Order Book Depth and Its Predictive Power for Intraday Momentum

A five-minute bar closes on its high with volume well above the session average. The range expanded, the close sits in the top eighth, and every screen built on OHLCV agrees that momentum is present.

That bar is an accurate record of what traded. It carries no record of the resting size that was still displayed and never got hit. By the time it printed, the offers above the last transaction may have been lifted, or pulled without ever being touched, or replaced by size an order of magnitude smaller than what stood there ninety seconds earlier. The chart reports continuation. The book, had anyone looked, may already have been describing a market with nothing left standing in front of the move.

What the bar keeps and what it discards

OHLCV is a summary of consummated prints. Four price extremes and a volume total encode several thousand events, and the discarded portion is systematically the passive side: the resting size that determined how expensive the aggression was, dropped from the summary precisely because it did the work of absorbing rather than the work of printing.

Depth carries the complementary record. A book snapshot describes displayed interest at prices that have not yet printed, which makes it a statement about occupancy rather than outcome. Intraday momentum interacts with that occupancy on every tick, which is why two bars with identical OHLCV can sit in front of two very different books.

Occupancy is the frame that makes this usable. Treat the levels above and below the last trade as space that is either occupied or vacant. A move through occupied space pays for every level it takes and leaves behind displayed size that has already been tested. A move through vacant space travels cheaply and arrives somewhere with nothing underneath it. On the chart, both produce the same candle.

Four reads the book supports

The following is the standard vocabulary available to any operator with a depth feed, and it is what the rest of this piece uses.

• Imbalance

  describes the asymmetry of displayed occupancy: resting size on one side measured against the other across some band of levels near the touch. A move running against a heavily occupied opposite side is doing different work from the same move running against a thin one.

• Thinning

  measures the rate at which resting size leaves a side without trading against it. Cancellations are information. When the follow-through side empties while price is still climbing, displayed size on that side has left without trading, and the book has changed without printing a single contract.

• Refresh

  measures replenishment. A level that is lifted and immediately restored to comparable size behaves like a wall; a level restored at a fraction of its prior size behaves like a retreat. Refresh separates absorption from exhaustion, and the tape displays both as the same aggressive volume.

• One-sided vanish

  is the fast case: withdrawal across several levels of a single side, largely without corresponding prints. This is the condition under which an ordinary-looking bar becomes a gap.

Each of these describes a state of the book. Treating them as states, rather than as standalone signals, governs everything downstream.

Where the read belongs in a system

The temptation with depth is to compress it into a single number and trade the number. That path produces a fragile artifact: a threshold fitted to one venue, one session type, and one liquidity regime, which then decays without announcing that it has decayed.

A public-safe system would give depth three architectural positions instead, none of which require the read to be correct on its own.

1. The first is regime.

   Depth conditions describe what kind of session is running. A book that stays occupied and refreshes through aggression is a different operating environment from one that empties on the first push, and strategies calibrated in the former are being run out of context in the latter. Depth state can determine which regime a system is even allowed to treat as current before any signal is evaluated.

2. The second is confirmation.

   Confirmation qualifies a trade the primary logic already proposed, using information the primary logic structurally cannot see.

3. The third, and the most defensible, is veto.

   A system that declines to act when the depth state contradicts the bar state surrenders some participation and avoids a specific failure: entering a move that has already consumed the liquidity that would have carried it. Veto authority is easier to validate than entry authority because the failure mode is bounded and the abstention leaves a reviewable record of why the system stood down.

In all three positions, depth functions as a second and structurally independent observation of the same moment, used to constrain what the first observation is permitted to do.

Where the read breaks

An operating model that omits its failure modes is incomplete, and this one has four worth naming.

• Resting size

  is a display, and displays can be placed with no intention of being filled. Posted size can appear and withdraw as price approaches it. A model that treats displayed size as committed size will be fed exactly what it is looking for.

• Non-displayed liquidity

  moves the error in the other direction. Iceberg orders and midpoint interest mean the visible book is a lower bound on size that can actually absorb. A level that looks vacant can absorb considerably more than it advertises, which biases thinning reads toward false positives in venues carrying significant hidden flow.

• Fragmentation

  limits the claim. A single venue’s depth is a partial view of an instrument traded across many, and consolidating depth is its own engineering problem with its own timing assumptions. A single-venue read should be labeled as one, rather than quietly promoted into a statement about the market.

• Staleness

  bounds everything else. Depth can go stale in the time it takes to act on it, and a snapshot that is already old may describe a book that no longer exists. Every depth-derived state needs an explicit validity period, and a system that cannot date its read should abstain rather than act on it.

These four argue for the same architecture. Depth belongs inside a system that can decline to act, rather than in front of one that must produce an answer.

The public model and the private specification

The extraction model described here is portable and public: read the book as occupancy, watch imbalance, thinning, and refresh, and use the resulting state to set regime, confirm a proposal, or withhold authorization. An operator can build that from published microstructure concepts and a depth feed.

The specification layer stays private, and it is where the actual work sits. Which levels are measured and how they are weighted, the intervals over which thinning is judged, the conditions that move a system between regimes, how flicker is filtered, how venues are combined, and how heavily the depth read counts against everything else in the stack — those are implementation decisions, and implementation is where an edge either lives or dies.

Three explicit non-claims. This piece does not report a shipped AlphaFlux Level-2 momentum predictor. It reports no performance, hit rate, or backtest result of any kind. It is not investment advice.

The test worth carrying

The bar leaves that question open. The book answers it imperfectly, inside a short validity period, and that imperfect answer remains more than the compression on the chart can provide. The discipline is in treating it as a constraint on action rather than a reason for it.

Disclosure note