Microstructure Noise and Bid-Ask Bounce

Tick-level predictability often reflects how trades print through the spread, not an exploitable signal about the efficient price.

The Intuition

At very short horizons, observed trade prices are noisy measurements of an underlying efficient price.

The reason is simple. Trades do not always occur at the same point relative to value:

• aggressive buys print near the ask
• aggressive sells print near the bid
• the midprice may barely move while trade prices alternate mechanically

That alternation creates apparent mean reversion in last-trade returns even when the efficient price is flat.

This is bid-ask bounce. It matters because:

• naive tick returns exaggerate noise
• ultra-short-horizon labels can look predictable for purely mechanical reasons
• executable markouts differ from midprice markouts
• short-horizon volatility estimates can be contaminated by spread effects

A Minimal Model

Write the observed trade price as

$$ p_t = v_t + \frac{c_t}{2} s_t + \eta_t, $$

where:

• $v_t$ is the latent efficient price
• $c_t$ is the bid-ask spread
• $s_t in {+1,-1}$ is trade direction
• $\eta_t$ captures additional microstructure noise

If the efficient price $v_t$ is roughly flat and trade direction alternates +1, -1, +1, -1, then trade prices oscillate between bid-side and ask-side prints. Under that stylized setup, the return series shows negative first-order autocorrelation even though there is no true mean reversion in value.

That is the core statistical intuition. In real markets, serial correlation in order flow can partly offset this pure-bounce effect, so the stylized mechanism should be read as a common component, not as a universal law.

Why Tick Returns Can Mislead

Suppose the midprice stays at 100.05 and the spread is 10 cents:

• ask = 100.10
• bid = 100.00

Now trades print:

100.10 -> 100.00 -> 100.10 -> 100.00

The last-trade return series looks like alternating gains and losses. A naive model may interpret that as a short-horizon reversal signal.

But nothing informational happened. The print process simply bounced between sides of the spread.

That is why Chapter 3 warns that very short-horizon "predictability" may be non-tradeable.

Midprice Returns Versus Trade-Price Returns

Quoted-midpoint returns often reduce part of this mechanical bounce:

$$ r_t^{\text{mid}} = \log m_t^{\text{quote}} - \log m_{t-1}^{\text{quote}}, \qquad r_t^{\text{trade}} = \log p_t - \log p_{t-1}. $$

These are not interchangeable.

• $r_t^{trade}$ reflects what printed
• $r_t^{mid}$ reflects how the quoted center moved

Neither is universally "correct." The right choice depends on the question.

• if you study short-horizon price formation, midprice often gives a cleaner state variable
• if you study realized execution outcomes, trade prices and executable markouts matter

The mistake is to use last-trade returns as if they were a noise-free target. Quoted midpoints are cleaner for this purpose, but they still contain noise from discreteness, quote revisions, and stale or flickering quotes.

A Tiny Simulation

This toy simulation shows the mechanism with a flat efficient price:

$python import numpy as np mid = np.full(8, 100.05) sign = np.array([1, -1, 1, -1, 1, -1, 1, -1]) trade = mid + 0.05 * sign trade_ret = np.diff(np.log(trade)) mid_ret = np.diff(np.log(mid)) # trade_ret alternates mechanically; mid_ret is identically zero $

mid_ret is zero throughout. trade_ret alternates mechanically.

That gap is the whole issue.

Consequences for Labels and Features

This matters most when people build short-horizon targets or flow features.

Bad label: predict next-tick last-trade return and treat accuracy as evidence of exploitable alpha. Better label: use quoted-midpoint change over a horizon less dominated by spread bounce, executable markouts based on bid/ask at decision time, or horizons long enough that mechanical alternation is not the whole target.

Similarly, order-flow or imbalance features can correlate with next-tick trade returns partly because both are tied to who crossed the spread, not because the signal predicts efficient-price movement.

Why This Is Not Just an Academic Detail

Suppose a model predicts that the next trade-price return will reverse. That can look impressive in backtests using print-to-print labels.

But if execution happens by crossing the spread, the model may be monetizing a pattern that disappears once:

• entry and exit are marked at executable prices
• transaction costs are included
• labels are moved from print space to tradeable markout space

So microstructure noise is not merely a nuisance for statisticians. It is one reason naive high-frequency alphas vanish in live trading.

A model that predicts quoted-midpoint movement is also not automatically profitable if execution still requires paying the spread to enter and exit.

In Practice

Use these rules:

• inspect first-order autocorrelation separately for trade-price and midprice returns
• treat next-tick return predictability with suspicion
• align target construction to the executable object you actually care about
• compare print-based and quote-based labels before trusting a short-horizon model
• remember that spread and signing effects can dominate the smallest horizons

Common Mistakes

• Treating last-trade returns as if they were clean measurements of value.
• Calling bid-ask bounce "mean reversion" without checking midprice behavior.
• Evaluating high-frequency models on print labels while claiming tradeable performance.
• Using ultra-short-horizon volatility estimates without considering spread contamination.
• Assuming that mechanical predictability survives transaction costs.

Connections

This primer supports Chapter 3's discussion of bid-ask bounce, bar construction, and tradability. It connects directly to trade signing, markout design, and later chapters that build labels and features from noisy market data.