Enabling a Large FMCG company with demand forecasting models

Retail

Enabling a Large FMCG company with demand forecasting models

05

Engineers

06

Months

81%

Accuracy with advanced models

Overview

A large FMCG company was looking for a platform that could help them forecast demand for their online e-commerce operations on Amazon, Target and Walmart for the US market. Demand forecasting is a key initiative in retail which helps category managers & merchandisers alike to understand customer demand patterns and accordingly stock inventory and make better marketing decisions. The current demand forecasting provided by Amazon did not meet the requirements of their team. Hence by fully utilizing the data and the domain expertise present at the client team, a more robust and accurate demand forecasting model could be arrived at.

Services

Data Engineering, Machine Learning, Visualization

Technology

Python, Apache Kafka, Apache Spark, PostgreSQL, React, Tableau, Docker, AWS

Location

New York, USA

Challenges

They were looking for a platform that could help them forecast demand for their online e-commerce operations on Amazon, Target and Walmart for the US market. Demand forecasting is a key initiative in retail which helps category managers & merchandisers alike to understand customer demand patterns and accordingly stock inventory and make better marketing decisions. The current demand forecasting provided by Amazon did not meet the requirements of their team. Hence by fully utilizing the data and the domain expertise present at the client team, a more robust and accurate demand forecasting model could be arrived at.

Only 60% Accuracy on basic models.
High variability in product sales

Solutions

Data Ingestion: The custom-developed platform was deployed within the client’s infrastructure. Data extracts from their SQL server data lake were filtered in PowerBI and the output excel files were directly uploaded to the developed platform. The platform automatically parsed the data and structured the data into a time series by aggregating the data into weekly and monthly series. The data was also partitioned by category, brand and SKU to generate individual time series for the models. This data was then pre-processed and cleaned for any inconsistencies and outliers and the time series statistics like Stationarity, Seasonality and trends were captured..
ML Models: The univariate demand forecasting models were first trained using ARIMA, S-ARIMA, Prophet and LSTM. These were fine-tuned to provide the required accuracy. Then multivariate models like Random Forest, VAR and Deep Learning were trained to achieve the desired accuracy as well as provide correlation between demand and market spend as well as Allocation %. The Output data was stored on the platform and made available via a single API.
Dashboarding: The Lightweight Application provided the final interface for the category managers to interact with the model data. This made use of the data access API layer and some of our pre-built components for model parameter control. With the Lightweight application deployed, client’s Team could search for a category, brand or SKU and then define a forecast period for which they would then be able to view a demand forecast in sales and units. This would also have upper and lower bounds for them to consider their planning activities A what-if scenario component was also provided in the app where the team could input a marketing send and view the affect on demand and vice versa. Similarly, it was also enabled for allocation % to check for out-of- stock events.

Modules implemented

Data Ingestion

This module was responsible for extracting data from different sources like SQL server and excel files.
It was then partitioned by category, brand and SKU to generate individual time series for the models.

ML Models

We used the models ARIMA, S-ARIMA, Prophet and LSTM. These were fine-tuned to provide the required accuracy.

Dashboarding

Dashboarding was developed using a light weight frontend framework with integration to data access layer. It provided the client with reporting and graphical insights

High Level Design Architecture

Need to estimate for your next project?

We at TechVariable do acknowledge that one size will not fit all. Hence, we work in collaboration with you to identify, analyze & then develop a solution that fulfills your needs. Either we will define the functional scope of your project to estimate the timeline and budget or you can create your own agile team from among our resources.

The Result

The custom-developed platform was able to achieve a significant improvement in demand forecasting accuracy. The univariate models were able to achieve an accuracy of 80%, while the multivariate models were able to achieve an accuracy of 90%. This improvement in accuracy has enabled the client to make better planning decisions and reduce inventory costs. The lightweight application has also been well-received by the client team and has been used to make better marketing decisions.

Previous slide

Next slide

See how our solutions are making a difference across industry

A-LOCK: Revolutionizing Data Privacy with a Unique Cloud Storage Solution

Our clients approached us with the vision of creating a unique cloud storage solution to provide enterprise-level privacy protection to daily users. They sought to develop a platform that would not store copies of users’ files, photos, videos, or memories, eliminating concerns over data breaches and unauthorized access.

70%

Increase in data privacy

04 Engineers

04 Months

An Innovative Augmented Reality and AI Solution for Online Clothing Stores

A Canada-based client was looking for a solution that could improve online sales for participating sellers, and in turn, make their marketplace platform more lucrative. TechVariable brought its raw idea to life by imagining and designing software that creates a user’s avatar based on information provided by the users themselves. Thus, the team took their first Augmented Reality project and delivered it in a record time of six months.

65%

Increase in Sales Efficiency

04 Engineers

06 Months

Powerful Insurance Claims Rejection Predictive Model

Health insurance plans offer comprehensive security when it comes to extensive medical-related expenditures. Yet, when filing an insurance claim for any service, various reasons may lead to a rejection of the claim, ushering in a financial burden on the client and a loss of monetary compensation for the service provider.