Bid-Ask Bounce: Microstructure Noise in Recorded Returns
Table of Contents
A stock trades at $49.95, then at $50.05, then back at $49.95. It looks as though the stock rose and fell. But if the best bid stayed at $49.95 and the best ask stayed at $50.05, the quotes never moved. The trades simply took place on different sides of the spread. This is bid-ask bounce, and it matters when you use recorded trades to measure short-term returns, volatility or trading costs.
What is bid-ask bounce?
Bid-ask bounce is the movement in recorded transaction prices caused by successive trades occurring on different sides of the bid-ask spread. A change in the last price can reflect which side initiated the trade even if the midpoint is unchanged.
The best bid is the highest displayed buying price in the market or feed you are looking at. The best ask is the lowest displayed selling price. The quoted spread is ask minus bid, and the midquote is their average. A buyer who wants to trade immediately can lift the ask; a seller can hit the bid, subject to the available size. Some orders execute inside the displayed spread. The midpoint is a useful benchmark, but it is neither a guaranteed fill nor a direct observation of what the stock is worth.[1] [7]
The distinction works both ways. If an order takes the last shares at the best ask, the next ask can be higher and the midpoint can rise even though the last fill was at the old ask. Farmer and coauthors documented this mechanism in London Stock Exchange order books from 1999 to 2002. The Fractal Markets Hypothesis article explains why a gap to the next quote matters. Here the question is how much movement the transaction series can show while the quotes stay put.[2]
Example: trades move while the midpoint stays at $50
Suppose five trades occur one second apart. There are enough shares to keep the bid at $49.95 and the ask at $50.05 throughout. The midpoint is ($49.95 + $50.05) ÷ 2 = $50.00, and the quoted spread is $50.05 − $49.95 = $0.10 per share. The first trade is at the bid, and each subsequent trade is on the opposite side. This deliberately alternating sequence lets us see the effect of trade location by itself.
| Second | Side | Trade | Trade change | Midquote |
|---|---|---|---|---|
| 0 | Bid | $49.95 | — | $50.00 |
| 1 | Ask | $50.05 | +$0.10 | $50.00 |
| 2 | Bid | $49.95 | −$0.10 | $50.00 |
| 3 | Ask | $50.05 | +$0.10 | $50.00 |
| 4 | Bid | $49.95 | −$0.10 | $50.00 |
The recorded price moves by a full spread each time, while every midpoint change is zero. That alternating pattern gives negative lag-one autocorrelation: a positive change tends to be next to a negative change. It does not show that a buyer can pay the bid and immediately sell at the ask. Crossing the spread in both directions would mean buying at $50.05 and selling at $49.95, losing $0.10 per share before fees.
Why the return reversals count toward volatility
The first and last trade prices are equal, so the total return is zero. But realized variance adds up squared returns: both the up move and the down move contribute. Here the one-second log returns alternate between plus and minus ln($50.05 ÷ $49.95), about 0.002 in decimal return units. The sum of their squares is approximately 4 × 0.0022 = 0.000016 in squared decimal-return units over the four-second window. The same calculation on the midpoint gives zero. On a two-second grid at seconds 0, 2 and 4, even the trade-based estimate is zero because all three sampled trades are at the bid. Another grid or trade sequence could leave some of the effect in place.[3]
The same sampling issue can show up in actual data. Hansen and Lunde found that Microsoft’s transaction-based realized variance in their 2000 sample was much higher at one-second sampling than at one-minute sampling. Some of their midquote estimates, by contrast, went down at finer intervals. An increase at the finest grids can point to microstructure noise, but the direction and size of the distortion depend on the data and the estimator.[4]
Roll’s implied spread: from negative covariance to dollars
Richard Roll asked whether you could estimate the spread with transaction prices alone. His idea was to use the way one trade enters two adjacent price changes: a high print raises the change into that trade and lowers the change out of it, all else equal. The signed products of adjacent changes measure that tendency. The formula uses covariance of price changes, in (dollars per share) squared, rather than autocorrelation, which has no units.[1]
Full Roll spread = 2 × √(−lag-one covariance of dollar price changes)
Apply the formula for a positive spread estimate only when the sample covariance is negative.
Why the population covariance is negative
In the simple model, the observed trade price Pt is the unseen efficient price Vt plus or minus half of a constant full spread S. Write the signed adjustment as (S/2)qt, where qt is +1 for a buyer-initiated trade at the ask and −1 for a seller-initiated trade at the bid. The model gives both signs equal probability and treats successive signs as independent. The efficient-price changes are uncorrelated at lag one and have zero covariance with the relevant trade-side terms. These are assumptions about the data, not conditions that alternating prints establish.[1]
The spread part of the change into trade t is (S/2)(qt − qt−1). In the preceding change it is (S/2)(qt−1 − qt−2). The only shared sign is qt−1. It has variance 1 and enters the two changes with opposite signs. Independence makes the other sign cross terms zero. Thus the model’s lag-one price-change covariance is (−S/2) × (S/2) × 1 = −S2/4. Solving for the positive S gives the formula above.[1]
For contrast, every adjacent product of the price changes in the fixed-quote example is −0.01 in (dollars per share) squared. Plugging that into Roll’s formula would produce $0.20 per share, twice the known $0.10 spread. The trades were forced to alternate; Roll’s model allows two trades in a row on the same side. The example explains why reversals happen, but it cannot validate the estimator.
How to calculate and interpret a Roll spread
- Start with consecutive transaction prices in dollars per share. Subtract each preceding price from the next to get dollar changes. Record whether observations are successive trades or regularly sampled last trades. Avoid bridging unrelated sessions or data gaps without documenting how you handle them.
- Form adjacent products. In the convention used here, subtract the mean of the full series of dollar changes from each change. Multiply each centered change by the preceding centered change, then divide the sum by the number of adjacent pairs. State your convention: some software uses a different divisor or centers the two lagged series separately.
- Read the sign before taking a square root. For negative covariance, take the square root of its absolute size and double it. The result is a model estimate of the full spread in dollars per share. Dividing by a stated reference price puts it on a relative scale.
For a separate hypothetical sample, suppose the consecutive dollar changes are $0, +$0.08, −$0.08, $0, +$0.08, −$0.08, $0, $0 and $0. Their mean is zero, so centering does not change them. Only two of the eight adjacent products are nonzero, both (+0.08) × (−0.08) = −0.0064. The sample covariance is (−0.0064 − 0.0064) ÷ 8 = −0.0016 in (dollars per share) squared. The full spread estimate is 2 × √0.0016 = $0.08 per share. At a $50 reference price, $0.08 ÷ $50 = 0.0016, or 0.16% (16 basis points). This small sample is enough to show the arithmetic, not to establish a reliable spread.
| Lag-one covariance | Roll result | Interpretation |
|---|---|---|
| Negative | Positive full-spread estimate | A model-based inference. Compare with observed quotes or fills if possible. |
| Zero | Formula gives zero at the boundary | This sample does not identify a positive spread. |
| Positive | No real-valued estimate | Report it as unavailable under the model. Taking an absolute value would disguise the sign. |
For a positive example, four dollar changes of +$0.04, +$0.04, −$0.04 and −$0.04 average to zero. Their three products sum to 0.0016 − 0.0016 + 0.0016 = 0.0016, so the covariance is 0.0016 ÷ 3 ≈ +0.000533 in (dollars per share) squared. For a zero case, +$0.04, $0, −$0.04 and $0 also average to zero and every adjacent product is zero. Neither result proves that trading is free. Harris showed that finite sample noise can produce positive covariance estimates even when Roll’s assumptions hold.[5]
Quoted spread vs. effective spread vs. Roll spread
These three measures answer different questions. Suppose the same $49.95 bid and $50.05 ask are in place, but in a separate example a buyer pays $50.02. The fill is $0.02 above a valid $50.00 reference midpoint. Twice this signed distance gives a $0.04 effective spread per share, even though the full quoted spread remains $0.10. For a sell, the signed calculation is twice the reference midpoint minus twice the fill price. The benchmark time matters: for example, the SEC’s study of execution quality used the midpoint when the order arrived.[7]
| Measure | Inputs | Question |
|---|---|---|
| Quoted spread | Best bid and ask | How far apart are displayed prices at that moment? |
| Effective spread | Fill, trade side and reference midpoint | How far from the benchmark did this trade execute, measured on a full-spread basis? |
| Roll implied spread | Series of transaction prices | What constant full spread would explain this negative covariance under the model? |
For a fill-quality check, use the actual trade and an appropriate quote benchmark where available. The Roll spread cannot tell you how well an individual order was filled. See trading execution for how orders interact with the spread, and implementation shortfall for costs measured against an investment decision price.
How to diagnose bounce in trade and quote data
- Compare synchronized trades and quotes. Use the same symbol, session, clock and venue or consolidated quote scope for your question. For example, the NYSE Daily TAQ specification distinguishes when a trade was reported to a securities information processor (SIP) from when the SIP published it. A publication timestamp is not automatically an execution timestamp. Check trade condition and correction flags, and use feed sequencing or a documented rule for timestamp ties. For a trade-matched diagnostic, do not look ahead to a future quote.[8] For a signed effective spread, use the recorded aggressor side when available. If you infer the side, label it as inferred and leave ambiguous midpoint trades unclassified.[9]
- Try fresh midquotes for measuring price movement. If the trade price oscillates inside unchanged valid quotes, a midpoint series removes that part of the oscillation. Check the age of quotes between updates and flag missing or invalid pairs. A midpoint can lag real moves or be distorted when the bid and ask refresh at different times.[4]
- Compare sampling intervals on the same window. If you are measuring realized variance, check both trade and midpoint series on several common time grids. A coarser grid can skip side switches, but it can also miss genuine price changes. If you carry the last observation forward, check how often and for how long it stays unchanged. There is no universal sampling interval that cleans every series.[3]
- Use a noise-aware method when fine observations matter. A realized kernel augments squared log returns with weighted products of returns at several lags. Under a simple observation-noise model, adjacent return errors share an opposite-signed price error. Their negative product can offset part of the extra variation that squared returns count. This is a way to estimate underlying variance, with weights and assumptions that matter; it does not estimate the spread.[10]
These checks complement tools for estimating historical volatility. Use the transaction series when the fill price is the thing you need to measure; consider valid quotes or a noise-aware variance method when the goal is the movement of the underlying price.
Common mistakes
- Using levels or percent returns in the dollar formula. The inputs are adjacent dollar changes. A covariance of decimal returns is dimensionless; its square root does not give dollars per share. A return-based version needs its own assumptions and unit conversion.[1]
- Calling a negative covariance a trading signal. A reversal in the recorded price need not be a reversal in the midpoint, and even an apparent reversal pattern may not be available at executable prices.
- Replacing trades with midquotes in the Roll calculation. Midquotes can be useful for variance, but they remove the trade-side variation Roll uses. A trade series that quietly substitutes midquotes for missing trades changes what the statistic measures.[11]
When the estimates can mislead
Orders can be persistent. Lillo and Farmer found that market order signs for Vodafone in their 1999–2002 London Stock Exchange sample remained positively correlated over many subsequent orders. In data like these, the assumption of independent trade sides is questionable. Trades inside the spread and changing spreads also weaken the interpretation of the fitted number as one constant cost.[6]
News can move both quotes. Larger price innovations alone need not change the population covariance if the innovations are uncorrelated across observations and with the relevant signs. They raise price-change variance, making the bounce less visible in autocorrelation. If moves have their own serial dependence or line up with trade flow, the covariance can pick up effects other than the spread. Check stability across comparable periods rather than treating any one negative reading as proof.[1] [11]
Removing bounce does not make a quote the true price. A stale midpoint can stay flat after a real change, and sampling less frequently can miss a move that reverses within an interval. A better input is the one that suits the question and keeps the data limitations visible.[3] [4]
Frequently Asked Questions
Sources
- Richard Roll: A Simple Implicit Measure of the Effective Bid-Ask Spread in an Efficient Market (1984), section I, footnote 5 and appendix A.
- Farmer et al.: What Really Causes Large Price Changes? (2004), sections 2 and 4.
- Aït-Sahalia, Mykland and Zhang: How Often to Sample a Continuous-Time Process in the Presence of Market Microstructure Noise (2005), sections 1, 3–5 and 8.
- Hansen and Lunde: Realized Variance and Market Microstructure Noise (2006), figure 1 and sections 2–3.
- Lawrence Harris: Statistical Properties of the Roll Serial Covariance Bid/Ask Spread Estimator (1990), abstract.
- Lillo and Farmer: The Long Memory of the Efficient Market (2004), figure 1 and sections V–VI.
- SEC: Report on the Practice of Preferencing (1997), empirical study subsection D on market order execution costs.
- NYSE: Daily TAQ Client Specification, version 4.3 (March 2026), Trades, Quotes and NBBO file fields.
- Lee and Ready: Inferring Trade Direction from Intraday Data (1991), introduction and section II.
- Barndorff-Nielsen et al.: Designing Realized Kernels to Measure the ex post Variation of Equity Prices in the Presence of Noise (2008), introduction and section 3.2.2.
- Joel Hasbrouck: Trading Costs and Returns for U.S. Equities: Estimating Effective Costs from Daily Data (2009), sections I.A and II.C.
Educational Use
The worked prices and calculations are hypothetical. They illustrate what the data can measure and are not trading recommendations.
