A New Hybrid Model based on Non-Gaussian Autoregressive Process and Neural Network Model for Financial Market Prediction

Loading...
Thumbnail Image

Authors

Park, Jihyun

Issue Date

2024

Type

Dissertation

Language

en_US

Keywords

Heavy Tails , Hybrid Models , Non-Gaussian Models , Outliers , Time Series Modeling

Research Projects

Organizational Units

Journal Issue

Alternative Title

Abstract

In financial modeling, the assumption of Gaussian noise, has been a classical assumption. Models based on this assumption have been widely used in various fields due to their interpretability, e.g, Autoregression (AR), Moving Agerage models, (MA), Autoregressive Integrated Moving Average (ARIMA), etc. However, the Gaussian assumption is often violated in practice. Consequently, the use of time series models with non-Gaussian distributions, machine learning, and hybrid approaches has increased as alternatives. In order to improve the predictive accuracy of time series models, different approaches have been applied such as Autoregressive models assuming non-Gaussian innovations and hybrid models with traditional time series models and modern non-linear technique. Most of the hybrid models in the literature combine a classic ARIMA model and a machine learning method. This study aims to bridge non-Gaussian models and machine learning methods. Specifically, we proposes a hybrid combination model based on the Autoregressive process with Laplace innovations and neural network. Traditional autoregressive models have strong interpretability and simple architecture. The main strength of neural network is their wide range of uses and their flexibility. By combining their relative advantages, we can capture unpredictable behaviors of the data. Furthermore, since we use a Laplace distribution error as an alternative to the classic autoregressive model with Gaussian innovations, this makes the model more robust in the presence of outliers. In the first stage, AR model fitted, and linear predictions and residuals obtained. Next, perform NN on the residuals. Then, calculate the predicted residuals. The final predicted values are obtained by the summation of linear predictions and non-linearly predicted values. We also evaluate the performance in terms of RMSE and MAE of single models and hybrid models for comparison purposes. Our simulation generates synthetic data from three different classic non-linear time series models. The empirical application to the real datasets (Leading economic indicators), the results exhibit the proposed hybrid model shows better accuracy of forecasting compared to the other benchmark models. Although further investigation is still needed, the results demonstrate that a hybrid model incorporating a non-Gaussian component can serve as a practical tool for financial time series modeling. We also conduct a comprehensive analysis of the bond market to gain preliminary insights and explore the behavior of residuals taken from autoregressive models with leading economic indicators.

Description

Citation

Publisher

License

CC BY-NC-SA

Journal

Volume

Issue

PubMed ID

DOI

ISSN

EISSN