LASSO-PCA-LSTM Framework for Predicting Indonesian Government Bond Yields with Bayesian-Optimized LSTM

Abstract

Accurate government bond yield forecasting is important for investment decision-making and financial risk management, yet it remains challenging because of nonlinear yield behavior and complex macroeconomic interactions. This study proposes a hybrid forecasting framework combining Least Absolute Shrinkage and Selection Operator (LASSO), Principal Component Analysis (PCA), Bayesian Optimization, and Long Short-Term Memory (LSTM) to predict Indonesian fixed-rate government bond yields (FR0074, FR0075, and FR0076). Historical bond prices were converted into yield-to-maturity values and integrated with nine domestic and global macroeconomic indicators. Four forecasting scenarios—LSTM, LASSO-LSTM, PCA-LSTM, and LASSO-PCA-LSTM—were evaluated across three prediction horizons (t+1, t+5, and t+30). LASSO identified four influential predictors: the USD/IDR exchange rate, BI Rate, foreign exchange reserves, and the Fed Funds Rate. The LASSO-PCA-LSTM framework achieved the strongest predictive performance at the t+1 and t+5 horizons, with average R² values of 0.8685 and 0.1760, respectively, whereas PCA-LSTM performed relatively better at t+30. These findings indicate that the effectiveness of feature selection and dimensionality reduction varies across forecasting horizons. By combining predictor selection, dimensionality reduction, and Bayesian hyperparameter optimization, the proposed framework preserves relevant macroeconomic information while reducing redundancy. The results demonstrate its potential for government bond yield forecasting and for supporting investment analysis, portfolio management, and monetary policy assessment in emerging financial markets.