The "Rough" HAR Model
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "The "Rough" HAR Model".
Paper summary: Ruth: So we're looking at this paper, "The Rough HAR Model," and it tackles how we approximate realized variance, which is basically how we measure volatility over time. The authors are proposing these discrete-time approximations to these rough continuous-time models Ruth They argue that the leading rough models can actually be viewed as autoregressive processes driven by fractional Gaussian noise.
Warren: That sounds like they're trying to find a simpler way to handle some really complex volatility patterns, which is what the abstract says. They show that when you look at this fractional Gaussian noise, its dependence really concentrates at the first lag when the Hurst parameter is below one half Warren And they claim that augmenting standard autoregressive or heterogeneous autoregressive models with a first-order moving average component approximates this roughness.
Iris: So what's the core mechanism here? They say that this first-order moving average component, which is an MA one, actually maps almost linearly into the Hurst parameter Iris This means they refer to these extensions as the "rough" AR and "rough" HAR models.
Ruth: Exactly. The authors show that when they fit these "rough" models to data simulated from fractional Brownian motion and the fractional Ornstein–Uhlenbeck process, they find that the estimated MA coefficient actually recovers the theoretical mapping to that Hurst parameter Ruth It's a direct link between an observed coefficient in a simple model and some underlying property of how persistent the returns are.
Warren: And when they look at ten ETFs, they report finding negative MA coefficients for every single asset Warren And these implied Hurst parameters line up closely with the estimates you get from those continuous-time models Warren That's pretty tight agreement between the two frameworks.
Iris: I wonder what variation is actually doing all this work here? They're using a discrete-time framework that inherits its structure from the ARMA toolbox, which is nice for estimation because it's established Iris But they are dealing with processes like fBm and fOU, which aren't even semimartingales, making estimation difficult in the continuous setting Iris
Ruth: That's a fair point. The authors acknowledge that these continuous-time fractional models are hard to apply empirically because of those estimation and extension barriers Ruth So they propose this discrete approach as a way around that difficulty, inheriting tools from the ARMA toolbox, which makes it more accessible Ruth
Warren: From a trading perspective, if you use these "rough" models on out-of-sample comparisons against their classical counterparts, they outperform them for nearly every asset and horizon Warren And the biggest gains they see are at short horizons Warren That suggests this approximation actually helps in making better short-term predictions.
Iris: That's interesting because usually, when you simplify a model like this, you expect some kind of loss of information or accuracy compared to the more complex continuous-time version Iris But they claim their accuracy is comparable to the rough continuous-time models Iris
Paper summary: Ruth: So it seems the simplification isn't a huge sacrifice in predictive power for these specific applications. They say that even when they look at measurement error, which can sometimes be confused with roughness, fixing that noise variance restores identification of an ARMA one with a negative MA coefficient Ruth The estimated MA coefficients stay around negative zero point four for every ETF even when you account for measurement error Ruth
Warren: So it sounds like the authors are saying that the roughness they see isn't just noise in their measurement, but something real that needs this specific ARMA structure to capture properly Warren It shifts the focus from just noise being there to needing a specific structure, which is important for strategy design.
Iris: The paper does mention that the negative MA coefficient in realized variance can also be caused by measurement error inducing an ARMA one point one structure in the observed series Iris They provide two pieces of evidence to show that just the noise itself isn't enough to explain all this roughness Iris
Ruth: And they give their second piece of evidence by using asymptotic theory for realized variance to estimate a variable, theta, in the presence of noise by fixing sigma squared w Ruth This process leads to that ARMA one point one structure with a negative MA coefficient Ruth So the identification comes from accounting for measurement error in a specific way.
Warren: It’s interesting that they use this framework to show that the HARQ model from Bollerslev, Patton, and Quaedvlieg actually delivers none of those gains you see in these rough models Warren That's a direct comparison against existing work.
Iris: And they find that the authors' rough models outperform their classical counterparts for nearly every asset and horizon in those out-of-sample comparisons Iris That suggests a tangible improvement over what is currently available.
Ruth: So to sum up, this paper, "The Rough HAR Model," proposes these simple ARMA-type models as discrete-time approximations for the rough continuous-time models of realized variance Ruth The main finding is that the estimated MA coefficients are negative for every ETF and the implied Hurst parameters match the continuous-time ones Ruth
Warren: And in terms of what this means, it shows that these "rough" models actually outperform their classical counterparts for nearly every asset and horizon when forecasting Warren Especially at short horizons where those gains are largest.
Iris: The paper concludes by comparing the performance of various models under both mean squared error and qlike loss functions across multiple horizons Iris And it leaves room for future extensions, like adding time-varying parameters or multivariate systems, which is good because the framework itself is linear and parsimonious Iris
Ruth: So in simple terms, they’ve created a more practical tool. They took a very hard problem in continuous time volatility modeling and made it into a simple discrete-time ARMA structure that works well empirically Ruth It gives us a way to get better forecasts without needing the heavy machinery of those fractional calculus models.
Conclusion: Ruth: So we're wrapping up on this paper, "The Rough HAR Model." Basically, they’ve taken these complex continuous-time volatility models and made them simpler, discrete-time approximations using ARMA structures.
Warren: Yeah, so the main thing is that they’ve shown you can get a decent forecast by just adding one moving average component to your standard autoregressive or heterogeneous autoregressive model. It's a shortcut for modeling roughness.
Iris: The core finding is that these "rough" models actually perform better than the classical ones when you test them on real data, like ten ETFs. They show that the estimated moving average coefficient maps directly to a specific parameter called the Hurst parameter.
Ruth: That mapping is key because it connects something in a simple model to how persistent or erratic the underlying volatility really is, which we've been trying to measure for years.
Warren: And they found negative coefficients for every single ETF, which implies that this roughness they are capturing is actually systematic across the board, not just random noise.
Iris: But here’s where I look at it closely—they have to account for measurement error in realized variance, because if you don't fix that noise variance, the identification breaks down and you get a different structure.
Ruth: Exactly. So while the rough models are powerful approximations of the continuous-time ideas, they aren't perfect without accounting for what we measure versus what’s actually happening in the market.
Warren: And despite those caveats about measurement error, they still show these rough models give better out-of-sample performance than the standard ones across almost every asset and time frame.
Iris: It means for someone building a trading strategy, this is a much simpler framework to implement than trying to solve fractional calculus problems directly.
Ruth: So it boils down to taking some high-level theory and making it into an ARMA structure that actually works in practice for forecasting realized variance. This suggests we might be able to get closer to understanding market microstructure without needing the heavy machinery of continuous time models.
Aarhus University · Aarhus Center for Econometrics (ACE) · Center for Research in Energy: Economics and Markets (CoRE) · Danish Finance Institute (DFI) · University of Macau
econ.EM
Submitted: 2026-09-18
Updated: 2026-11-18
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: This paper proposes discrete-time approximations to rough continuous-time models of realized variance and shows that augmenting autoregressive (AR) and heterogeneous autoregressive (HAR) models with…
Key concepts
- Rough Models
- These are discrete-time approximations of continuous-time fractional models like fBm and fOU. The key idea is that the dependence structure of these processes concentrates its influence at the first lag, which allows for a simplified model structure.
- MA(1) Component
- A first-order moving average component with a negative coefficient acts as a natural proxy for fractional Gaussian noise (fGn). Combining this MA(1) term with an AR structure creates the 'rough' models that capture the essential roughness of the underlying continuous-time process.
- Hurst Parameter (H)
- This parameter characterizes the long-term dependence in fractional processes. The paper shows that when H is below 0.5, the Wold representation of fGn concentrates its dependence at lag one, meaning a negative MA coefficient maps almost linearly to this Hurst parameter.
Terminology
Summary
This paper proposes discrete-time approximations to rough continuous-time models of realized variance and shows that augmenting autoregressive (AR) and heterogeneous autoregressive (HAR) models with a first-order moving-average (MA(1)) component approximates roughness, leading to rough
AR and rough
HAR models that outperform classical counterparts in forecasting
Model Construction
The authors propose that the leading rough models can be viewed as autoregressive processes driven by fractional Gaussian noise The task is then to approximate the fGn, and they show that the Wold representation of fGn concentrates its dependence at the first lag when the Hurst parameter is below 0.5 A first-order moving average (MA(1)) with a negative coefficient is a natural proxy for fGn Combining the MA(1) component with an AR(1) or a HAR structure yields what they call the “rough” AR and “rough” HAR models
Theoretical Foundation
The continuous-time fractional models like fBm and fOU process can be viewed as an autoregressive process of order one (AR(1)) driven by fractional Gaussian noise (fGn) The dependence structure of fGn is distinctive, characterized by a sharp drop in the autocovariance function around the origin This sharp drop after the first lag is also the signature of a low-order MA process The Wold coefficients for fGn decay at a hyperbolic rate, and when H is in the rough range, this decay implies that the expansion can be truncated after lag one
Empirical Validation
The authors show that when they fit the “rough” models to data simulated from fBm and fOU, the estimated MA coefficient recovers the theoretical mapping to the Hurst parameter When applied to ten ETFs, they find negative MA coefficients for every asset, and the implied Hurst parameters align closely with those from continuous-time models In out-of-sample comparisons, the “rough” models outperform their classical counterparts for nearly every asset and horizon
Forecasting Performance
The rough models outperform their classical counterparts for nearly every asset and horizon, with the largest gains at short horizons, and their accuracy is comparable to that of the rough continuous-time models The log-RHAR model has a lower QLIKE than the log-HAR model for every ETF at every horizon The results are reported in Tables 4 and 5, showing that the rough models outperform their classical counterparts for nearly every asset and horizon
Limitations and Extensions
The continuous-time fractional models are difficult to apply in empirical work due to estimation and extension barriers The discrete-time framework proposed here inherits its extensions off the shelf from the ARMA toolbox The model is linear and parsimonious, allowing for use of a wide range of established estimation techniques and easy extension through time-varying parameters or structural breaks
Measurement Error Consideration
The negative MA coefficient in RV can also arise from measurement error, which induces an ARMA(1,1) structure in the observed series The authors provide two pieces of evidence suggesting that noise alone cannot fully explain the roughness observed in realized variance Fixing the noise variance restores identification, and this leads to an ARMA(1,1) structure with a negative MA coefficient The estimates remain around −0.4 for every ETF when accounting for measurement error
Conclusion
In this paper, they propose simple ARMA-type models as discrete-time approximations to rough continuous-time models of realized variance The estimated MA coefficients are negative for every ETF, and the implied Hurst parameters obtained from the “rough” AR are in line with those of the continuous-time models In forecasting, the “rough” models outperform their classical counterparts for nearly every asset and horizon The simplicity of the framework leaves ample scope for extensions, such as additional predictors, time-varying parameters, and multivariate systems
Disclosure Statement
The authors declare no conflicts of interest The data are from VOLARE and are publicly available at https://volare. p. 6 p. 6 p. 6 p. 6 p. 6 p. 6 p. 6unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks. p. 0 p. 0 p. 0 p. 0 p. 0 The appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. p. 3 p. 3 p. 3 p. 3 p. 3 The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. p. 2 p. 2 p. 2 p. 2 p. 2 p. 2 p. 2 The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks. The appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks. The appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks. The appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks. The appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks and the appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks The data are from VOLARE and are publicly available at https://volare. p. 2unime.it.
Pruning redundant reasoning tokens
The authors propose that the Wold representation of this noise concentrates its dependence at the first lag when the Hurst parameter is below one half Augmenting the autoregressive (AR) and heterogeneous autoregressive (HAR) models with a first-order moving-average (MA(1)) component therefore approximates the roughness, and the MA coefficient maps almost linearly into the Hurst parameter This leads to rough
AR and rough
HAR models
How the training signal is built
The starting point is that fBm and the fOU process can be viewed as an autoregressive process of order one (AR(1)) driven by fractional Gaussian noise (fGn) The task is then to approximate the fGn, and they show that the Wold representation of fGn concentrates its dependence at the first lag when the Hurst parameter is below 0.5 The first Wold coefficient is negative, and close to linear in the Hurst parameter Hence, a first-order moving average (MA(1)) with a negative coefficient is a natural proxy for fGn
Where the gains come from
The rough models outperform their classical counterparts for nearly every asset and horizon, with the largest gains at short horizons The first lag therefore does almost all of the work, justifying cutting the expansion at the first lag The estimated MA coefficients are negative for every asset, and the implied Hurst parameters obtained from the “rough” AR are in line with those of the continuous-time models In forecasting, the “rough” models outperform their classical counterparts for nearly every asset and horizon
Measurement Error Analysis
The authors provide two pieces of evidence suggesting that noise alone cannot fully explain the roughness observed in realized variance The HARQ model of Bollerslev, Patton, and Quaedvlieg (2016) delivers none of the gains of the rough models The second piece of evidence uses asymptotic theory for RV to estimate θ in the presence of noise Fixing σ 2 w restores identification, and this leads to an ARMA(1,1) structure with a negative MA coefficient The estimated MA coefficients are negative for every ETF when accounting for measurement error
The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons The simple framework leaves ample scope for extensions, such as additional predictors, time-varying parameters, and multivariate systems
The authors declare no conflicts of interest The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks and the appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. The data are from VOLARE and are publicly available at https://volare.unime.it. The references list the seminal works such as Andersen, Bollerslev, Diebold, and Labys (2001) and Bennedsen, Christensen, and Christensen (2026). The paper includes simulation studies comparing the rough models to continuous-time benchmarks across various horizons and detailed out-of-sample forecast accuracy results for individual stocks and the appendix provides additional empirical and technical details, including parameter recovery results for the fOU process with alternative κ values and a summary of how to recovering Wold coefficients from the spectral density via Szegő factorization. The paper concludes by comparing the performance of various models under MSE and QLIKE loss functions across multiple horizons and presenting results for a panel of 40 individual stocks. The data are from VOLARE and are publicly available at https://volare.unime.it.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Discrete-time approximation of rough continuous-time models allows for estimation using standard ARMA toolbox extensions, specifically by augmenting autoregressive (AR) and heterogeneous autoregressive (HAR) models with a first-order moving-average (MA(1)) component to approximate roughness. This enables the estimation of
rough
AR andrough
HAR models on log RV data, which are easier to estimate by standard off-the-shelf software. -
The MA coefficient in the rough models maps almost linearly into the Hurst parameter, allowing for a direct mapping between this coefficient and H. This relationship is specifically noted as being
monotonic and almost linear
when plotting c1 with H in simulations, suggesting that c1 isparticularly informative about the Hurst parameter.
-
The rough AR model provides a specification that is
linear and involves only four parameters, omega, rho, theta, and sigma,
making it suitable for use with a wide range of established estimation techniques and easily extendable through time-varying parameters or structural breaks. -
The rough models outperform their classical counterparts in out-of-sample comparisons for nearly every asset and horizon, with
substantial gains at short horizons,
demonstrating improved accuracy comparable to continuous-time rough models like fBm and fOU. -
The model confidence set (MCS) procedure can be used to identify the best performing models by assigning p-values, which helps in selecting the most robust specification among competing models under different loss functions like MSE and QLIKE.
Sources
- Modeling and Forecasting Realized Volatility with Multivariate Fractional Brownian Motion
- VOLatility Archive for Realized Estimates (VOLARE)
- Optimal Estimation for General Gaussian Processes
Related papers
- Stationarity with Occasionally Binding Constraints
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs