ASF

Price as Geometry

March 17, 2026|
A
A

Does the price process have memory? The answer depends on the timescale at which it is asked. On BTC and ETH in 2025 the returns of a few milliseconds are close to uncorrelated, returns between about 50 ms and 30 s are positively autocorrelated, and returns of a minute and longer behave approximately as a random walk. No single model captures all three regimes. The transition between them is a geometric fact about coarse-graining.

Other markets show the same pattern. Lo and MacKinlay applied variance ratio tests to weekly CRSP equity returns in 1988 and rejected the random walk null.[1] Cont's survey of stylised facts documented a resolution-dependent pattern (autocorrelation of returns that vanishes beyond about 20 minutes, heavy tails) across many markets.[21] Lo's 1991 study used a modified rescaled range statistic robust to short-range dependence and found no evidence of long-range memory in US stock returns.[22] This post adds a geometric framework in which these patterns are statements about coarse-graining, together with measurements at millisecond resolution that show the fine structure invisible to daily studies.

The organising variable throughout is the temporal resolution : the coarse-graining map that projects the microscopic trade-by-trade price process onto a sequence of binned log-returns. Every statistical property examined is a function of : stationarity, volatility scaling, autocorrelation, spectral structure and information flow between assets all change as varies.

The empirical measurements use BTC/USDT and ETH/USDT trades from Binance spot over the calendar year 2025: 441.6 million BTC trades (1.21 million a day) and 470.0 million ETH trades (1.29 million a day). ETH trades have microsecond timestamps. BTC trades have millisecond timestamps on 345 of the 365 days, which fixes 1 ms as the finest resolution used. The ETH file is empty on 19 January and 6 May, and ETH statistics use the other 363 days. Crypto markets trade continuously for 24 hours with no opening or closing auction, and the trade record is public. Every statistic below is computed on a calendar grid: the last trade price is carried forward through bins with no trade, and each day counts once (the reason is given in Section 4). The scripts are in the site repository under scripts/price_geometry/.

1. Philosophical Priors

No market ontology is assumed here. The price of any liquid asset is the output of a deterministic system: a large number of interacting agents, each choosing to submit, cancel or execute orders according to their own information and objectives. The mid-price at any moment is the projection of that high-dimensional deterministic system onto a single observable scalar.

The apparent stochasticity of prices is an artefact of observation. Binning trades into millisecond buckets and computing log-returns is a coarse-graining operation. The resulting time series inherits statistical properties that depend on the coarse-graining timescale. At fine resolution the price at time reflects individual order decisions that each have structure. At coarse resolution it averages over so many decisions that the central limit theorem pushes the residual towards Gaussian noise with independent increments.

This prior has a precise mathematical form developed in Section 3. Asking "is the market stationary?" or "does price follow a random walk?" without specifying the resolution is ill posed. Every answer below is relative to a specific .

Any trading strategy, risk model or derivative pricing formula is implicitly a claim about which stochastic model is closer to true at the resolution at which the strategy operates. A market-making strategy at 10 ms makes a different claim from a momentum strategy at one hour. This post measures which claims are empirically defensible and at which resolutions.

Section 8 returns to this framing: the Fisher-Rao metric on the family of price return distributions is itself a function of , and the market's information geometry changes as the clock coarsens.

2. On-ramp to the same question

This section introduces the key objects without set-builder notation. Readers already comfortable with log-returns and the variance ratio test may skip to Section 3.

Mid-price. A limit order book quotes two prices for a traded asset. At any moment there is a best ask (the lowest price at which someone is willing to sell) and a best bid (the highest price at which someone is willing to buy). The mid-price is their average,

The mid-price filters out the alternating "uptick/downtick" pattern of trades bouncing between the bid and the ask. The data set used below has trades and no quotes. Every measurement therefore uses the last trade price in place of the mid-price. The bounce it adds is small on these two markets: consecutive trades of opposite aggressor side less than 1 ms apart differ by a median of of the price for BTC and for ETH, one price step of 0.01 USDT in both cases.

Order book. The spread is the loss on an immediate round-trip: buy at the ask, sell at the bid. Depth refers to the quantity available at each price level. A market order that exceeds the depth at the best bid or ask walks the book and moves the mid-price. Depth and order flow create price impact: the tendency of large orders to move prices against the trader who placed them.

Log-return. At resolution , the log-return at time is

We use log-returns in place of percentage returns for two reasons. First, they are additive over time: the log-return over two periods is the sum of the two single-period log-returns. Multi-scale analysis is then algebraically clean. Second, they are approximately symmetric around zero for small moves and percentage returns have an upward bias for large positive moves. At millisecond timescales the distinction is negligible and at longer horizons it is material.

Volatility and the prediction. The realised volatility at resolution over a window of periods is

If returns were independent and identically distributed, the volatility at resolution would satisfy . This is the square-root-of-time scaling, the signature of a process with no memory. It is the null hypothesis. Section 4 tests it with the variance ratio statistic.

Stationarity. A time series is stationary if its statistical properties do not change over time. A river in steady flow is stationary: the distribution of water heights at noon today is the same as at noon a year ago. A river during a flood is not. For price processes, stationarity means that the distribution of log-returns at any given resolution does not drift as time passes. Whether BTC log-returns are stationary is a resolution-dependent question: they may be stationary at a one-hour timescale while exhibiting systematic trends at a five-year timescale.

Random walk. The canonical null hypothesis in finance is that the price follows a random walk: tomorrow's price is today's price plus an unpredictable shock drawn independently from some fixed distribution. Under the random walk null, no strategy that looks only at past prices can forecast future returns. The variance ratio equals one at every timescale, the autocorrelation of returns is zero at every lag, and the Hurst exponent is exactly one-half.

Consequences. Every trading strategy, risk model and option pricing formula assumes either that the random walk null holds at the relevant timescale or that it fails in a specific exploitable direction. Market-making strategies assume short-timescale mean reversion and would lose money if the null were exact. Trend-following strategies assume persistent autocorrelation and would lose money if the null were exact. This post establishes empirically and geometrically which assumptions are warranted at which resolutions.

3. Coarse-Graining Map

We now develop the precise mathematical framework. The temporal resolution is a parameter that changes the statistical model at every level.

Let be a probability space supporting the microscopic price process , where is the mid-price at time .

Definition (Filtration and coarsened filtration).

The natural filtration of the price process is the family , encoding all information up to time . For a resolution parameter , the -coarsened filtration is

the information available at the most recent grid point at or before .

Definition (Coarse-graining projection).

The coarse-graining projection at resolution is the conditional expectation operator

When is an integer multiple of the -grid is a subset of the -grid and . The tower property then gives . Without the divisibility condition the filtrations need not be nested (take , , ).

Definition (Resolution manifold).

The resolution manifold of the price process is the one-parameter family

where is the law of the coarse-grained process with . The resolution manifold is fibered over the positive real line .

The question "is the market stationary?" asks whether varies with calendar time for a fixed . A distinct question is whether the family is constant in once returns are rescaled by . If so the process has scale-invariant statistics: the random walk null and geometric Brownian motion both satisfy this. Otherwise the process has resolution-dependent structure.

Theorem (Markov property under coarse-graining).

If is a time-homogeneous Markov process with transition semigroup , the sampled sequence is a Markov chain with one-step kernel and -step kernel . If the observed price is a function of an unobserved Markov state (the full order book), the sampled price need not be Markov at any .

Proof. Since the process changes only at the grid points , and at a grid time. For bounded measurable and the Markov property of gives

The right side is a function of alone. Hence it also equals . This is the Markov property of the sampled chain with -step kernel . The Chapman-Kolmogorov identity is the semigroup property.

For the second statement, conditioning on the past of is coarser than conditioning on the past of , and depends on the earlier observations whenever they contain information about the hidden part of .

Corollary (Closure of GBM under coarse-graining).

Geometric Brownian motion is closed under : the coarse-grained log-return process has i.i.d. increments with variance scaling as . Hence and the variance ratio for all .

Proof. Under GBM, the log-price satisfies by Ito's formula. For any two times , the log-return over is

Since by the definition of Brownian motion, the -increment has distribution

so . Increments over non-overlapping intervals are independent: for with , the increments are driven by disjoint pieces of the Brownian path.

The -return is the sum of consecutive -returns:

By independence and identical distribution of the summands,

Substituting into the variance ratio definition,

This holds for every and every . Hence GBM is closed under coarse-graining with exact unit variance ratio at every scale.

The corollary connects theory and empirics. The conclusion returns to it.

4. Random Walk Null

We now operationalise the null hypothesis and test it at 22 resolutions, from 1 ms to 4 hours.

Definition (Geometric Brownian motion).

The geometric Brownian motion (GBM) model postulates

where is a standard Brownian motion. Under GBM, the log-price follows . Hence log-returns at any resolution are i.i.d. Gaussian.

Definition (Variance ratio statistic).

Let denote unit-period log-returns and the -period return. The Lo-MacKinlay variance ratio at horizon is

Under GBM, at every by the Corollary in Section 3.

Theorem (Asymptotic distribution of VR under the random walk null).

Under the null hypothesis that are i.i.d. with finite fourth moment, and with observations,

[1]

Proof sketch. The variance ratio can be written as

where (demeaned). Under the null are i.i.d. with mean zero and variance and the numerator is a linear combination of sample second moments at overlapping horizons. Expanding the -period variance estimator,

where are the sample autocovariances. Under the null, for all , and by the CLT for sample autocovariances of i.i.d. sequences, jointly for , with uncorrelated components. Applying the delta method and summing the covariance contributions from the weighting in the expansion yields variance

The sum identity is the standard sum-of-squares formula. The Lindeberg CLT applied to the triangular array of summands completes the argument and gives the stated Gaussian limiting distribution.[1]

Definition (Hurst exponent).

The Hurst exponent characterises long-range dependence via the variance scaling

The three regimes are: (GBM, uncorrelated increments), (mean-reverting, negatively correlated increments), (trending, positively correlated increments).

Theorem (DFA consistency).

The detrended fluctuation analysis (DFA) estimator of the Hurst exponent converges to for fractional Gaussian noise and for a semiparametric class of stationary long-range dependent processes.[2]

Proof sketch. The argument below is for stationary Gaussian increments with long-memory covariance . DFA proceeds as follows. Given observations , form the profile . Partition into non-overlapping windows of size . In each window, fit a local polynomial trend and compute the root-mean-square residual . The DFA scaling hypothesis is

so is estimated as the slope of the log-log regression of against .

The convergence argument has two steps. First, one shows that concentrates around its expectation at rate by an LLN argument applied to the squared detrended residuals across windows. Second, one identifies via the structure of the covariance function of the profile process: for a long-memory process with covariance , the variance of the detrended profile over a window of size grows as by the Karamata theorem for regularly varying functions. The log-log slope estimator is then consistent because it is a continuous function of the in the OLS regression of on across the window sizes , as with . Bardet and Kammoun also show that a trend in the data distorts the DFA estimate.[2]

Empirical results. With a unit period of 1 ms compares the variance of -millisecond returns with times the variance of 1 ms returns. For each day the ratio is computed from that day's returns alone, and the curve is the mean over days. Pooling squared returns across the year would let a few crash days decide the answer. On 10 October 2025 trades on both pairs bounced by up to 10% within single milliseconds, and that day has 85% of the year's 1 ms BTC variance. With squared returns pooled the BTC ratio at 1 s is 0.26. The day-by-day mean is 2.22. The local Hurst exponent is half the slope of against . It applies the variance-scaling definition above scale by scale. The 95% bands come from resampling days.

Variance ratio and local Hurst exponent for BTC and ETH at 22 resolutions
Top: variance ratio relative to 1 ms for BTC (orange) and ETH (blue) at resolutions from 1 ms to 4 hours with 95% bootstrap bands over days. The dashed line is the random-walk value . Bottom: local Hurst exponent . Both assets rise above between about 20 ms and 30 s and settle at or slightly below beyond 2 minutes.
DFA log-log plot for BTC and ETH showing Hurst exponents
DFA fluctuation function of 1 s returns for BTC and ETH over 2025. Each day's returns are divided by that day's standard deviation. The dashed lines are least-squares fits over windows from 1 minute to 1 day, whose slopes are the Hurst estimates .

Interactive: drag the slider to read the variance ratio and the local Hurst exponent at each resolution.

Result. The random walk null fails at sub-minute scales, and the failure is persistence. From 1 ms to 10 ms the local exponent is close to (BTC 0.51 to 0.53, ETH 0.50 to 0.52). It peaks at 200 ms (BTC 0.60, ETH 0.59) and falls back through between 30 s and 2 minutes. The variance ratio rises accordingly: at 1 s it is 2.22 for BTC and 1.88 for ETH. It peaks at 3.7 and 2.6 between 30 s and 2 minutes and drifts down to 3.2 and 2.2 at 4 hours. Beyond 2 minutes the local exponent lies between 0.46 and 0.50. This weak mean reversion over minutes to hours is resolved by the 95% bands at several scales. The DFA estimates over windows of 1 minute to 1 day are for BTC and 0.491 for ETH. Nothing in these data shows mean reversion at millisecond resolution on a typical day. The hourly result agrees with Lo's finding of no long-range memory in equity returns.[22] Return autocorrelation confined to short intraday horizons is one of the stylised facts surveyed by Cont.[21]

Equity comparison. Lo and MacKinlay's weekly CRSP study found variance ratios above 1 for portfolios, largest for small-cap portfolios, and slightly negative autocorrelation for individual stocks.[1] The BTC and ETH ratios above 1 sit at sub-minute horizons, four orders of magnitude shorter than the weekly horizon of that study.

5. Mean Reversion as Geometry

The random walk null fails at short timescales. We survey the main stationary alternatives, framed geometrically.

Definition (Ornstein-Uhlenbeck process).

The Ornstein-Uhlenbeck (OU) process is the solution to

The drift term is the restoring force: it pulls the process back towards the long-run mean with speed . The stationary distribution is .

Theorem (Unique stationary Gaussian Markov process).

The OU process is, up to affine transformation, the unique continuous-time stationary Gaussian Markov process.[4]

Proof sketch. We characterise all continuous-time stationary Gaussian Markov processes. A process that is simultaneously Gaussian and Markov is determined by its mean function and covariance function .

Stationarity requires (constant) and for some function .

Markov property for a Gaussian process is equivalent to requiring that the partial correlation between and (for ) vanishes when conditioning on . For a Gaussian process this means

Using the Gaussian conditional covariance formula and the stationarity condition , this becomes the functional equation

Setting , and , the equation reads . This is Cauchy's multiplicative functional equation on . Under mild regularity (measurability of , which holds when the process is continuous in probability), the only non-zero solutions are for some .

The exponential covariance is the covariance of the stationary OU process with stationary law . The case gives a constant covariance. It describes the degenerate process and is excluded. Thus up to the affine reparametrisation the OU process is the unique member of this class.[4]

The OU process has flat geometry. More general diffusions can have curved geometry.

Definition (Black-Scholes as flat geometry).

Under the change of variables , the Black-Scholes PDE

becomes . With time to expiry and for suitable constants , this reduces to the heat equation on . The Black-Scholes formula is the transformed payoff integrated against the Gaussian heat kernel on flat : zero curvature and vanishing Christoffel symbols. The Ito correction becomes a constant drift that the exponential substitution removes.[5]

Definition (Curved Riemannian diffusion).

A gradient diffusion on a Riemannian manifold has generator , where is the Laplace-Beltrami operator. In local coordinates it solves the Ito SDE

where is an orthonormal frame (). The term in the Christoffel symbols is the Ito correction:

The correction depends on the chart as well as on the metric. It vanishes in Cartesian coordinates on flat space and in harmonic coordinates on any manifold, and it is non-zero in polar coordinates on the flat plane.

Proposition (Chart drift absorbed by an OU fit).

Let be a one-dimensional ergodic diffusion with stationary law and drift , where is the Ito correction of the chart and . Assume . The continuous-observation least-squares OU estimator converges almost surely:

Proof. Write in the numerator. The ergodic theorem gives , and . The martingale term tends to zero by the strong law for martingales, since its quadratic variation grows linearly in . Hence , and substituting gives the limit.

The bias is the slope of the regression of the chart drift on . It depends on the chart through the Christoffel symbols and does not measure curvature: every one-dimensional metric is flat, yet is non-zero in any chart where . An OU fit overestimates when and are negatively correlated under and underestimates it when they are positively correlated.

One model in the table below needs a definition first. Fractional Brownian motion (fBm) with Hurst parameter is the unique (up to scaling) continuous Gaussian process with stationary increments satisfying . Standard Brownian motion is the special case , where increments are independent. For the increments are correlated across all lags: positively for (long-range dependence) and negatively for (anti-persistence). The covariance matrix of fBm increments is Toeplitz with entries ; for this is:

At we have for all and the covariance reduces to the identity of standard Brownian increments. At the off-diagonal entries are negative and encode the anti-persistent structure of rough volatility. Gatheral, Jaisson and Rosenbaum estimated the regularity of log-volatility from realised variance of major equity indices and found (far below the Brownian value).[7] At the sample paths of fBm are far rougher than Brownian motion. Fractional Brownian motion with is not a semimartingale and the classical Ito calculus does not apply to it, and a theory of integration against such paths requires the rough paths framework.[8]

The table below summarises the geometric interpretation of the standard models.

ModelGeometric interpretation
GBM / Black-ScholesFlat , BS formula = Gaussian heat kernel,
Avellaneda-StoikovFlat mid-price; affine inventory correction; symmetric optimal quotes
OU / VasicekFlat , quadratic potential , Gaussian stationary measure
HestonState space with metric from the vol process; Ito correction from Christoffel symbols modifies BS drift
Rough volatilityLog-vol is fBm with ; path space requires rough paths calculus
General curved diffusion arbitrary; Ito correction from gives observable drift signature

The Avellaneda-Stoikov market-making model[6] is an example. It assumes a mid-price following arithmetic Brownian motion on flat and derives the optimal bid and ask quotes by solving a Hamilton-Jacobi-Bellman equation. The reservation price is

a linear (affine) correction for inventory with risk aversion . The optimal spread is

The linearity of the inventory correction and the symmetry of the optimal quotes about the reservation price are both consequences of the flat geometry. On a curved state space the Hamilton-Jacobi-Bellman equation acquires curvature corrections, the inventory correction becomes nonlinear and the optimal bid and ask are no longer symmetric about . Whether these corrections matter at millisecond scales is open.

6. Autocorrelation Structure Across Scales

The variance ratio test detects memory in aggregate. The autocorrelation function resolves it by lag.

Definition (Autocorrelation function).

For a wide-sense stationary process , the autocorrelation function at lag is

the normalised inner product on .

Theorem (Wiener-Khinchin theorem).

For a wide-sense stationary process with integrable autocorrelation function, the (normalised) power spectral density and the autocorrelation function form a Fourier pair:

[14][15]
Definition (Short and long memory).

A stationary process has short memory if and long memory if the sum diverges. Fractional Gaussian noise (the stationary increment process of fractional Brownian motion) with Hurst exponent is the canonical long-memory model.

Theorem (Spectral characterisation of long memory).

For a stationary process with autocovariance and , the power spectral density satisfies

[3]

Empirical results. Signed volume is the aggressive buy quantity minus the aggressive sell quantity in a bin, zero in a bin with no trade. We compute its autocorrelation and that of log returns at resolutions of 1 ms, 10 ms, 100 ms and 1 s and lags from 1 ms to 1 s, each as the mean over days of the daily estimate. The heatmap shows both dimensions at once.

Signed-volume and return autocorrelation at 1 ms resolution for BTC and ETH
Autocorrelation at 1 ms resolution of signed volume (left) and log returns (right) with 95% bootstrap bands over days. Signed volume has lag-1 autocorrelation 0.051 for BTC and 0.066 for ETH and stays positive at every lag. The return autocorrelations are below 0.015 in magnitude at every lag.
Autocorrelation heatmaps of signed volume and returns across lag and resolution for BTC and ETH
Autocorrelation by lag (-axis) and resolution (-axis) for signed volume (top) and log returns (bottom). Red is positive and blue negative. Grey cells have a lag shorter than the resolution and are undefined. Both quantities are positively autocorrelated in almost every cell. The lag-1 return autocorrelation grows from 0.014 (BTC) at 1 ms to 0.14 at 100 ms.

Interactive: switch between signed volume and returns, and brush a region to read the mean autocorrelation over the selected cells.

Result. Order flow is persistent. The lag-1 signed-volume autocorrelation is 0.051 for BTC and 0.066 for ETH at 1 ms and 0.051 and 0.042 at 1 s. Within the 1 ms series it falls to about 0.006 by lag 10 ms. Returns at 1 ms are nearly uncorrelated: the lag-1 value is 0.014 for BTC and for ETH, with a band from to 0.002 that contains zero. Coarser bins show the persistence of Section 4: the lag-1 return autocorrelation is 0.045 and 0.023 at 10 ms, 0.14 and 0.13 at 100 ms and 0.13 and 0.10 at 1 s. At 1 minute it is 0.002 for BTC and for ETH, and at 1 hour and . Cont lists the absence of return autocorrelation beyond short intraday horizons among the stylised facts of equity and currency markets.[21]

Equity comparison. Roll derived a mechanical source of negative autocorrelation: in a market with bid-ask spread and trades arriving at random on either side, the serial covariance of transaction-price changes is whatever the underlying fundamental process.[23] The bounce shrinks with the spread. On Binance spot in 2025 consecutive opposite-side trades less than 1 ms apart differ by one price step of 0.01 USDT, a relative spread of for BTC and for ETH. Neither pair shows the negative lag-1 value of Roll's mechanism at 1 ms. ETH, with the wider relative spread, has a lag-1 value consistent with zero and BTC a positive one.

7. Wavelets and Geometric Harmonics

The Wiener-Khinchin theorem decomposes variance by frequency. Wavelets refine this by isolating variance at specific timescales without losing time localisation.

Definition (Multiresolution analysis).

A multiresolution analysis (MRA) of is a nested sequence of closed subspaces satisfying the Mallat axioms: completeness, closure under translation and dilation, and generation by a single scaling function . The complementary subspaces (where denotes the orthogonal complement: is the subspace of orthogonal to , so that ) are generated by the mother wavelet .[9]

Theorem (Wavelet variance decomposition).

For a second-order stationary process , the total variance decomposes over wavelet scales as

where is the wavelet variance at scale and is the -th level wavelet coefficient.[10]

Proof sketch. Take in discrete time with unit spacing, mean and spectral density on . Then . The level- coefficients of the maximal overlap discrete wavelet transform are for a filter with squared gain . Since sums to zero, has mean zero and

The level- filter is approximately band-pass for the octave , and for an ideal filter bank the integral is exactly the spectral mass of that octave:

The pyramid algorithm gives for every , where is the squared gain of the level- scaling filter. As , for every . When is integrable and has no atom at zero, dominated convergence gives .[10]

The wavelet variance profile is the time-localised analogue of the power spectrum: it shows which scales contain the variance. It is also defined for processes whose increments of some order are stationary (such as fractional Brownian motion), where the ordinary power spectrum is not.

Definition (Graph Laplacian and diffusion map).

Given a point cloud , define the weight matrix and the graph Laplacian where is the diagonal degree matrix. As , the rescaled operator converges to a multiple of the Laplace-Beltrami operator on the underlying manifold , exactly so for uniform sampling and in general after the density normalisation of Coifman and Lafon. The eigenfunctions of are the geometric harmonics.[11]

Theorem (Diffusion distance).

Euclidean distance in the full diffusion map coordinates equals the diffusion distance

where is the -step transition kernel of the random walk on the graph and is its stationary distribution. Two points are close in diffusion distance if there are many short paths between them on the manifold.[11]

Proof sketch. Let be the Markov matrix of the random walk and its symmetric conjugate. Then has orthonormal eigenvectors with real eigenvalues . The stationary distribution is , and are right eigenvectors of normalised so that . In particular is constant. The diffusion map embeds each point into via

The squared Euclidean distance in this embedding is

The spectral expansion of gives . The functions are orthonormal in . Hence

where the term drops out because is constant. With this is , and truncating at discards terms that are small when is small. As and , the Markov matrix converges to the heat semigroup of the Laplace-Beltrami operator, and the diffusion distance converges to the diffusion distance on the underlying manifold.[11]

The diffusion distance is a notion of geometric proximity intrinsic to the data manifold. For price data two market states are "similar" when the conditional distributions of future returns are close in .

Empirical results. We apply a Daubechies-4 (D4) discrete wavelet transform to each day's 1 ms log returns and compute the wavelet variance at twelve levels, whose scales ms run from 1 ms to 2.048 s. Each level's variance is divided by the day's 1 ms return variance. The shares then sum to one up to the variance left beyond level 12. For white noise the share at level is .

Wavelet variance decomposition for BTC and ETH
Share of 1 ms return variance at each wavelet scale for BTC (orange) and ETH (blue), averaged over the days of 2025, on a log axis. The dashed line is white noise. Both assets follow it up to 32 ms and lie above it from 64 ms (BTC) and 128 ms (ETH).

Interactive: watch the 30-day profile of times the variance share evolve through 2025 (white noise is flat at 1). Press play or drag the scrubber.

Result. The finest scale has half the variance (BTC 0.49, ETH 0.50) and each level up to 32 ms has close to half the share of the one below, as for white noise. From 64 ms for BTC and 128 ms for ETH the shares exceed the white-noise value: at 2.048 s by a factor of 2.3 for BTC and 2.0 for ETH. The excess is the persistence of Section 4 seen scale by scale. Across days, the excess at the four coarsest levels is negatively correlated with the day's 1 ms return variance (correlation for BTC and for ETH on the logarithm of the variance). Volatile days are closer to white noise at sub-second scales and calm days have relatively more coarse-scale variance.

Equity comparison. Andersen, Bollerslev, Diebold and Ebens computed daily realised volatility for all 30 DJIA component stocks from 5-minute returns and found strongly persistent, long-memory volatility. They did not decompose return variance by wavelet scale.[25] The persistence they report concerns volatility over days. The excess coarse-scale variance measured here concerns signed returns below a few seconds.

8. Information Geometry and Cross-Asset Flow

The statistical models that describe the price process at different resolutions are themselves elements of a geometric space. This is the domain of information geometry.

Definition (Statistical manifold).

The statistical manifold of price return distributions at resolution is the family

where parameterises the distribution of the log-return . As varies, the manifold changes: at fine it needs parameters for the autocorrelation of returns; at coarse it is approximately a Gaussian family parameterised by .

Definition (Fisher-Rao metric).

The Fisher information matrix at is the matrix whose entry is the expected product of the -th and -th score functions:

defining a Riemannian metric on called the Fisher-Rao metric. The matrix is symmetric and positive semi-definite. This is the unique (up to scale) monotone metric on the space of probability distributions.[12]

For the Gaussian family with , the two score functions are and . Their expectations give a diagonal matrix:

The diagonal structure means and are orthogonal coordinates on : estimating the mean gives no Fisher information about the variance and vice versa. In the coordinates the metric is . The Fisher-Rao geodesic distance on this family is therefore times the hyperbolic distance between the points of the upper half-plane, and the family has constant curvature .

Theorem (Cramer-Rao bound as geodesic constraint).

Any unbiased estimator of satisfies

Equality holds only for exponential families in the mean-value parametrisation, where the maximum likelihood estimator attains it. In regular models the MLE attains the bound asymptotically. The Cramer-Rao bound is a statement about the geometry of : no estimator can achieve precision greater than the inverse Fisher information, just as no path on a Riemannian manifold can be shorter than the geodesic distance.

Proof (scalar case). Let be an unbiased estimator of , so for all . Let denote the score function.

Step 1 (zero-mean score). Differentiating the identity with respect to under the integral sign (justified by dominated convergence under regularity conditions),

Step 2 (covariance identity). Differentiating with respect to ,

Since , this gives .

Step 3 (Cauchy-Schwarz). By the Cauchy-Schwarz inequality for covariances,

where is the Fisher information (using Step 1: ). Substituting gives

Equality. The Cauchy-Schwarz inequality is tight if and only if is proportional to almost surely. This means for some function . For an exponential family parametrised by its mean , the score is . The sufficient statistic satisfies exactly this condition and is the MLE, which therefore achieves the bound.

Definition (Transfer entropy).

The transfer entropy from process to process at resolution is

where denotes conditional Shannon entropy. It quantifies how much knowing the past of reduces uncertainty about the future of , beyond what is already known from 's own past.

Theorem (Transfer entropy as KL divergence).

Transfer entropy equals the conditional Kullback-Leibler divergence, averaged over the conditioning variables:

The Fisher-Rao metric is the unique Riemannian metric for which is the natural divergence to second order in the parameter displacement [12, Chapter 3]. Transfer entropy is a coordinate-free measure of information flow on the statistical manifold .

Proof. Denote , , for brevity. Starting from the definition,

Expanding the conditional entropies using the definition ,

where the expectation is over the joint distribution . Writing this expectation as a sum over all values,

which is exactly the KL divergence between the conditional distribution of given and the conditional distribution of given alone, averaged over the marginal . This is the conditional divergence in the statement.

Non-negativity. By Gibbs' inequality, with equality if and only if almost everywhere. Gibbs' inequality follows from the convexity of : by Jensen,

Equality holds in Jensen if and only if is constant almost surely, i.e., . Therefore , with equality if and only if knowing provides no additional information about beyond what already provides, i.e., .

For the two-asset BTC/ETH system, the transfer entropies at a given resolution organise into an information flow matrix:

The diagonal entries are zero by definition (a process has no transfer entropy to itself). The dominant direction of information flow at resolution is read off the larger off-diagonal entry. The crossover is the value of at which the two off-diagonal entries are equal and the dominant direction switches.

Empirical results. We compute and at nine resolutions from 1 ms to 1 s. Both return series sit on the same clock grid and the two returns in a pair end at the same instant. Each return is coded by its sign (down, unchanged, up), the history is one bin, and the joint frequencies are pooled over the year. A second estimate with the source series shifted by 60 s measures what the estimator returns for a source with no timing information.

Transfer entropy between BTC and ETH at nine resolutions
Transfer entropy (orange) and (blue) at resolutions from 1 ms to 1 s on a log axis with 95% bootstrap bands over days. Dotted lines are the same estimates with the source shifted by 60 s. The two directions agree within their bands at 1 and 2 ms. From 5 ms upwards ETH to BTC is the larger.

Interactive: drag the slider to change resolution and compare the two directions of information flow.

Result. Information flows in both directions and grows with resolution, from about bits at 1 ms to (BTC to ETH) and bits (ETH to BTC) at 1 s. The shifted-source estimates are 4% of the measured values at 1 ms and 15% to 28% at 1 s. Pooled over the year the two directions at 1 ms are (BTC to ETH) and bits (ETH to BTC) with overlapping bands, and at 2 ms they differ by 3%. Day by day the direction does change between 1 and 5 ms: BTC to ETH is the larger on 62% of days at 1 ms, ETH to BTC on 60% at 2 ms, 88% at 5 ms and at least 98% from 10 ms. At 1 s ETH to BTC is 3.7 times the reverse. The companion post (BTC/ETH Lead-Lag, March 2026) reports a reversal near 15 to 20 ms in a different statistic, the order in which large moves arrive. Sign-coded transfer entropy with a one-bin history places the change of direction near 2 ms and has ETH ahead at every coarser resolution. The flow matrix changes in size and asymmetry with , and a single flat model with fixed parameters cannot match it at every resolution.

9. Machine Learning as Data-Driven Geometry

This section surveys the choices that standard machine learning methods make implicitly about the geometry of the price process, framed through the resolution lens above. It has no new empirical results. The synthesis in Section 10 draws only on the empirical findings of Sections 4, 6, 7 and 8.

9.1 Gaussian Process Regression

Definition (GP kernel as Riemannian metric).

A Gaussian process prior over price paths is specified by a positive semi-definite kernel . The kernel defines a quadratic form on square-integrable test functions,

which is the covariance of the Gaussian functionals and . The inner product of the reproducing kernel Hilbert space is the dual form, given by the inverse of the integral operator with kernel . Either form is a choice of flat metric on a space of price paths. A kernel is stationary if depends only on the lag , and non-stationary otherwise.

Proposition (Stationary GP misspecification).

A GP on returns with a stationary kernel fixes the ACF at every resolution: the ACF of the binned process at resolution is determined by through the formula below. If , as for the squared exponential, Matérn and rational quadratic kernels, the binned ACF is non-negative at every lag and every resolution. Such a kernel cannot produce the weak negative return autocorrelation at minutes to hours of Section 4, and a single must match the four rows of the return heatmap of Section 6 at once.

Proof. Let be the return over the -th bin. Stationarity of gives

which depends on , and only. The denominator is positive for a non-degenerate kernel. If the numerator is non-negative, and for all and .

Non-stationary kernels can in principle capture resolution-dependent structure, but specifying them requires an explicit model for how the ACF changes with . This model is the resolution manifold of Section 3.

9.2 Neural SDEs

Definition (Infinitesimal generator of a neural SDE).

For the Ito neural SDE

the infinitesimal generator acting on a twice-differentiable test function is

The generator encodes the full first- and second-order local behaviour of the process and determines the transition semigroup via .

Theorem (Resolution dependence of the learned generator).

Let be the maximum likelihood estimator for the neural SDE trained on observations at timestep . Then approximates the generator of the coarse-grained process , not the generator of the underlying price process . In general, for .[16][17]

Proof. The Euler-Maruyama discretisation at step gives the transition

The MLE minimises the KL divergence between the empirical transition density at step and the model transition density. By Definition 3, is the -step transition density of the coarse-grained process . The Euler-Maruyama likelihood is Gaussian with mean and variance . To leading order the fit therefore matches

Both ratios are statistics of the -step increments. For a diffusion they converge to its drift and squared diffusion coefficient as and differ from them at order . For the observed price, which need not be Markov (Theorem 3), the variance ratio differs from 1 (Section 4). Then the average of at step is times its value at step , and the learned generators differ.

9.3 Reservoir Computing

Definition (Echo state property).

A reservoir recurrent network with weight matrix , input weights , and nonlinearity updates its state via

The network has the echo state property if the state is asymptotically a unique function of the input history , independent of the initial state. For a sufficient condition is that the largest singular value of is below 1. The usual design rule is spectral radius . It is necessary for the property under zero input and not sufficient in general. For the linearised reservoir the influence of on decays asymptotically like .[18][19]

Proposition (Memory timescale versus input autocorrelation time).

The reservoir's effective memory timescale at input sampling rate is

The reservoir faithfully propagates autocorrelation structure only when , where is the autocorrelation decay time of the input process at resolution . When the reservoir forgets correlated history before it can influence the readout; when the reservoir saturates with stale history.

Proof. The contribution of to the current state passes through applications of . In the linearised reservoir it is bounded by , and by Gelfand's formula. Setting this to of its initial value gives steps, corresponding to elapsed time . The input autocorrelation contributes a signal of amplitude at lag ; the reservoir transmits this signal only for . If , substantial autocorrelation at lag is discarded. If , the reservoir retains state from lags where and adds noise to the readout.

From Section 6, the return ACF at falls from about 0.14 at lag 100 ms to 0.07 at 200 ms and 0.02 to 0.03 at 500 ms. The decay time is roughly 150 to 300 ms. A reservoir intended to exploit this sub-second persistence must be tuned to . This requires between and . At where the ACF is near zero, the optimal reservoir is nearly memoryless.

9.4 Signature Methods

Definition (Path signature).

For a continuous bounded-variation path , the depth- signature is the collection of iterated integrals

over all multi-indices and depths . The full signature is the limit as .

Theorem (Chen's identity).

For a path and intermediate time , the signature satisfies

where denotes the tensor product of the free tensor algebra. For time-augmented paths the linear functionals of are dense in the space of continuous functions on the space of paths, uniformly over compact sets.[8][20]

Chen's identity says the signature is a homomorphism from paths under concatenation to the tensor algebra under multiplication. Universality is a separate fact. The shuffle identity makes the product of two linear functionals of the signature another linear functional. The signature of a time-augmented path determines the path. Stone-Weierstrass then applies to this point-separating algebra of functions on path space.

Proposition (Signature depth and resolution).

Let be the price path coarse-grained at resolution with per-step volatility . The depth- term of the signature of one step has magnitude . The effective depth at which the -th term falls below a tolerance relative to the depth-1 term is

which is increasing in for . Coarser resolution requires deeper signatures per step.

Proof. On one step the piecewise-linear path moves by the increment of size , and the signature of a straight segment is . The depth- term therefore has norm , and the ratio of successive depth norms is at most . Setting and solving gives the stated formula, up to the factorials.

At the 2025 data give (BTC) to (ETH), and for : at millisecond resolution each step is described by its increment. At with to , : depth-3 and higher terms are negligible, and each hourly step is well described by its level-1 and level-2 signature components.

9.5 Common Thread

Proposition (Resolution as a universal model hyperparameter).

For any model class trained on observations of the price process at resolution , let be the model minimising KL divergence to the true coarse-grained law . Suppose for every (the class is well specified at every resolution). Then , and is non-constant whenever the rescaled laws depend on .

Proof. The KL divergence is non-negative and is zero if and only if the two distributions are equal. Under well-specification its minimum over is zero and is attained only at . Hence whenever in the resolution manifold (Definition 3). A variance ratio or autocorrelation that changes with (Section 6) shows that the rescaled laws differ.

The kernel length scale, SDE timestep, reservoir spectral radius and signature truncation depth are all parameterisations of this same hyperparameter. A principled approach to price modelling must either restrict to a specific resolution band or model the resolution dependence by treating as a parameter of the statistical model.

10. Synthesis

The four empirical investigations of Sections 4, 6, 7 and 8 measure one structure.

Variance ratio and Hurst. Relative to 1 ms the variance ratio rises to 2.22 (BTC) and 1.88 (ETH) at 1 s and peaks at 3.7 and 2.6 between 30 s and 2 minutes, far outside the bootstrap bands around 1. The local Hurst exponent is close to below 10 ms, near 0.6 at 200 ms, and crosses between 30 s and 2 minutes. Beyond that it lies between 0.46 and 0.50, and DFA over 1 minute to 1 day gives 0.495 and 0.491.

Autocorrelation structure. Signed volume is positively autocorrelated at every resolution and lag measured (0.051 and 0.066 at lag 1 ms). Returns at 1 ms are close to uncorrelated. The lag-1 return autocorrelation is 0.13 and 0.10 at 1 s and within 0.03 of zero at 1 minute and 1 hour.

Wavelet decomposition. Half of the 1 ms variance sits at the 1 ms scale and the profile follows white noise up to 32 ms. Coarser scales have up to 2.3 times the white-noise share, and the excess is smaller on volatile days.

Information geometry. Transfer entropy grows by two orders of magnitude from 1 ms to 1 s. Its direction changes near 2 ms on a day-by-day count, and from 5 ms upwards ETH to BTC dominates, by a factor of 3.7 at 1 s. The Gaussian family, a hyperbolic half-plane of curvature in the Fisher-Rao metric, describes the marginal law of returns only where the returns are close to independent, which here means scales of a minute and longer.

Summary. No single model is adequate across all resolutions. Between about 50 ms and 30 s both markets are persistent: returns and order flow are positively autocorrelated, wavelet variance exceeds the white-noise profile and ETH leads BTC in transfer entropy. At a minute and longer they are close to a random walk with a slight tendency to revert over minutes to hours. A model calibrated at one end of this range misstates the variance at the other by a factor of up to 3.7.

11. Conclusion

We began with the observation that the mid-price of a traded asset is the projection of a deterministic high-dimensional system onto an observable scalar, and the apparent stochasticity is generated by the temporal coarse-graining we perform when we bin trades into intervals of duration .

The empirical findings of this post quantify how the projected noise changes with . Between about 50 ms and 30 s the noise has memory: it is positively autocorrelated and has directional structure between assets. From a minute upwards the memory is nearly gone: the returns are close to uncorrelated. The local Hurst exponent rises from at 1 ms to about 0.6 at 200 ms and returns to between 30 s and 2 minutes.

The resolution dependence arises from how a deterministic agent-interaction system looks when observed through a temporal filter. Agents with information, objectives and constraints make the individual order decisions that form the price process. The randomness enters only when many such decisions are aggregated into a single log-return. At fine aggregation, the individual structure still shows through. At coarse aggregation, the central limit theorem has washed it away.

A model that specifies its resolution and makes claims only within that band makes a claim it can support. A model that claims resolution independence (the same drift, diffusion coefficient or kernel from 1 ms to 1 hr) makes an empirically false claim, as this post documents.

The state-space metric, the Fisher-Rao geometry of the statistical manifold, the diffusion distance on the data manifold and the wavelet variance decomposition of path space formalise the same idea: the model is a function of . The measurements here do not identify a curvature. They show that the flat random-walk model with scaling fails between 50 ms and 30 s and is an adequate approximation from a minute upwards. The task is to know which timescale applies.

References

  1. Lo, A. W. and MacKinlay, A. C. (1988). Stock market prices do not follow random walks: Evidence from a simple specification test. Review of Financial Studies, 1(1), 41-66.
  2. Bardet, J.-M. and Kammoun, I. (2008). Asymptotic properties of the detrended fluctuation analysis of long-range-dependent processes. IEEE Transactions on Information Theory, 54(5), 2041-2052.
  3. Beran, J. (1994). Statistics for Long-Memory Processes. Chapman and Hall. Spectral characterization of long memory.
  4. Doob, J. L. (1953). Stochastic Processes. Wiley. Chapter 11: characterization of Gaussian Markov processes.
  5. Black, F. and Scholes, M. (1973). The pricing of options and corporate liabilities. Journal of Political Economy, 81(3), 637-654.
  6. Avellaneda, M. and Stoikov, S. (2008). High-frequency trading in a limit order book. Quantitative Finance, 8(3), 217-224.
  7. Gatheral, J., Jaisson, T., and Rosenbaum, M. (2018). Volatility is rough. Quantitative Finance, 18(6), 933-949.
  8. Lyons, T. J. (1998). Differential equations driven by rough signals. Revista Matematica Iberoamericana, 14(2), 215-310.
  9. Mallat, S. G. (1989). A theory for multiresolution signal decomposition: The wavelet representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 11(7), 674-693.
  10. Percival, D. B. and Walden, A. T. (2000). Wavelet Methods for Time Series Analysis. Cambridge University Press.
  11. Coifman, R. R. and Lafon, S. (2006). Diffusion maps. Applied and Computational Harmonic Analysis, 21(1), 5-30.
  12. Amari, S. (1985). Differential-Geometrical Methods in Statistics. Springer. Fisher-Rao metric and the KL divergence as natural divergence to second order (Chapter 3).
  13. Amari, S. (2016). Information Geometry and Its Applications. Springer.
  14. Wiener, N. (1930). Generalized harmonic analysis. Acta Mathematica, 55, 117-258.
  15. Khinchin, A. (1934). Korrelationstheorie der stationaren stochastischen Prozesse. Mathematische Annalen, 109, 604-615.
  16. Li, X., Wong, T.-K. L., Chen, R. T. Q., and Duvenaud, D. (2020). Scalable gradients for stochastic differential equations. AISTATS 2020.
  17. Kidger, P., Foster, J., Li, X., Oberhauser, H., and Lyons, T. (2021). Neural SDEs as infinite-dimensional GANs. ICML 2021.
  18. Jaeger, H. (2001). The echo state approach to analysing and training recurrent neural networks. GMD Technical Report 148.
  19. Maass, W., Natschlaeger, T., and Markram, H. (2002). Real-time computing without stable states: A new framework for neural computation based on perturbations. Neural Computation, 14(11), 2531-2560.
  20. Chevyrev, I. and Kormilitzin, A. (2016). A primer on the signature method in machine learning. arXiv:1603.03788.
  21. Cont, R. (2001). Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance, 1(2), 223-236.
  22. Lo, A. W. (1991). Long-term memory in stock market prices. Econometrica, 59(5), 1279-1313.
  23. Roll, R. (1984). A simple implicit measure of the effective bid-ask spread in an efficient market. Journal of Finance, 39(4), 1127-1139. Derives Cov(Δpt, Δpt-1) = -s²/4 for the bid-ask bounce; applied to NYSE stocks.
  24. Andersen, T. G., Bollerslev, T., Diebold, F. X., and Ebens, H. (2001). The distribution of realized stock return volatility. Journal of Financial Economics, 61, 43-76. Realized volatility for all 30 DJIA components from 5-minute returns.