<abstract xmlns="http://www.w3.org/1999/xhtml"><p> Background: Since high-frequency data have become available, an unbiased volatility estimator, i.e. realized variance (RV) can be computed. Commonly used models for RV forecasting suffer from strong persistence with a high sensitivity to the returns distribution assumption and they use only daily returns. Objectives: The main objective is measurement and forecasting of RV. Two approaches are compared: Heterogeneous AutoRegressive model (HAR-RV) and Feedforward Neural Networks (FNNs). Even though HAR-RV-type models describe RV stylized facts very well, they ignore its nonlinear behaviour. Therefore, FNN-HAR-type models are developed. Methods/Approach: Firstly, an optimal sampling frequency with application to the DAX index is chosen. Secondly, in and out of sample predictions within HAR models and FNNs are compared using RMSE, AIC, the Wald test and the DM test. Weights of FNN-HAR-type models are estimated using the BP algorithm. Results: The optimal sampling frequency of RV is 10 minutes. Within HAR-type models, HAR-RV-J has better, but not significant, forecasting performances, while FNN-HAR-J and FNNLHAR- J have significantly better predictive accuracy in comparison to the FNN-HAR model. Conclusions: Compared to the traditional ones, FNN-HAR-type models are better in capturing nonlinear behaviour of RV. FNN-HAR-type models have better accuracy compared to traditional HAR-type models, but only on the sample data, whereas their out-of-sample predictive accuracy is approximately equal.</p></abstract>

Background: Since high-frequency data have become available, an unbiased volatility estimator, i.e. realized variance (RV) can be computed. Commonly used models for RV forecasting suffer from strong persistence with a high sensitivity to the returns distribution assumption and they use only daily returns. Objectives: The main objective is measurement and forecasting of RV. Two approaches are compared: Heterogeneous AutoRegressive model (HAR-RV) and Feedforward Neural Networks (FNNs). Even though HAR-RV-type models describe RV stylized facts very well, they ignore its nonlinear behaviour. Therefore, FNN-HAR-type models are developed. Methods/Approach: Firstly, an optimal sampling frequency with application to the DAX index is chosen. Secondly, in and out of sample predictions within HAR models and FNNs are compared using RMSE, AIC, the Wald test and the DM test. Weights of FNN-HAR-type models are estimated using the BP algorithm. Results: The optimal sampling frequency of RV is 10 minutes. Within HAR-type models, HAR-RV-J has better, but not significant, forecasting performances, while FNN-HAR-J and FNNLHAR- J have significantly better predictive accuracy in comparison to the FNN-HAR model. Conclusions: Compared to the traditional ones, FNN-HAR-type models are better in capturing nonlinear behaviour of RV. FNN-HAR-type models have better accuracy compared to traditional HAR-type models, but only on the sample data, whereas their out-of-sample predictive accuracy is approximately equal.

Neural Network Approach in Forecasting Realized Variance Using High-Frequency Data

Business Systems Research Journal

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Background: Since high-frequency data have become available, an unbiased volatility estimator, i.e. realized variance (RV) can be computed. Commonly used...

Neural Network Approach in Forecasting Realized Variance Using High-Frequency Data

Published Online: Jul 28, 2018

Page range: 18 - 34

Received: Jan 29, 2018

Accepted: Apr 21, 2018

DOI: https://doi.org/10.2478/bsrj-2018-0016

Keywords
high-frequency data, realized variance, nonlinearity, long memory, jumps, leverage, feedforward neural networks, Heterogeneous AutoRegressive model

© 2018 Josip Arnerić, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Neural Network Approach in Forecasting Realized Variance Using High-Frequency Data

Published Online: Jul 28, 2018

Page range: 18 - 34

Received: Jan 29, 2018

Accepted: Apr 21, 2018

DOI: https://doi.org/10.2478/bsrj-2018-0016

Keywordshigh-frequency data, realized variance, nonlinearity, long memory, jumps, leverage, feedforward neural networks, Heterogeneous AutoRegressive model

© 2018 Josip Arnerić, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Keywords
high-frequency data, realized variance, nonlinearity, long memory, jumps, leverage, feedforward neural networks, Heterogeneous AutoRegressive model