Forecasting residential electricity demand a single hour ahead is a well-defined but persistently difficult problem, since household load fluctuates with occupant behavior in ways that simple statistical models struggle to capture. This paper compares three regression approaches for short-term residential load forecasting (STRF) - Linear Regression, Random Forest Regression, and Extreme Gradient Boosting (XGBoost) - under a single, controlled experimental pipeline. Data come from the publicly available Individual Household Electric Power Consumption dataset [1], originally logged at one-minute resolution between 2006 and 2010; after cleaning and aggregating to hourly resolution, the working dataset contains roughly 34,584 hourly observations, and the forecasting target is set one hour ahead. Alongside calendar-based predictors (hour of day, day of week, a weekend flag), two lagged variables, Lag_1 and Lag_24, were engineered to encode short-term momentum and daily periodicity in consumption. The dataset was split chronologically, 75% for training and 25% for testing, so that no future information leaks into the training process. Accuracy was assessed with three complementary metrics: MAE, MSE, and R². Across all three, XGBoost is the clear winner: relative to the Linear Regression baseline, it cuts MAE by 50.1% and MSE by 68.6%, and lifts R² from 0.5272 to 0.8518 - a 61.6% relative gain. These results indicate that boosting-based ensembles are considerably better suited than linear or purely bagging-based models at capturing the nonlinear, time-dependent structure of household electricity use, which is directly relevant for applied forecasting in smart energy systems.