Order Book Feature Engineering

Question

You are building a model that predicts the direction of the next mid-price move over the next few seconds, using the state of the limit order book.

Propose at least five features you would compute from the order book and recent trades, explain the economic intuition behind each one, and discuss how you would normalise them so that a single model can be used across many different instruments.

Accepted Answer

The order book is the natural feature source at short horizons, because it displays the standing supply and demand. Good features compress it into stationary, comparable numbers. A standard starter set:


 	Book imbalanceWith $Q_b$ the resting size at the best bid and $Q_a$ at the best ask: \begin{equation} I = \frac{Q_b - Q_a}{Q_b + Q_a} \in [-1, 1] \end{equation} If far more size rests on the bid than on the ask, the ask queue is more likely to be eaten through first, so the next mid move is more likely to be up. Imbalance is the single most documented short-horizon predictor in the microstructure literature — and note that it is already normalised by construction.
 	MicropriceThe mid-price ignores where the size sits; the microprice weights each price by the opposite side's queue: \begin{equation} P_{micro} = \frac{P_a Q_b + P_b Q_a}{Q_b + Q_a} \end{equation} When the bid queue is much larger, $P_{micro}$ leans towards the ask — the "fair" value is higher than the mid. The gap $P_{micro} - P_{mid}$, expressed in fractions of the spread, is a clean directional feature.
 	Bid-ask spread (in ticks). A regime variable rather than a directional one: a wide spread signals uncertainty, thin liquidity or information asymmetry, and changes how every other feature should be interpreted — which is exactly what interaction-capable models (trees) exploit.
 	Trade flow imbalanceSigned aggressor volume (buys minus sells) over a short trailing window, ideally as an EWMA. Aggressive orders pay the spread to trade now, which is the classic fingerprint of informed or urgent flow; recent net flow tends to predict continuation.
 	Depth beyond the top of bookCumulative size at the first few levels and the "slope" of the book on each side. A side that is shallow deeper down is easier to push through; a thick wall absorbs flow.
 	Activity and volatilityrealised volatility of the mid over a trailing window, message/trade arrival rates, time since the last trade, and the rate at which the best queues are being depleted or replenished.

Normalisation — the part that makes one model work everywhereRaw prices and raw sizes are meaningless across instruments: 500 shares is huge for one stock and nothing for another, and a 1-cent move is large for a $10 stock but noise for a $1000 one. The rules:

 	Never use raw price levels as features — use returns, differences in ticks, or distances measured in spreads or in units of recent volatility. Levels are non-stationary; differences are approximately stationary.
 	Express sizes relativelyas ratios (the imbalance above), or scaled by a rolling average of depth or by average daily volume.
 	Standardise per instrument with rolling statistics — z-score each feature using means and standard deviations computed strictly from past data (full-sample scaling is lookahead bias, as covered in the backtesting question).
 	Prefer dimensionless constructions from the start (ratios, signs, percentile ranks): they pool naturally across instruments, which multiplies the effective training data per model parameter.

Two closing remarks that score points. First, consider event time: sampling features per book update or per trade, rather than per wall-clock second, makes their distributions far more stable between quiet and busy periods. Second, every feature must be computable incrementally — an O(1) update per market data event (the EWMA pattern), not a recomputation over a window — because at production latencies you do not get to loop over history. And a word of caution: features built from resting orders can be manipulated (spoofing), so robust versions cap the influence of any single order and weight by queue position or order age.

Title	Category	Subcategory	Difficulty	Status
Encoding Categorical Variables (and the Target-Encoding Trap)	Machine Learning	Feature Engineering	Medium	Unlock
Handling Missing Values	Machine Learning	Feature Engineering	Easy	Unlock

The Interview Guide for Quants and Traders

Order Book Feature Engineering

What do you think of this page?