How Accurate Are Trade Classification Rules?

Knowing the tick rule and quote rule exist is only half the story — the harder, more useful question is how often they're actually right. When exchanges occasionally release datasets that tag every trade with its true aggressor side, researchers can check classification rules against ground truth, and the answer shapes how much you should trust any order-flow signal built on top of inferred trade direction.

The analogy: a witness guessing who threw the first punch

Imagine a security guard who didn't see the start of a scuffle but has to guess who started it, based only on who's now standing closer to the door. Most of the time that's a decent clue — the aggressor usually ends up nearer the exit. But sometimes the guess is wrong: maybe the person who threw the first punch got knocked back, or the fight moved before the guard looked over. Trade classification rules are the same kind of after-the-fact guess, and their accuracy depends heavily on how "clean" the scene is when the guess gets made — a busy, chaotic room makes the guard's guess much less reliable than a quiet one.

What the accuracy studies actually find

Comparisons against ground-truth data — using NYSE TAQ trades matched to proprietary order records — consistently find the quote rule and Lee-Ready style combined rules classify roughly 80-90% of trades correctly in normal, liquid conditions. That sounds high, but it means one in five to one in ten trades is mislabeled, and the errors are not random noise: they cluster in specific, predictable situations.

$$\text{Accuracy} = \frac{\text{trades correctly classified}}{\text{total trades}}$$

In plain English: accuracy is simply the fraction of trades where the rule's guessed direction matches the true, known direction — and the number that matters isn't the headline average, it's how accuracy varies across conditions.

Accuracy is systematically worse for trades that print exactly at the midpoint, for trades in fast-moving or volatile markets where the recorded quote may already be stale by the time the trade prints, and for markets with sub-penny or very tight ticks where the tick rule has little price movement to work with. Accuracy is also worse for trades that occur off the primary exchange or through non-standard order types, since the quote used as the reference may not be the one the trade actually executed against.

Worked example: accuracy across two regimes

A researcher checks classification accuracy on a large-cap stock during a calm trading week and during an earnings-day volatility spike, using proprietary order data as ground truth.

• Calm week: quotes are stable, spreads are 1-2 cents, and the quote-rule classification matches true direction on 89% of trades. The 11% of errors cluster almost entirely at the midpoint, where the rule effectively has no information.
• Earnings day: quotes update dozens of times a second as the market digests news, and reported trade timestamps lag the true execution time by tens of milliseconds. Classification accuracy on this same stock drops to 76% — a substantial degradation exactly on the day when knowing who was buying and who was selling would be most valuable for understanding the reaction.

That gap matters practically: a signal built on classified order flow that looks reliable in backtests over calm periods can degrade sharply, and silently, during exactly the volatile episodes a trader most wants to understand.

Treat the boundary above as an analogy for a trade classifier's decision line between "buy" and "sell": in calm regimes the two classes are well separated and the boundary is easy to draw correctly; as volatility rises the classes blur together near the midpoint, and no boundary — however well designed — separates them cleanly.

What this means in practice

Anyone using classified trade direction as an input — for volume delta, order flow imbalance, or informed-trading estimates — should validate classification accuracy on their specific data and asset class rather than assuming textbook accuracy figures apply, and should expect that accuracy specifically during high-volatility periods. Where possible, use datasets with genuine aggressor flags instead of inferred labels; where that's not possible, treat classification error as an additional noise term layered on top of whatever signal the order flow itself carries.