Research on ETF prediction based on Bi-LSTM-Bi-GRU

• Propose a parallel Bi-LSTM–Bi-GRU model for more accurate ETF price forecasting. • Outperforms the latest models, such as GRU, TCN, and Transformer, on key metrics (MAE, RMSE, R²). • Effectively integrates Bi-LSTM’s long-term trend modeling with Bi-GRU’s short-term volatility capture. • Demonstrates robust performance during periods of high volatility and price turning points. • Shows strong generalization across different markets and ETF categories. Accurate prediction of ETF price trends is of great significance to investors' decision-making, risk management of financial institutions, and the stable operation of the market. In view of the non-stationarity of ETF price series and the complexity of the relationship between volume and price, as well as the limitations of traditional prediction methods in dealing with complex non-linear relationships and multi-dimensional input features, a combined model based on Bi-LSTM-Bi-GRU is proposed. The model design adopts a dual-branch parallel structure: the Bi-LSTM branch extracts long-term trend dependencies from the price series through bidirectional long short-term memory units, while the Bi-GRU branch captures short-term fluctuations driven by trading volume using the gating mechanism. The outputs of the two branches are fused through feature concatenation to integrate long- and short-term information. The model architecture includes an input layer, Bi-LSTM layer, Bi-GRU layer, feature fusion layer, fully connected layer, and output layer. Through feature fusion, the model can leverage the different features extracted by LSTM and GRU to enhance the accuracy of ETF price prediction. In this experiment, the training, validation, and testing sets were divided in a 6:2:2 ratio. The mean squared error (MSE) was used as the loss function. Hyperparameters were optimized through grid search, and learning rate scheduling and early stopping mechanisms were introduced to enhance training stability and generalization capability.Experimental studies have shown that the Bi-LSTM-Bi-GRU model significantly outperforms the comparison models in terms of MAE, RMSE, and R2, especially during periods of severe market volatility when its bidirectional structure can more accurately predict price turning points.