The role of feature engineering in multivariate time series forecasting: Evidence from housing and mortgage rate data

Abstract

This study investigates the effectiveness of feature engineering in enhancing the accuracy of multivariate time series forecasting, addressing limitations of conventional statistical and machine learning models. Two U.S. housing-related datasets were analysed using regression, machine learning, and deep learning approaches. Feature engineering techniques, including lag, rolling window, expanding window, and seasonality, were applied, and feature importance was validated using Random Forest, XGBoost, and Decision Tree classifiers. Forecasting was performed with regression models, VAR, and LSTM, while Root Mean Square Error (RMSE) was employed to validate and compare model performance. Results demonstrate that feature engineering significantly improves forecasting accuracy across all models, with LSTM consistently outperforming VAR, Random Forest, Decision Tree, and XGBoost regressors. Accuracy was further enhanced by optimising the number of engineered features included in forecasting. The findings highlight that effective feature engineering, combined with rigorous model validation, can achieve high forecasting accuracy without relying on overly complex models, making this approach valuable for practical applications such as housing market prediction and financial time series analysis. Unlike prior studies focused mainly on model sophistication, this work emphasises the role of feature engineering and feature selection in improving multivariate time series forecasting and provides novel insights into balancing model complexity with feature optimisation.