Synthetic Retail Time-Series for Privacy-Preserving Demand Forecasting: An End-to-End Pipeline, Benchmark Evaluation, and Privacy Analysis

Access to high-quality transactional data is a persistent bottleneck in retail demand forecasting due to commercial confidentiality and privacy regulations such as GDPR. Synthetic data offers a promising alternative, yet its utility for training complex machine learning pipelines remains under-explored, and no standardized validation protocol exists for synthetic time-series in forecasting contexts. In this paper, we propose and evaluate a complete, privacy-preserving retail demand forecasting pipeline driven by a controllable synthetic data generator. Our generator replicates core retail dynamics including daily and weekly seasonality, SKU-specific trends, promotional uplifts, holiday effects, and exogenous foot traffic, modelled via Poisson processes and sinusoidal components. We validate dataset fidelity using the Fidelity-Utility-Privacy (FUP) framework through autocorrelation analysis, distributional diagnostics, STL decomposition, and uplift verification. A feature-rich forecasting pipeline is constructed using leakage-safe lag features, rolling windows, and calendar encodings, and three model families are benchmarked using time-ordered train-test splits and rolling-origin cross-validation: Linear Regression (baseline), Random Forest, and LightGBM/XGBoost (gradient boosting). Results demonstrate that tree-based ensembles substantially outperform linear baselines, with Random Forest achieving MAE of 6.20, RMSE of 8.15, and MAPE of 11.7% on short-horizon forecasts. The pipeline is deployed as an interactive Streamlit decision-support dashboard supporting SKU selection, horizon adjustment, and promotion/holiday scenario testing. This work addresses three persistent gaps in the literature: the integration gap, the synthetic validation gap, and the real-time deployment gap. We further discuss the privacy properties of parametric synthetic generation and motivate future extensions toward formal differential privacy guarantees, energy/smart-grid domain transfer, and HPC-scale evaluation using national supercomputing facilities. An interactive Streamlit dashboard for this research is available at: https://retail-forecasting-hvdzvesi4u9l6fs5tvdoyi.streamlit.app/