Meta-LSTR: Meta-Learning with Long Short-Term Transformer for futures volatility prediction

Futures are essential instruments in financial markets. Accurately predicting futures volatility is crucial for calculating value-at-risk and comprehensively assessing financial uncertainty. However, the rapid changes in the futures market, the continuous emergence of new commodities, and the close interaction with spot markets create a complex market environment. This results in futures data having intricate characteristics of limited historical data, non-stationary, and non-linear, posing significant challenges for accurately predicting volatility. We propose a futures volatility prediction framework, Meta-Learning with Long Short-Term Transformer (Meta-LSTR) to tackle these challenges. To improve the understanding of market dynamics, we construct a Long-Short Term Transformer network. In conjunction with a de-stationary module and market-side information, the network can effectively capture multi-scale non-stationary features and non-linear temporal dependencies. To enhance the efficiency of limited data utilization, we employ a meta-learning approach to extract common knowledge across different varieties of futures. Comprehensive experiments using Chinese market data highlight the effectiveness of the Meta-LSTR model in futures volatility prediction. Compared to other state-of-the-art methods, the proposed Meta-LSTR model reduces prediction error by over 21.99%. •We propose Meta-LSTR, using Meta-SGD and Transformer for fast volatility adaptation.•We construct a Transformer to extract multi-scale non-stationary features.•We introduce a feature enhancement method with price-volume data and market factors.•Our model achieves superior performance in the Chinese futures market.