Time Series Forecasting with LightGBM under Data Scarcity: An Application to Romania’s Inland Gas Consumption

Developing forecasting models capable of learning from small datasets is increasingly valuable for scenarios with limited computational resources and tight time constraints. In this case study, we processed monthly data on Romania’s inland gas consumption—a primary benchmark reflecting the country’s industrial growth, level of technological advancement, and reliance on non-renewable energy sources. This research tests the extent to which LightGBM, a gradient boosting framework, can predict seasonal patterns in monthly gas consumption. To aid the machine learning framework in better understanding the series pattern, we applied a first-order differencing to the data. By combining hyperparameter tuning, cross-validation, and tailored feature engineering (including lagged variables and rolling-window statistics), the analysis thoroughly evaluates LightGBM’s performance under data-scarce conditions. Model accuracy was assessed using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE), demonstrating the extent of LightGBM’s predictive capacity across multiple horizons despite constrained data settings. These findings offer insights into the feasibility of employing fast-adapting, lightweight machine learning techniques for reduced time series datasets, while minimizing both computational effort and processing time.