This study presents a comparative analysis of time series forecasting (TSF) models for retail demand prediction using a synthetic dataset. The data was designed to reflect real-world dynamics such as seasonality, trends, promotions, and price elasticity. The objective is to evaluate the performance of statistical, machine learning (ML), and deep learning (DL) models across short-, medium-, and long-term forecast horizons. Four models were tested: Naive Forecaster, ARIMA, XGBoost, and LTSFLinearForecaster. Using expanding window backtesting, model accuracy was assessed with standard metrics: RMSE, MAE, and MAPE. Results show that forecasting performance varies notably by product category and horizon. Classical models perform well in stable conditions but degrade with complexity. XGBoost improves accuracy but shows high variability. LTSFLinearForecaster delivers a more stable performance. But no single model dominates across all conditions.