Business requirement was to design and develop a solution to predict the optimum quantity of perishable food items for the catering application of an airline.
Machine Learning and Statistical Computing was used to derive the intelligence and accuracy in the predictions.
Optimization Techniques were used in conjunction with the model predictions to obtain a trade-off between stock-out and wastage percentages
On an average, around 25% reduction was achieved in the overall item level wastages, across all legs in which the application operates
Technology: Mcube product (leveraging R and Python libraries), integration with the catering system of the airlines
- An ensemble modeling approach was used for forecasting demand (at an item-sequence level).
- Minimum Absolute Percent Error was used as the model selection criteria (for each item-sequence combination). The validation period, against which the efficacy of the model was tested was 6 months. The model was built using three years of cart level transactional data
- Factors like day of week, time of day, domestic/international flight, etc. were also incorporated into the model
- The results of the forecasts were further fed into an optimization module such that the wastage (at an item level) was restricted to 25% (or less), and there was no undersupply for more than 5% of the total food items predicted across all operating sectors of the airline.
- The APIs deriving the predictive recommendations were used for consumption and integration with the catering application
- The APIs were delivered with authentication enablement for security reasons, and the response was in JSON format