The single largest distributor of mobile and lifestyle products in the Middle East, providing end-to-end supply chain solutions for all categories of Information Technology products (PCs, PC building blocks, networking, software and enterprise solution products) and Consumer and Lifestyle products (Telecom, Digital Lifestyle products, Entertainment products and Digital Printing Machines) to over 100 international brands and relationship with major brands like Apple, Samsung, Dell, HP etc. to name a few. The distributor is amongst the leading supply chain solution providers in Middle East, Africa, Turkey and CIS region for leading manufacturers of Information Technology, Telecom and Lifestyle products.
The customer is distributing more than 500 SKU’s of mobile phones across multiple regions like UAE, Qatar, Saudi Arabia etc. Currently team is predicting the inventory of each SKU for a certain period based on their experience and instinct. It was difficult for them to predict the inventory every day because of multiple factors needs to be considered while predicting the inventory, also the volume of SKU is huge which made it a routine and tedious process to get the predictions for each unit. Also, the customer intended to bring to the forefront use cases like aging inventory, thereby minimizing the inventory cost.
The customer required a Forecasting solution which predicts the Inventory of all the SKU’s based on historical data, across all the countries, enhancing data driven decision making processes, empowering the organization to be future ready.
Typically for solving any Machine Learning problem we need domain expertise, Data and Technology. Citrus had multiple sessions with customer’s domain experts to understand the use case and the availability of the data. After the thorough understanding of the use case and data Citrus proposed Amazon DeepAR plus algorithm to create the Inventory Prediction Model which will fulfill all the customer requirements.
In order to build the model, created a Data Pipeline to get the Sales and Inventory historical as well as live data from customer SAP Hana database to Amazon S3.
Citrus’s team of advanced data scientists, data engineers and business analytics aided the customer with data preprocessing, training and deployment and rest API creation.
The team, performed Exploratory data analysis and visualization of the dataset in order to understand its properties and prepared the data for further pre-processing. After pre-processing, selected specific products according to quantity and quality of historical data, the team eliminated duplicates sales transactions and did outlier analysis in order to identify historical anomalies and deep dive into the business reasons and implications of the same, so that the suggested issues could be implemented in line for future predictions of the model.
The team post data analysis, Input the missing day sales volume with zero as there was no sales transaction happened during that day. Therein, the team built the models based on different versions of the pre-processed data to predict the sales demand of any SKU. Citrus undertook experiments with different algorithms and services like AWS DeepAR plus, Aws Forecast etc. and respective hyper-parameter combinations. Selected the best Model and their hyper-parameters and did further research in optimizing their performance.
The best model was therein deployed after testing and created an endpoint for the further usage. Proficient in Server less technologies, the team created a Lambda function which will give inventory predictions for any SKU based on the predicted demand. The Predictions include inventory upper bound, lower bound and actual predicted value. The Endpoint is exposed for input requests through Lambda function. Created a Rest API using API Gateway and attached the prediction lambda to API for the inference request as well.
For Implementing end to end Inventory Prediction solution for Redington it took 3 Months starting from March 2020. Divided the project into three phases and in each phase delivered the above requirements. By end of phase 1 we selected the SKU’s which has good quality and quantity of data and identified the key features that are responsible for building the model. In second phase prepared the data as required by the model and tried with different forecasting models finally selected DeepAR plus algorithm which worked best. By the end of phase 3 created the model and endpoint which provides real time predictions and automated end to end model training process.