A quoted price is a measurement, not a primitive observation of value.
Atamus Capital does not publish proprietary strategy rules, signals, feature definitions, datasets, data transformations, model architectures, candidate-generation methods, training procedures, parameters, investment universes, portfolio construction methods, execution processes, trade-level information, position-level information, or performance results. This note examines the mathematics of quoted-price formation through closed-form calculations, deterministic numerical experiments, and controlled Monte Carlo experiments under fully specified probability laws. No internal data source, execution venue, order-routing logic, cost estimate, sampling rule, signal horizon, implementation assumption, or model-development workflow is disclosed. The mechanisms discussed below are public market-structure models for reasoning about observed prices. They are not a description of Atamus Capital's internal research architecture.
A quoted price is not a direct observation of value. It is the output of a trading mechanism: a bid, an ask, a midpoint, a transaction direction, a matching rule, and a sequence of discrete prints. This distinction matters because observed returns can contain structure that is unrelated to information, forecasting ability, or economic value. We examine the simplest mechanical example: bid-ask bounce. In the Roll model, transaction prices alternate around an efficient price, inducing negative first-order autocovariance in observed price changes. That autocovariance identifies the effective spread under restrictive assumptions. The same mechanism contaminates volatility measurement. When observed prices equal efficient prices plus microstructure noise, high-frequency realized variance has expectation equal to integrated variance plus a term that grows linearly with the sampling frequency. More observations do not automatically mean more information. The article develops these results analytically, validates them through closed-form calculations, deterministic numerical experiments, and controlled Monte Carlo experiments under fully specified probability laws, and separates market-structure measurement from proprietary strategy design.
1. The price we observe is a measurement system
Many financial datasets present a price as though it were a primitive object. It is not. A printed transaction price is a measurement of a trading process. It depends on the efficient price, the quoted spread, the side demanding liquidity, queue priority, latency, hidden liquidity, auction rules, and reporting conventions. Even before any predictive model is considered, the observed return series has already passed through a market mechanism.
Atamus begins from a simple distinction:
A model that treats those three objects as interchangeable can mistake market mechanics for information. The most elementary example is bid-ask bounce. If a buy prints near the ask and a sell prints near the bid, consecutive transaction prices can move down and then up even when the efficient price has barely moved. The resulting return series may show negative serial dependence. That dependence is not necessarily mean reversion in value. It can be the arithmetic of the quote.
Let \(M_t\) denote an efficient log price or midpoint, and let \(P_t\) denote the observed transaction log price. A stylized transaction-price model is
where \(c>0\) is the half-spread in log-price units and
indicates whether the transaction prints near the bid or ask. The full effective spread is
This notation is deliberately minimal. It is not a complete market model. It is a controlled experiment showing how quoted-price mechanics can create statistical structure in observed returns even when the efficient price itself is a martingale.
A quoted price is not the efficient price.
Controlled quoted-price path
PROPRIETARY
Figure 1. The model-implied benchmark path shows an efficient midpoint, bid and ask rails, and transaction prints. The transaction series can move because the trade sign changes, even if the midpoint is stable over the same interval. No market data or Atamus Capital data is used.
View data
| Layer | What it shows |
|---|---|
| Efficient midpoint | Unobserved reference value in the controlled price path. |
| Bid and ask rails | Quoted transaction bounds separated by the fixed spread. |
| Transaction prints | Observed prices generated by trade sign, not changes in value alone. |
The lesson is not that spreads explain all short-horizon behavior. They do not. The lesson is narrower and more important: before a return can be interpreted, the measurement equation that produced it must be understood.
2. The Roll model and the sign of first-order covariance
Assume the efficient price follows a random walk:
with
Assume also that the trade signs \(Q_t\) are independent, symmetric, and independent of \(U_t\):
Observed transaction-price changes are
The one-period observed variance is
The first-order autocovariance is
Thus the effective spread is identified by
This is the Roll estimator in its cleanest form. It is also a warning. A negative first-order covariance in transaction returns can be produced mechanically by the spread. It is not, by itself, evidence of a predictable reversal in fundamental value.
The associated first-order autocorrelation is
The autocorrelation becomes more negative when the spread is large relative to efficient-price innovation variance. It becomes closer to zero when efficient-price movement dominates the spread.
The Roll covariance fingerprint.
Analytical Roll identities
PROPRIETARY
Figure 2. Under the Roll assumptions, increasing the half-spread makes the first-order covariance more negative and increases the observed one-period return variance. The full-spread estimate is \(2\sqrt{-\gamma_1}\). The values are analytical, not estimated from market data.
View data
| Default half-spread | 4 bps |
|---|---|
| Default effective spread | 8 bps |
| Roll covariance identity | \(\gamma_1=-c^2\) |
| Full-spread estimate | \(2\sqrt{-\gamma_1}\) |
Several caveats matter. The Roll estimator assumes that quote-side effects are independent and symmetric, that the efficient price innovation is serially uncorrelated, and that the spread component is the dominant source of the negative first-order covariance. Real markets can violate these assumptions through order-flow persistence, inventory control, adverse selection, discreteness, price clustering, hidden liquidity, auction effects, and changes in the spread itself. The formula is therefore not a universal spread oracle. Its value is that it reveals a structural fingerprint: quoted-price mechanics can induce negative serial covariance without any change in expected value.
3. How bid-ask bounce contaminates volatility
The same calculation explains why transaction returns can overstate short-horizon variance. The observed one-step variance is
The permanent innovation variance is \(\sigma_U^2\). The spread contributes \(2c^2\) to one-period observed variance even though it is mostly reversed through the next transaction. The long-run variance of cumulative price changes makes the cancellation visible. Over \(n\) observations,
so
The spread term does not scale linearly with \(n\) in the long-horizon cumulative price change. It appears at the endpoints. By contrast, summing squared high-frequency transaction returns counts the bounce repeatedly.
This distinction is central to realized variance. Let \(X_t\) be a continuous efficient log-price martingale benchmark with integrated variance
For the calculation below we treat \(IV_T\) as deterministic over the horizon, or equivalently condition on the volatility path. Suppose observed log prices are
where \(\varepsilon_i\) is zero-mean observation noise with
independent across observations and independent of \(X\). The naive realized variance at sampling grid \(0=t_0<t_1<\cdots<t_n=T\) is
Since
we obtain
Under the stated assumptions,
and therefore
The bias grows with the number of observations. This is the paradox of high-frequency measurement: sampling more often reduces discretization error in an ideal frictionless price, but it also accumulates microstructure noise.
Realized variance accumulates microstructure noise.
Analytical realized-variance expectation
PROPRIETARY
Figure 3. The chart plots \(\mathbb E[RV_n]/IV_T\) under the additive-noise model for a daily integrated volatility of 2 percent and several disclosed noise levels. The values are analytical outputs of \(IV_T+2n\eta^2\). They are not empirical estimates.
View data
| Daily integrated volatility | 2.00% |
|---|---|
| Displayed noise levels | 0.5, 1, 5, 10, and 20 bps |
| Expected realized variance | \(IV_T+2n\eta^2\) |
| Displayed quantity | \(\mathbb E[RV_n]/IV_T\) |
For example, with daily integrated volatility equal to 2 percent and observation noise equal to 5 basis points, sampling 390 times in a day gives
The expected realized variance is almost 49 percent higher than the integrated variance. Sampling every second in a 6.5-hour session gives \(n=23{,}400\), and the same noise assumption implies
The exact numbers depend on the stylized assumptions. The direction of the effect under the stated additive-noise model does not. Independent observation noise contributes a term proportional to the number of observations.
4. More data can increase measurement error
In a noiseless diffusion model, higher-frequency sampling improves realized variance. With observation noise, the problem changes. The statistic must balance discretization error against noise accumulation.
Under a simplified Gaussian benchmark, suppose the efficient-price increments are independent normal variables with
and the observation noise is independent Gaussian with variance \(\eta^2\). The observed return vector
is Gaussian with covariance matrix \(\Sigma_n\) satisfying
and all other off-diagonal entries equal to zero.
Because \(RV_n=R^\top R\), the Gaussian quadratic-form identity gives
and
Therefore
The mean-squared error for estimating \(IV_T\) is
This expression is not a universal sampling rule. It is a transparent benchmark showing why the naive rule "sample as frequently as possible" is wrong when prices contain microstructure noise.
The sampling-frequency tradeoff.
Exact Gaussian MSE benchmark
PROPRIETARY
Figure 4. The curves show the exact Gaussian mean-squared error of naive realized variance under the disclosed additive-noise model. The optimum sampling frequency changes with the noise level. This is a measurement example, not an Atamus Capital production rule.
View data
| Benchmark | Exact Gaussian MSE of naive realized variance. |
|---|---|
| Controls | Observation-noise standard deviation in basis points. |
| Displayed components | RMSE, squared-bias contribution, and variance contribution. |
| Purpose | Show why optimal sampling frequency depends on noise. |
The deeper implication is that high-frequency data does not remove the need for modeling. It moves the modeling problem closer to the exchange. The market researcher must decide what the statistic is intended to estimate: transaction-price variation, efficient-price variation, executable price variation, or portfolio-level realized risk after costs. Those are different estimands.
5. The estimator has assumptions
The Roll estimator is elegant because it turns a covariance into a spread:
But the square root only exists when \(\widehat\gamma_1<0\). In finite samples, even if the true covariance is negative, the sample covariance can be positive. A positive estimate does not necessarily mean the model is impossible. It can mean the sample is short, the spread is small relative to efficient-price variance, or the Roll assumptions are violated.
For a sample \(R_1,\ldots,R_n\), define
The Roll estimate is therefore naturally written as the piecewise quantity
Reporting zero when \(\widehat\gamma_1\geq0\) is a modeling convention, not a mathematical identity.
Roll estimates are structural, not magic.
Controlled Monte Carlo experiment
PROPRIETARY
Figure 5. The controlled Monte Carlo experiment uses a true full spread of 8 basis points and efficient innovation standard deviation of 5 basis points. Short samples produce wide estimator dispersion and sometimes positive first-order covariance. The estimator is informative only under its assumptions.
View data
| True effective spread | 8 bps |
|---|---|
| Efficient increment standard deviation | 5 bps |
| Replications per sample length | 30,000 |
| Sample-length controls | 50, 100, 250, 500, 1,000, and 2,500 returns |
The point is not to reject the estimator. The point is to respect it. A covariance-based spread estimate is a structural inference from a model. It should not be treated as direct observation.
6. Noise share and the limit of naive measurement
The additive-noise formula gives a simple decomposition of expected realized variance:
The expected fraction attributable to microstructure noise is
The threshold where microstructure noise contributes half of expected realized variance is obtained by solving \(\omega_n=1/2\):
The threshold where it contributes 90 percent is
For \(IV_T=0.0004\) and \(\eta=5\) basis points,
This means that under the disclosed assumptions, more than half of expected realized variance comes from observation noise once the sampling grid exceeds approximately 800 observations per day.
When noise becomes most of measured variance.
Analytical noise-share heatmap
PROPRIETARY
Figure 6. The heatmap shows \(\omega_n\) across sampling frequencies and observation-noise levels. The calculation is analytical and model-based. It is not fitted to any instrument or strategy.
View data
| Daily integrated variance | \(0.02^2\) |
|---|---|
| Noise-share formula | \(\omega_n=2n\eta^2/(IV_T+2n\eta^2)\) |
| Noise grid | 0.5, 1, 2, 5, and 10 bps |
| Sampling grid | 1 to 23,400 observations per day |
This is why market-structure research is not a cosmetic layer placed after alpha research. It shapes the meaning of the observations themselves. The same historical print sequence can imply different conclusions depending on whether the researcher studies transaction prices, quote midpoints, sampled prices, filtered prices, or execution prices.
7. What the quoted price can and cannot tell us
The quoted price contains information, but not all of it is information about value. Some of it is information about the market mechanism. That distinction matters for four reasons.
First, short-horizon serial dependence can be mechanical. Bid-ask bounce can induce negative autocovariance even when the efficient price is a martingale.
Second, one-period transaction variance is not the same as permanent price variance. The spread inflates observed return variance but largely cancels across longer cumulative changes.
Third, high-frequency realized variance can become increasingly biased as sampling frequency rises. More observations can mean more repeated counting of observation noise.
Fourth, estimators based on observed prices are estimators of a model. They require assumptions. When those assumptions fail, the output can be precise and still be wrong.
The Roll model is useful precisely because it is simple enough to derive. It teaches that market data is not raw reality. It is a recorded interaction between value, liquidity, quotes, orders, and time.
8. The institutional implication
A systematic investment process that ignores market structure can confuse measurement artifacts with opportunity. A process that understands market structure still does not have a strategy. It has a cleaner map of what the data can mean.
Atamus treats this distinction as foundational. Before a return series can support any claim about signal, risk, or implementation, the observed price must be interpreted as a measurement. The quoted price is not merely where analysis begins. It is part of the object being analyzed.
The conclusion is simple:
Some variation is information. Some is liquidity. Some is measurement. Some is convention. Serious research has to separate them before it can claim to have found anything at all.
References
[1] Roll, R. (1984). A Simple Implicit Measure of the Effective Bid-Ask Spread in an Efficient Market. The Journal of Finance, 39(4), 1127-1139.
[2] Glosten, L. R., and Milgrom, P. R. (1985). Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics, 14(1), 71-100.
[3] Hasbrouck, J. (2007). Empirical Market Microstructure: The Institutions, Economics, and Econometrics of Securities Trading. Oxford University Press.
[4] O'Hara, M. (1995). Market Microstructure Theory. Blackwell.
[5] Hansen, P. R., and Lunde, A. (2006). Realized Variance and Market Microstructure Noise. Journal of Business & Economic Statistics, 24(2), 127-161.
[6] Zhang, L., Mykland, P. A., and Ait-Sahalia, Y. (2005). A Tale of Two Time Scales: Determining Integrated Volatility With Noisy High-Frequency Data. Journal of the American Statistical Association, 100(472), 1394-1411.
[7] Ait-Sahalia, Y., Mykland, P. A., and Zhang, L. (2005). How Often to Sample a Continuous-Time Process in the Presence of Market Microstructure Noise. The Review of Financial Studies, 18(2), 351-416.
Disclaimer
Research notes published by Atamus Capital are provided for general informational and research purposes only. They do not constitute investment advice, trading advice, a recommendation, an offer to sell, or a solicitation to buy any security, fund interest, account, or investment product.
This note does not disclose Atamus Capital's proprietary strategies, signals, feature definitions, datasets, data transformations, model architectures, candidate-generation methods, training procedures, parameters, portfolio construction methods, execution processes, investment universe, research thresholds, model-development workflow, or investment decisions.