A0327
Title: A novel hybrid ensemble approach to forecast freight rates volatility
Authors: Morten Risstad - Norwegian University of Science and Technology (Norway) [presenting]
Malvina Marchese - City University of London (United Kingdom)
amir alizadeh - city university (United Kingdom)
Abstract: The contribution is to the forecasting literature in three important ways. First, the performance of a variety of machine learning algorithms in forecasting freight rates volatility is investigated. Second, an extensive forecasting comparison between traditional GARCH models and machine learning methods is conducted. The aim is not only to make a complete comparison across traditional and ML methods but also to provide evidence of when and why some of these methods improve the accuracy of forecasting volatility. Findings suggest that substantial incremental information about future volatility can be extracted with ML from additional volatility predictors with minimal noise fitting if regularization is applied. In contrast, the GARCH-MIDAS and GARCH-X models yield only minor improvements. However, traditional GARCH models do a better job of capturing the long-range persistence of the volatility. When deep learning models are compared to the benchmark FIGARCH model, the average MSE for FIGARCH is lower, and the crucial driver for this superior performance is the more effective way to capture fractional integration. Thus, the findings prompt the proposal of a novel hybrid ensemble stacking algorithm that combines GARCH models and tree-based algorithms. Its superior forecasting performance is established at several horizons, including 1, 5, 22 and 60 days ahead according to statistical (MCS and SPA tests) and economic loss functions (value at risk).
