when $I$ is near 0 (balanced book), the next move is roughly a coin flip, about 51% up; when $I$ is near $-0.8$, the next move is up only about 35% of the time.

Right now the book shows $Q^{bid} = 4{,}500$ shares and $Q^{ask} = 500$ shares:

$$I = \frac{4{,}500 - 500}{4{,}500 + 500} = \frac{4{,}000}{5{,}000} = 0.80 .$$

Based on the historical relationship, this reading puts the estimated probability of the next tick being up around 65% rather than 50% — a real, exploitable edge on any single prediction, though a modest one, and the queue can refill or empty within milliseconds, so this estimate has a very short shelf life.

The binomial distribution above is the right frame for queue-imbalance predictions: each individual tick is a probabilistic coin flip tilted by the current imbalance, not a certainty — profitable use of the signal comes from making the same tilted bet many thousands of times, not from any one prediction being reliable.

What this means in practice

Queue imbalance is cheap to compute (it needs only the top of book, not the full depth or trade tape), updates every time the book changes, and has one of the best-documented short-horizon predictive relationships in the microstructure literature — which is exactly why market makers use it constantly to skew their own quotes, and why its edge, like most easily observed signals, is thin and gets competed away quickly by faster participants. It predicts the direction of the next small tick, not the size of any subsequent move, and it says nothing about horizons beyond a few seconds.