Perishable food supply chains are exposed to pronounced demand volatility driven by promotions, weather, andfestivaleffects;consequently, traditional statistical forecasters often yield biased or lagged signals that propagate into stockouts, overstocking,andavoidable waste. This work presents an integrated decision-support framework that couples multi-horizon, exogenous-awaretimeseriesforecasting with inventory optimization tailored to perishable goods. The forecasting layer benchmarks classical models against machinelearning (gradient-boosted trees, ensembles) and deep learning architectures (LSTM/GRU, Temporal Fusion Transformer, PatchTST),explicitly incorporating external covariates to capture non-linear and non-stationary demand regimes. Forecast distributions thenfeedanoptimization layer that applies the Economic Order Quantity model for relatively stable items and the Newsvendor formulation, aswellaslinear/mixed-integer programs, for short-life products across SKU–store hierarchies. The system is evaluated using statistical accuracymetrics (MAPE, RMSE, Bias) and operational key performance indicators (service level, fill rate, holding cost, and wastagepercentage).An interactive dashboard operationalizes these components, enabling scenario analysis and proactive alerts for stockout or overstockrisk.By jointly improving forecast fidelity and translating predictions into implementable replenishment rules, the frameworktargetsmeasurable reductions in waste and cost while sustaining customer service levels in real-world retail contexts. Keywords: Time Series Forecasting; Perishable Inventory; Temporal Fusion Transformer; LSTM/GRU; Exogenous Regressors;Newsvendor Model; Linear Programming; Decision-Support