Market Microstructure Noise

The defining statistical signature of microstructure noise is negative first-order autocorrelation in high-frequency returns. This is primarily driven by the bid-ask bounce, where transaction prices oscillate between the bid and ask quotes without any fundamental change in the asset's value. A trade at the ask followed by a trade at the bid creates an artificially negative return, which is immediately followed by a reversal. This induces a mean-reverting pattern in observed prices, making the noise component distinctly predictable over very short horizons, unlike the random walk of the efficient price.

Microstructure noise is fundamentally shaped by the minimum price increment (tick size) mandated by an exchange. Because prices can only exist on a discrete grid, the observed price is a rounded version of the underlying continuous efficient price. This rounding error introduces a uniformly distributed noise component bounded by the tick size. The variance of this noise is directly proportional to the square of the tick size, meaning that assets with larger relative tick sizes (e.g., low-priced stocks) exhibit proportionally higher levels of microstructure noise contamination.

The intensity of microstructure noise is not constant; it exhibits a strong deterministic intraday pattern. Noise variance is typically highest during the opening and closing auctions and immediately following macroeconomic announcements, periods characterized by elevated information asymmetry and order flow toxicity. It follows a distinct U-shape or L-shape curve throughout the trading day. This heteroskedasticity requires noise-robust estimators, like the Two-Scales Realized Volatility (TSRV) estimator, to adapt their sampling windows to the local noise level rather than applying a constant correction factor.

The signal-to-noise ratio is a direct function of the sampling interval. As the sampling frequency increases, the variance of the efficient price innovation scales linearly with time, while the variance of the microstructure noise remains constant. Consequently, at very high frequencies (e.g., tick-by-tick), the observed price variance is dominated entirely by noise. This is the bias-variance trade-off of realized volatility estimation: sparse sampling avoids noise but discards data, while dense sampling captures more signal but requires explicit noise correction. The optimal sampling frequency is the point where the mean squared error of the estimator is minimized.

A critical assumption of early noise models is that microstructure noise is independent of the efficient price innovation. However, empirical evidence shows this is often violated. In the presence of inventory control by market makers, a large buy order that moves the efficient price upward also creates an inventory imbalance, causing the market maker to lower their quotes to attract offsetting sell flow. This induces a negative correlation between the efficient price and the noise. More complex models, such as those incorporating adverse selection, can even generate a positive correlation, requiring estimators that are robust to this dependence structure.

In modern fragmented markets, microstructure noise is not generated by a single venue. The same asset trading simultaneously across multiple lit exchanges and dark pools creates venue-specific noise components. Observed prices from the National Best Bid and Offer (NBBO) can exhibit staleness and fleeting arbitrage opportunities. The noise in a consolidated tape is a complex composite of the individual venue noises, and cross-venue lead-lag relationships introduce additional serial dependence. This fragmentation requires noise models that account for a multi-dimensional observation vector rather than a single price series.

A primary source of market microstructure noise caused by transaction prices oscillating between the bid and ask quotes. When a trade occurs at the bid, the next trade is statistically more likely to occur at the ask, creating a negative serial correlation in observed returns. This mechanical bouncing inflates realized volatility estimates and distorts covariance calculations, even when the true efficient price is unchanged. The magnitude of the bounce is directly proportional to the bid-ask spread.

The dispersion of trading activity across multiple competing lit exchanges, dark pools, and systematic internalizers. Fragmentation introduces noise because the consolidated tape aggregates quotes and prints from venues with varying latencies and fee structures. A single large order may execute as hundreds of small trades across 12+ venues, creating phantom volume spikes and spurious autocorrelation in the observed price series. This effect is amplified by Regulation NMS in US equities.

Price variation introduced by the minimum tick size mandated by an exchange. Since prices can only move in discrete increments, the observed price is a rounded version of the latent efficient price. This rounding error introduces a moving average component into the return series. Discreteness noise is particularly severe in low-priced securities where the tick size represents a larger percentage of the price, and it biases estimators of integrated variance.

Noise generated when market makers adjust their quotes to manage inventory risk. After absorbing a large buy order, a market maker may lower their bid to discourage further selling and raise their ask to encourage selling, creating temporary price pressure unrelated to fundamental value. These inventory-driven quote adjustments introduce transient mean reversion in mid-quote returns. The effect decays as the market maker unwinds their position through subsequent trades.

Price variation caused by the adverse selection component of the spread. Market makers widen spreads to protect against informed traders, and the resulting quote adjustments create noise in the observed price. When an informed trader's order arrives, the permanent price impact component reflects new information, while the temporary component—the price concession required to attract liquidity—manifests as microstructure noise. Distinguishing these components is the central challenge of transaction cost analysis.

At very high sampling frequencies, microstructure noise dominates the signal, causing realized volatility estimates to explode upward. The signature plot—a graph of realized volatility against sampling frequency—reveals this bias. As the sampling interval shrinks below approximately 5 minutes, noise-induced variance overwhelms the true price variance. The optimal sampling frequency balances the bias from noise against the variance from sparse sampling, a trade-off formalized in the two-scale estimator of Zhang, Mykland, and Aït-Sahalia.