<p>Accurate forecasting of residential electricity consumption (REC) is essential for effective energy planning, particularly in countries with growing demand and limited infrastructure. This study compares the performance of classical statistical and machine learning models for REC forecasting in Paraguay. Specifically, we evaluate the Seasonal Autoregressive Integrated Moving Average (SARIMA) model alongside three machine learning approaches: Extreme Gradient Boosting (XGBOOST), Neural Basis Expansion Analysis for Time Series (NBEATS), and Random Forest (RF). Forecasts were produced for short-term (12 months) and medium-term (24 months) horizons. Unlike most studies that assess performance using a single evaluation period, we compute error metrics (MAPE, MAE, RMSE) across multiple validation windows to examine the temporal stability of each model. RF consistently achieved the highest accuracy in short-term forecasts, benefiting from its ability to capture nonlinear patterns without relying on strict statistical assumptions. In the medium term, although all models showed a reduction in accuracy, RF maintained competitive performance and temporal consistency, supporting its use for extended horizons. While SARIMA occasionally performed competitively, its residuals often violated key assumptions such as normality and independence, particularly in non-stationary contexts. This work represents the first empirical application of these methods to Paraguay’s residential electricity data and introduces a replicable framework for model evaluation in settings with limited exogenous data, as commonly encountered in developing countries. Although centered on Paraguay, the findings are applicable to other countries facing similar challenges in energy forecasting and planning.</p>