This study explores diffusion-based generative modeling as a data augmentation strategy to improve forecasting in data-scarce scenarios. Using the ETTh1 multivariate energy dataset, we evaluate point and quantile forecasting across multiple forecasting architectures (XGBoost, LSTM, BiLSTM). Synthetic samples are generated via the Diffusion-TS framework for the neural models only, and incorporated at varying synthetic-to-real ratios. Results show that BiLSTM models benefit substantially in point forecasting, achieving up to a 15.3% improvement (i.e., reduction) in RMSE and 8.1% in MAE, and similarly improve quantile accuracy, while LSTM models degrade under all ratios. Bias-variance analysis reveals that diffusion-based augmentation mainly reduces variance at moderate levels but introduces bias when excessive synthetic data are used.