CNN-HyperLSTM-TransformerXL: A Hybrid Deep Learning Model for Enhanced Futures Market Prediction

Being able to predict the trends in the futures market accurately is super important for adjusting economic policies and keeping the market stable. But because of the nonlinear and non-stationary nature of futures data, traditional methods have a hard time effectively building models for the extended temporal relationships and complex patterns in sequential data. This paper comes up with a new mixed deep learning model called CNN-HyperLSTM-TransformerXL (CH-TXL). It combines the good points of three advanced neural structures: convolutional neural networks, hyper-long short-term memory networks, and Transformer-XL. CNN gets local features to make it better at handling noise. HyperLSTM uses a dynamic gating system to make the memory more flexible. Transformer-XL takes advantage of extended attention to catch global dependencies. Also, the Hurst index is brought in as a feature to measure the self-similarity and long-term memory of the time series. It's calculated using the aggregated variance method. Using the futures data of the CSI 500 Index, the hyperparameters of the model are optimized through Bayesian optimization. The experimental results show that CH-TXL is better than models like CNN-LSTM, ARIMA-LSTM, and Transformer when it comes to prediction accuracy and generalization. Especially when dealing with long sequences and complex patterns. This model converges really well and the loss value stays low and stable, which shows that it works well.