Order Book Imbalance: Building a Strategy on Order Flow

The basic OBI idea — buy when bids outweigh asks — is well known and does not work. The reasons it does not work are specific and fixable, and each fix is a design decision rather than a parameter to optimise. This is the practitioner treatment; what OBI is covers the definition.

• Depth window must be fixed before anything else — 5 levels and 20 levels can disagree in sign.
• Thresholds must be per-instrument percentiles, not a global constant.
• Require persistence: an imbalance that survives several seconds is a different object from an instantaneous spike.
• Confirm with executed flow. The book states intent; the tape states what happened.
• The cost check is the gate: at a 0.3% stop and 0.15% round-trip, you pay 0.5R per trade before you are right about anything.

Why the naive version fails

"Go long when OBI exceeds +0.4" fails for four separate reasons, and they compound:

1. The signal is instantaneous. A one-snapshot reading captures a book state that may not exist a second later.
2. The threshold is arbitrary. +0.4 is common on a thin book and rare on a deep one. A single constant across a universe means you are effectively trading a different signal on every instrument.
3. Resting orders lie. Cancellation is free and spoofing is real, so the largest imbalances are the most suspect.
4. The costs are brutal. OBI's predictive horizon is short, which pushes you toward tight stops, and tight stops carry the highest cost per unit of risk.

Take the last one seriously before building anything. Round-trip cost in R is:

cost_in_R = round_trip_fraction × entry_price / |entry_price − stop_price|

At 0.15% round trip with a 0.3% stop — an ordinary OBI-scale stop — you pay 0.5R per trade in fees alone. Your gross expectancy must exceed 0.5R before you make a cent. Most short-horizon book signals do not come close, and this is the single most common reason an OBI strategy that backtests well loses money live.

Fix 1: choose a depth window and never vary it

Depth is the highest-leverage decision in the whole design and the one most often left implicit.

For a strategy spanning several instruments, a percentage band is materially better than a level count. Twenty levels on an instrument with a coarse tick covers a much wider price region than twenty levels on a fine-tick instrument, so a level-based window silently measures different things on different symbols.

Fix 2: calibrate thresholds per instrument

Replace the global constant with a percentile of that instrument's own recent distribution. Collect OBI readings over a rolling window and define your trigger as, for example, the 90th percentile.

This makes the signal self-normalising. On a stable deep book the 90th percentile might be +0.25; on a volatile thin one it might be +0.65. Both mean the same thing — unusually imbalanced for this instrument right now — which is the property you actually wanted.

Fix 3: require persistence

An imbalance present in one snapshot and an imbalance sustained across several seconds are different objects. The first is often a single large order about to be cancelled; the second is a book that has genuinely repriced.

Require the threshold to hold across N consecutive observations, or use a short moving average of the reading rather than the instantaneous value. This removes most spoof-driven triggers at the cost of entering slightly later — a trade worth making, since the signals it removes are disproportionately the fake ones.

Fix 4: confirm with executed flow

The single most valuable addition. The book states intent; the tape states what actually happened.

Compare resting imbalance against aggressive flow — the volume that traded into the bid versus into the ask over a recent window. Four combinations:

Requiring agreement between the two removes a large share of spoof-driven signals, because manufacturing resting depth is cheap while manufacturing executed volume is not.

Three strategy shapes

Continuation with a trend filter

Take strong bid imbalance only when the higher-timeframe trend is up. Enter on the imbalance, exit on a fixed target or when the imbalance normalises. The trend filter removes the counter-trend trades that dominate the loss distribution.

Absorption reversal

Heavy aggressive selling into a bid wall that does not break is a defended level. If the seller exhausts and price holds, the resulting bounce can be sharp. Requires the persistence and flow-confirmation machinery above; without it you are catching knives.

Imbalance divergence

Price making higher highs while the imbalance deteriorates means the advance is being carried by aggression rather than by resting demand. The slowest of the three and the least sensitive to spoofing, because it depends on a trajectory rather than on a level.

Venue differences are not noise

The same asset has a different book on every venue. Depth, tick size, participant mix and how much of the book is displayed all vary, so an imbalance on one exchange may be absent on another.

Two implications. First, calibrate thresholds per venue, not just per instrument. Second, cross-venue disagreement is itself a signal: an imbalance visible on one venue and nowhere else is more likely to be a single participant than a market-wide repricing.

The order-book APIs and their WebSocket diff-stream semantics differ meaningfully between the three major venues — including how sequence gaps must be handled to keep a local book correct. See the API comparison.

Testing an OBI strategy honestly

Book-based strategies are harder to backtest than candle strategies, and the ways they mislead are specific:

• You need book snapshots, not candles. Reconstructing historical books requires stored depth data. Without it there is nothing to test.
• Your own order changes the book. At meaningful size you are part of the imbalance you are measuring.
• Latency is part of the strategy. A signal acted on 500ms later is a different strategy, and on a seconds-horizon signal that difference is decisive.
• Fills are optimistic. Assuming a fill at the touch when the book is thin overstates results in exactly the conditions the strategy targets.
• Charge the costs. With the cost-in-R arithmetic above, this is not a refinement — it usually decides the verdict.

Where OBI fits

OBI is a strong confirmation layer and a weak primary signal. Its best use in most systems is as a gate on an entry generated elsewhere: take the structural setup only when the book is not leaning against it. That formulation captures most of the value while avoiding the trade frequency — and therefore the cost — that sinks OBI-as-trigger strategies.

TradeFloor computes OBI across all three connected venues and exposes both the level and OBI divergence as trigger conditions, combinable with up to four other conditions in a single rule, so the confirmation-layer pattern is expressible directly.

// questions or corrections · [email protected] · more essays · /blog