The Fifth Elephant 2023 Monsoon

Link to presentation: https://docs.google.com/presentation/d/1ZaA3TdTqBHurUJV7ngEhxVvZOkTM5D0vMI21qdI48kI/edit#slide=id.p1
Problem statement:
Instamart, the quick commerce grocery delivery service of Swiggy gives unparalleled convenience of being able to order, from a huge assortment, across fresh fruits & vegetables/ dairy/FMCG products and accessories for household requirements, parties or festivities, pretty much at any time of the day and also through late night (from 6am to 3pm) and get the delivery in ~10–15 min. Instamart follows a dark store model where micro-fulfilment centers are established to fulfill the grocery orders of a certain geographical area of a few kilometers of radius. Efficient demand planning ensures that the sufficient units of each of the products are ‘available’ in the closest pod (dark store) for customers to order throughout the day, while making sure not stocking up too many units which can eventually lead to ‘wastage’. For efficient planning, ML based forecasting techniques are used to predict the daily ‘demand’ of an item for a given store (referred as SKU). But the demand forecasting for Instamart, or instant grocery delivery systems in general, have a handful of challenges that traditional forecasting methodologies can not resolve.

Firstly, due to the hyper-local nature of demand planning, there is high variation of demand across geographies, items and days – which leads to frequent ‘out-of-stock’ for some of the SKUs even before the pod is closed for the day. Hence, the historically observed time series data for building forecasting models is not the accurate representation of the ‘true’ demand, rather it is a truncated demand. The frequent absence of true demand makes the model development and evaluation challenging, especially when we are dealing with number of SKUs in the order of 10^4.

Business implication:
To track the efficiency of the demand planning, the business team tracks two metrics primarily: 1) availability – it measures the proportion of the day a SKU was available for the users to order, and 2) wastage – it approximately quantifies the units over-stocked and eventually led to wastage. Not being able to accurately evaluate the model performance using traditional metrics such as wMAPE on back testing data can lead to deploying forecasting models in production that can either underpredict or overpredict the true demand which means lower availability (i.e., revenue opportunity loss) or higher wastage respectively.

Solution:
In this presentation we will go over our approach of ‘Adaptive Metric Alignment’ for accurate model evaluation which is more closely aligned with the business metrics. We will cover the following topics in our presentation:

1. Introduction to demand planning process for Swiggy Instamart
2. Demand forecasting challenges for quick commerce grocery delivery services
3. Drawbacks of traditional model evaluation methods
4. Adaptive metric alignment: estimated availability and wastage
   Production implication and conclusion