Macroeconomic driven cement price forecasting in Sri Lanka using econometric and deep learning models

Abstract

Cement price stability is fundamental to the performance of the construction sector, which significantly contributes to national economic development in Sri Lanka. However, recent macroeconomic instability characterized by exchange rate depreciation, inflationary pressures, and fuel price volatility has intensified cement price fluctuations, particularly during the 2022 economic crisis. Despite the growing need for reliable forecasting tools, limited research has evaluated the combined application of econometric and deep learning models for cement price prediction in highly volatile emerging economies. This study develops and compares forecasting models using a 20-year monthly macroeconomic dataset (2006 to 2025). An Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX) model was established as a linear baseline, incorporating key predictors including Inflation, Exchange Rate, Diesel Price, and Global Coal Prices identified through Pearson correlation analysis. Advanced deep learning architectures, namely Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Attention-based LSTM (ATT-LSTM), were implemented to capture non-linear temporal dependencies. Model performance was evaluated using Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²). Results indicate that while the GRU model achieved the highest accuracy under optimized sequence-length configurations, the LSTM demonstrated greater stability and robustness during crisis-period testing. Overall, deep learning architectures outperformed the ARIMAX baseline in capturing cement price volatility during the Sri Lankan economic crisis. These findings can support contractors, quantity surveyors, and policymakers in improving procurement planning and managing construction cost risks during periods of economic instability.