The Bangladesh garment industry is highly vulnerable to short-term disruptions, causing frequent backorders and demand–supply imbalances. This study evaluates and compares multiple ML algorithms to identify the most effective models for backorder prediction and risk-aware demand forecasting, with the aim of reducing backorder frequency and maintaining balanced inventory levels. An initial industry survey established the relevance of the problem, followed by the collection of a real-world dataset from Bangladeshi garment manufacturers. For backorder prediction, XG Boost, ANN, Random Forest, MLP, Cat Boost, Ada Boost, and Light GBM were trained and tested, with performance measured using accuracy, precision, MAE, MAPE, and RMSE. Products were then categorized into high, medium, and low-risk groups based on backorder frequency and demand forecasting was carried out for high and medium-risk categories using XG Boost, Random Forest, and Light GBM. An additional inventory status indicator was calculated to measure the gap between forecasted and actual inventory sufficiency across models and scenarios. Sensitivity analysis with ±10% and ±20% demand variations was performed to assess model robustness under uncertainty. Results show that MLP is the most suitable for backorder classification, while XG Boost and Random Forest perform strongly under stable demand conditions. However, under volatile demand, Light GBM outperformed other models, reducing backorder rates and providing better inventory coverage. Risk-based product segmentation enabled focused forecasting strategies, and sensitivity analysis highlighted model behavior under fluctuating demand. These findings provide actionable insights for manufacturers to select appropriate ML models for minimizing backorders, maintaining inventory balance, and improving supply chain resilience in disruption-prone environments.