Integrating Machine Learning and Scenario Planning for Enhanced Demand Forecasting in Bangladesh’s Garment Industry

Abstract

The garment industry in Bangladesh is highly unstable and volatile in terms of exporting due to the cyclic nature of demand based on seasonal order, concentration of buyers, and disruption in the supply chain. The traditional forecasting models like the ARIMA and the Holt-Winters are very popular due to their interpretability and comparative ease of use though they cannot capture non-linear and non-regular order behaviour. Simultaneously, machine learning models have been brought up as a helpful instrument in enhancing forecast accuracy, and scenario planning has been identified as a means of aiding in decision making when there is uncertainty. This thesis studies how machine learning can be integrated with scenario planning to forecast the demand in the garment industry in Bangladesh. Based on this, the research question of the study is as follows: How can machine learning models, combined with scenario planning, improve the accuracy and adaptability of demand forecasting in Bangladesh’s garment industry compared to traditional forecasting methods?

The theoretical basis of the thesis is based on forecasting theory, Systems Theory, the Dynamic Capabilities View, and literature on supply-chain resilience. The literature discusses that conventional statistical models can be applied to predict stable and seasonal demand trends but are less effective in situations where demand is influenced by sudden changes, uncertainty on the buyer side and structural volatility. Machine learning models, especially tree-based ensemble models, like Random Forest and XGBoost, are more flexible in the modelling of non-linear demand behaviour. The accuracy of forecasts is, however, not enough to make operational decisions in uncertain supply-chain conditions. Scenario planning is therefore a complementary decision-support layer, which transforms results of forecasts into scenarios of the stress test in the form of demand declines, demand peaks and buyer cancellation exposure.

Quantitative case-study approach is applied under the internal company-order data of a garment exporter in Bangladesh. The forecasting target is the monthly order quantity, and the traditional statistical models are compared with the Random Forest and XGBoost using a chronological holdout test period. The lagged order quantities, rolling demand momentum, and calendar-based indicators are used to form machine learning models. The best model’s baseline forecast is then subjected to scenario planning via deterministic stress tests of a demand drop, a demand spike, and the cancellation of the largest buyer order exposure. The results suggest that machine learning can be used together with scenario planning to enhance accuracy of forecasting and flexibility in management. The paper concludes that a unified forecasting-scenario planning framework may aid in production planning, capacity calculation, buyer-risk evaluation, and information-intensive production planning in the garment sector of Bangladesh