Building a Forecasting Model to Predict Donuts Sales to Reduce Food Waste
With increasing urbanization and trend towards fast food and take-aways, outlets like Dunkin Donuts are all set to thrive within the Pakistani food industry. Backed by global presence and considered to be the pioneer of fresh and creamy donuts within Pakistan.
But while globally Dunkin Donuts deploys AI on multiple levels, Dunkin Donuts Pakistan had some challenges at hand, the biggest one being inventory management and waste reduction. High gross margins and low production costs often lead companies to view overproduction as a solution since the gross margins exceed the throwaway losses. Even though they were taking several steps to assess demand forecast as accurately as possible, there was still a high percentage of stock being thrown away because of low shelf life. They did have a very streamlined supply chain with orders being placed by area managers the night before, and shipment from production centres to stores occurring twice a day, allowing greater flexibility of stocks. Store managers would also coordinate with each other every evening to transfer additional stock amongst the branches. Despite all this, they had a 20% throwaway wastage, which was our task to reduce.
Challenge
The key test was not just to reduce the waste percentage for the client, but to do so under some challenging circumstances. Firstly, the client did not have a centralized system for recording sales. Instead, a rider was sent to each store to collect the physical receipts which were then entered into spreadsheets manually, leading to data delays of up to a week, and our algorithm had to account for these delays.
Secondly, the algorithm was expected to compete against human intuition and communication (that of the area managers), who could make immediate adjustments if they could foresee a sudden shortfall or spike approaching, such as asking for extra production or coordinating with other managers. The algorithm, however, had to make predictions without such additional information.
Solution
The algorithm is the product of a series of data cleaning, feature engineering and advanced statistical modelling techniques. The model used in this case is an ensemble model which relies on 3 or more models at the base level and a custom-built optimizer at the top level. The custom-built optimizer works to ensure that all business constraints are taken care of, and the predictions returned are the ones that best meet those constraints. LFD has, hence, taken the focus away from model KPIs like accuracy and instead shifted it to actual business constraints, problems and profitability.
LFD tackled this with advanced forecasting and modelling techniques, relying on not just the client's data but also open-source data such as weather, while accounting for the vast difference between production costs and selling price to ensure that underpredictions were avoided as far as possible. Compared to the client's previous one-size-fits-all model of forecasting based on averages and adjusted for outliers, our model generated multiple models for each store, and combined model predictions by applying weights to individual models, thus improving accuracy by a huge margin.
Result
LFD has a strong commitment to reducing unnecessary expenses for clients. As such, all algorithms are made on open-source software such as R Programming while all dashboards and portals are built on opensource platform. The case for this algorithm is no different as it is completely open source with LFD only charging for its own added intellectual property.
Despite having no access to discretionary human decisions and delayed reporting, the algorithm still managed to outperform the client's system, and reduce wastage by 25%. With faster data transfer and more detailed historical reporting, there is no doubt that the wastage reduction percentage can be even higher through our model.
A new phase of forecasting
Because this forecasting model considers multiple internal as well as external factors, it is currently the most advanced and adaptive demand forecasting model within the Pakistani industry. Built up with multiple variables and working on real time data plus incorporating for any data lag, our model has futuristic implications for all businesses that deal inventory management, supply chain and demand forecasting.
Our algorithms are a key turning point for the way 'intelligent forecasting' and 'educated guesses' are made within various industries in Pakistan, and the business value generated from this, along with social implications for the government and social sector (such as vaccine forecasting, predictive modelling of schools, to name a few) can be manifold.
