Food distributors struggle to forecast demand
Although consumers are far more aware of the end of the global food chain—the grocery store and the dinner table—food distributors are a crucial part of every country’s chain of supply. In an economy that faces constant change in the form of labor shortages, seasonal shifts, proliferation of products, highly perishable items, and even sudden catastrophes like the current pandemic, food distributors need to accurately predict demand so that they can provide adequate, optimized supplies of fresh food to stores.
Current demand forecasting tools rely on pre-built statistical models that apply general rules to a retailer’s transaction data. This approach has three major shortcomings:
They are myopic in the sense that by considering just historical transaction data, these tools miss out on major demand drivers like consumer sentiment, competitor strategy, demographics, and other major world events like severe weather conditions or a global pandemic like the one we are facing today.
The statistical forecasting models are not customized to each client and are too generic. When a deviation of just 1% could cost a retailer millions, precision of the forecasting models matters.
And this is critical: unless products are broken down into their constituent attributes and forecasting is done at an attribute level, a forecasting tool can never know why a size 11 Nike Air Jordan sells 45% more than a size 9 or why strawberry flavored Kellogg’s Raisin Bran doesn’t perform as well as the original.
About the client
This top 3 Japanese food distributor is known as a wholesale provider of frozen items, alcohol, chilled foods, and bakery items, as well as shelf-stable grocery products. With more than 4,000 employees, they are an important part of the Japanese food economy. The Japanese Food Distributor was clear on what they wanted. They had a current logic for deciding on assortments and shelf placements for each retail store they distribute to. They needed someone that can first automate that process and then make it even better through AI. CrowdANALYTIX could help with both.
Our approach
We first started by building models to extract key product attributes like ingredients, nutrition as well as tags like Gluten-free or not, Organic or not, etc. Once the auto-tagging algorithms were built and about 28 product attributes extracted, we started working on the assortment recommendations and shelf planning algorithms.
The first phase of assortment recommendations and shelf planning algorithms were basically a rules-based engine to automate the current manual approach that the client was already using. The output of the model also gave us baseline data and feedback that could be used to train and tune a more robust and self-learning model in the second phase.
Phase two models would be built by comparing multiple machine learning and statistical approaches to identify the models that give us the highest level of precision and give us a lift of at least 5-10% on the revenue and margins that are achieved through the version of the models in Phase 1.
Impact
Prior to implementation of our solution, the client was spending approximately 100,000 hours every year on determining assortments and shelf plans for each retail store they serviced. The savings from this alone were more than 50%. In addition, in early tests, we have seen an increase in revenues of close to 2.5% and are confident that this will be sustained or increased over time.
