Order Book Features For Machine Learning

A raw book snapshot — every price and size at every level, on both sides — is dozens of correlated numbers that change on every message. Feeding that directly into a model wastes most of its capacity re-learning that "bid size at level 1" and "bid size at level 2" move together. Feature engineering compresses the book into a smaller set of numbers that each capture a distinct, economically meaningful signal.

The core features, computed on a real book

Book: bid 250 @ 99.98, offer 180 @ 99.99, next levels bid 400 @ 99.97, offer 220 @ 100.00.

Spread: $99.99 - 99.98 = 0.01$ — how expensive it is to trade immediately.

Order-flow imbalance at touch: $I = \frac{Q^b}{Q^b+Q^a} = \frac{250}{250+180} = 0.581$ — more resting size on the buy side, a mild lean toward price rising (see The Microprice for why).

Multi-level imbalance (using two levels each side): total bid size $250+400=650$, total offer size $180+220=400$, giving $I_2 = 650/1050 = 0.619$ — slightly stronger than the touch-only reading, suggesting the buy-side lean isn't just a touch-level fluke.

Book slope (how fast size grows moving away from the touch): bid size roughly doubles from level 1 to level 2 (250→400, +60%), offer size barely grows (180→220, +22%) — a steep bid slope relative to the offer slope again points the same direction, toward buy-side support.

Each of these is a single number, cheap to compute on every book update, and together they capture most of what a deep model trained on the raw levels would otherwise have to rediscover on its own.

Where it's used. These features are the standard inputs for short-horizon mid-price direction models — predicting whether the next move is up, down, or flat over the next few hundred milliseconds to a few seconds. They're also used as inputs to fill-probability models for resting limit orders, and as sanity-check features alongside end-to-end deep models trained on raw book images, which tend to rediscover imbalance and slope as their most important learned signals anyway — see Deep Learning On Limit Order Books.