Paper: arXiv 2406.08041
Authors: Francesco Audrino, Jonathan Chassot
Abstract
We investigate the predictive abilities of the heterogeneous autoregressive (HAR) model compared to machine learning (ML) techniques across an unprecedented dataset of 1,455 stocks. Our analysis focuses on the role of fitting schemes, particularly the training window and re-estimation frequency, in determining the HAR model’s performance. Despite extensive hyperparameter tuning, ML models fail to surpass the linear benchmark set by HAR when utilizing a refined fitting approach for the latter. Moreover, the simplicity of HAR allows for an interpretable model with drastically lower computational costs. We assess performance using QLIKE, MSE, and realized utility metrics, finding that HAR consistently outperforms its ML counterparts when both rely solely on realized volatility and VIX as predictors. Our results underscore the importance of a correctly specified fitting scheme. They suggest that properly fitted HAR models provide superior forecasting accuracy, establishing robust guidelines for their practical application and use as a benchmark. This study not only reaffirms the efficacy of the HAR model but also provides a critical perspective on the practical limitations of ML approaches in realized volatility forecasting.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper relies on straightforward autoregressive and statistical models with minimal advanced mathematics, but demonstrates high empirical rigor through extensive backtesting on a large dataset (1,445 stocks), comprehensive hyperparameter tuning, and robust error metrics (MSE, QLIKE, utility).
Research Flowchart
flowchart TD
A["Research Goal<br>Assess HAR vs. ML Models<br>for Volatility Forecasting"] --> B["Dataset<br>1,455 Stocks<br>Realized Vol & VIX"]
B --> C["Modeling Approaches"]
C --> D["HAR Model<br>Linear Benchmark"]
C --> E["Machine Learning<br>Models"]
D --> F["Optimized Fitting<br>Window & Re-estimation"]
E --> G["Extensive<br>Hyperparameter Tuning"]
F --> H{"Performance Evaluation<br>Metrics: QLIKE, MSE, Utility"}
G --> H
H --> I["Key Findings<br>HAR Outperforms ML<br>Low Computational Cost<br>Correct Fitting is Crucial"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes that a carefully fitted HAR model provides superior forecasting accuracy compared to machine learning methods.
Source sentence
Our analysis focuses on the role of fitting schemes, particularly the training window and re-estimation frequency, in determining the HAR model's performance
Assumes 4
- HAR models are estimated using a rolling window with regular re‑estimation. model
Source sentence
HAR models are typically estimated in a rolling window fashion, where the model is re-estimated at regular intervals using a fixed (or increasing) number of past observations.
- Transaction costs are proxied by the rolling median bid‑ask spread over the last nine months. execution
Source sentence
In this setting, we follow Bollerslev et al. (2018) and use the rolling median of the bid-ask spread over the last nine months as a proxy for transaction costs.
- Realized utility is interpreted as the percentage of wealth an investor would pay for access to the model’s forecasts. other
Source sentence
the values may be interpreted as the percentage of wealth that an investor would be willing to pay in order to access a portfolio that leverages the model’s forecasting capabilities compared to a portfolio that uniquely allocates capital in the risk‑free asset.
- The analysis uses only realized volatility and VIX as predictors for all models. data
Source sentence
HAR consistently outperforms its ML counterparts when both rely solely on realized volatility and VIX as predictors.
Data
- CRSP U.S. Stock Database market-prices public
Source sentence
We query high-frequency data for a total of 6,061 stocks. Among these, 1,445 stocks remain consistently present throughout the nine years.
- NYSE Trade and Quote (TAQ) database order-book commercial
Source sentence
We source high-frequency intraday trade data for all selected stocks from the TAQ database from January 1, 2015, to October 31, 2023.
Costs
Modeled
rolling median of the bid‑ask spread over the last nine months used as proxy
Source sentence
In this setting, we follow Bollerslev et al. (2018) and use the rolling median of the bid-ask spread over the last nine months as a proxy for transaction costs.
Finds 4
- HAR consistently outperforms ML models in predictive accuracy across QLIKE, MSE, and realized utility metrics. predictive-accuracy
Source sentence
HAR consistently outperforms its ML counterparts when both rely solely on realized volatility and VIX as predictors.
- The HAR model estimated via WLS yields the highest realized utility, even when transaction costs are included, regardless of VIX inclusion. performance
Source sentence
the HAR model estimated via WLS markedly outperforms the other models, regardless of whether VIX is included.
- Larger training windows generally lead to lower RMSE for the HAR model. statistical
Source sentence
larger training windows generally lead to lower prediction errors.
- ML models underperform relative to HAR, with the gap more pronounced for DJIA and Nasdaq‑100 subsets. negative
Source sentence
the underperformance of ML models is even more pronounced in these subsets.
Checked by others
Supported by
- NeuralFactors: A Novel Factor Learning Approach to Generative Modeling of Equities
Compare the findings
This paper: Larger training windows generally lead to lower RMSE for the HAR model.
That paper: Increasing training data length improves performance, contradicting the belief that non‑stationarity requires shorter windows.
- Stock Price Prediction Using a Hybrid LSTM-GNN Model: Integrating Time-Series and Graph-Based Analysis
Compare the findings
This paper: Larger training windows generally lead to lower RMSE for the HAR model.
That paper: Expanding window training improved adaptability, mitigated concept drift, and enhanced generalisation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (46)
- Stochastic Volatility and Realized Stochastic Volatility Models (2023) — link
- Out-of-sample volatility prediction: Rolling window, expanding window, or both? (2024) — link
- Approximation capabilities of multilayer feedforward networks (1991) — link
- The Elements of Statistical Learning (2009) — link
- Ridge Regression: Biased Estimation for Nonorthogonal Problems (1970)
- Random Forests (2001) — link
- Greedy function approximation: A gradient boosting machine (2001) — link
- Risk Everywhere: Modeling and managing volatility (2018) — link
and 38 more
- Bagging Predictors (1996) — link
- Forecasting of Realised Volatility with the Random Forests Algorithm (2018) — link
- Forecasting and trading high frequency volatility on large indices (2018) — link
- Prediction of realized volatility and implied volatility indices using AI and machine learning: A review (2024) — link
- Novel hybrid model based on echo state neural network applied to the prediction of stock price return volatility (2021) — link
- Lassoing the HAR Model: A Model Selection Perspective on Realized Volatility Dynamics (2016) — link
- Complete subset least squares support vector regression (2021) — link
- A hybrid approach of adaptive wavelet transform, long short-term memory and ARIMA-GARCH family models for the stock index prediction (2021) — link
- Regression Shrinkage and Selection Via the Lasso (1996) — link
- Statistical Learning with Sparsity (2015) — link
- Good volatility, bad volatility: Signed jumps and the persistence of volatility (2015) — link
- A Machine Learning Approach to Volatility Forecasting (2023) — link
- Volatility Models and Their Applications (2012)
- Realised variance forecasting under Box-Cox transformations (2017) — link
- Direct Versus Iterated Multiperiod Volatility Forecasts (2019) — link
- Predicting excess stock returns out of sample: Can anything beat the historical average? (2008) — link
- HAR Modeling for Realized Volatility Forecasting (2012)
- Forecasting global equity market volatilities (2020) — link
- Global equity market volatility spillovers: A broader role for the United States (2016) — link
- A Practical Guide to harnessing the HAR volatility model (2021) — link
- Modeling and forecasting realized volatility (2003) — link
- The distribution of exchange rate volatility (2001)
- Estimating quadratic variation using realized variance (2002) — link
- Realized volatility forecasting: Robustness to measurement errors (2021) — link
- Is implied volatility more informative for forecasting realized volatility: An international perspective (2020) — link
- Increasing the information content of realized volatility forecasts (2023) — link
- Trading and non-trading period realized market volatility: Does it matter for forecasting the volatility of US stocks? (2020) — link
- Forecasting realized volatility with machine learning: Panel data perspective (2023) — link
- The Model Confidence Set (2011) — link
- Volatility Forecasting with Machine Learning and Intraday Commonality (2023) — link
- Econometric analysis of realized volatility and its use in estimating stochastic volatility models (2002) — link
- Volatility forecast comparison using imperfect volatility proxies (2011) — link
- Realized kernels in practice: trades and quotes (2009) — link
- The impact of sentiment and attention measures on stock market volatility (2020) — link
- Deep Learning (2016) — link
- Realized Volatility Forecasting with Neural Networks (2020) — link
- A simple approximate long-memory model of realized volatility (2009) — link
- Does anything beat 5-minute RV? A comparison of realized measures across multiple asset classes (2015) — link
0 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.