Realized Volatility Forecasting for New Issues and Spin-Offs using Multi-Source Transfer Learning

TL;DR

This paper tackles volatility forecasting for assets with sparse histories, such as new issues and spin-offs, by introducing a model-agnostic multi-source transfer learning framework that selects informative source subsequences via dynamic time warping and integrates them with target data. The approach is evaluated with HAR, FNN, and XGBoost across standard and extended predictor sets, showing that multi-source transfer learning, especially with XGBoost and extended predictors (XGB-EXT), consistently improves 1-day-ahead realized variance forecasts relative to targets trained only on the scarce data or naively pooled data. The study also analyzes the properties of selected source subsequences and demonstrates the method’s value immediately after listing, as well as the benefits of progressively richer predictor sets as data accrue. Overall, the results highlight the practical potential of transfer learning to enhance volatility forecasts in data-scarce financial settings and suggest avenues for further optimization and broader application.

Abstract

Forecasting the volatility of financial assets is essential for various financial applications. This paper addresses the challenging task of forecasting the volatility of financial assets with limited historical data, such as new issues or spin-offs, by proposing a multi-source transfer learning approach. Specifically, we exploit complementary source data of assets with a substantial historical data record by selecting source time series instances that are most similar to the limited target data of the new issue/spin-off. Based on these instances and the target data, we estimate linear and non-linear realized volatility models and compare their forecasting performance to forecasts of models trained exclusively on the target data, and models trained on the entire source and target data. The results show that our transfer learning approach outperforms the alternative models and that the integration of complementary data is also beneficial immediately after the initial trading day of the new issue/spin-off.

Figures (6)

• Figure 1: Training and evaluation data sets for new issues and spin-offs across different sample periods $s$. Blue intervals represent exclusive training periods, while red intervals indicate evaluation periods during which all realized variance forecasting models are re-estimated every 5 trading days. An overview of the training and evaluation periods used in our forecast evaluation starting immediately after the first trading day is provided in Figure \ref{['fig:short_train_eval_schema']}.
• Figure 2: Example MTL model subsequence selection process after the 150th trading day of a new issue/spin-off. Depicted is the selection of subsequences from a single source asset for a subsequence size of $m=22$. The date of the source data observation $b$ corresponds to the 150th trading day of the new issue/spin-off. The DTW distance is computed based on the realized variance components. The number of selected subsequences is determined by the threshold parameter $\epsilon$; in the depicted example, only the most recent source subsequence is selected. This process is repeated at each estimation step for the progressively extended target and source series while maintaining temporal alignment.
• Figure 3: Training and evaluation data sets for new issues and spin-offs immediately following their first trading day across different sample periods ($s=1$, $s=5$, and $s=22$). Blue intervals represent exclusive training periods, while red intervals indicate evaluation periods during which all realized variance forecasting models are re-estimated daily in this setting.
• Figure 4: Average selection proportions of subsequences by MTL-25 models for each target asset, aggregated across all (re-)estimation steps in $s=s^{*}$. The selection rates are categorized based on the temporal distance between forecast origins and subsequence start dates, grouped into 100-trading-day intervals.
• Figure 5: Average selection proportions of subsequences by MTL-50 models for each target asset, aggregated across all (re-)estimation steps in $s=s^{*}$. The selection rates are categorized based on the temporal distance between forecast origins and subsequence start dates, grouped into 100-trading-day intervals.