A0817
Title: Realized volatility forecasting for new issues and spin-offs using multi-source transfer learning
Authors: Andreas Teller - Friedrich Schiller University Jena (Germany) [presenting]
Uta Pigorsch - University of Wuppertal (Germany)
Christian Pigorsch - Friedrich Schiller University Jena (Germany)
Abstract: The special case of forecasting realized volatility of financial assets with limited historical data is considered, such as new issues or spin-offs. Typically, realized volatility forecasting models rely on a sufficient history of data. For new issues and spin-offs, however, an extensive data history is not directly available. Therefore, the proposal is to forecast the realized variance of assets with limited historical data based on multi-source transfer learning. Specifically, complementary source data of financial assets is exploited with a substantial historical data record by selecting source time series instances most similar to the target data of the respective new issue or spin-off. Based on these instances and the target data, heterogeneous autoregressive models, feedforward neural networks, and extreme gradient boosting models are estimated. Their forecasting performance is compared to forecasts of the same models trained exclusively on the target data and to a simplified training data pooling approach that includes the entire source and target data. Results indicate that integrating complementary data can significantly improve the accuracy of realized variance forecasts for new issues and spin-offs, even shortly after their initial trading day. In particular, the proposed transfer learning approach shows superior performance compared to models trained solely on target asset data and those that additionally incorporate the complete source data.
