Machine Learning & RPA For Retail
Increasing Accuracy in Inventory Forecasting
Saving money and increasing product availability
Currently many retailers do not have an effective inventory forecasting system that is able to predict future minimum inventory levels based on information such as current sales velocity, anticipated sales and historical sales data. Today, Inventory forecasting systems use present stock levels and generate re-order requests once the levels are low. Advoqt has created an accurate forecasting system that predicts future sales down to the SKU and specific day of the year.
Advoqt Technology Group drives Digital Transformation for mid-sized businesses. Seasoned consultants, proprietary Machine Learning and Robotic Process Automation technology, and a vetted ecosystem of niche partners allow us to deliver integrated solutions that give you a measurable competitive advantage.
As part of discovery, Advoqt analyzed historical datasets and identified three distinct categories of product/store combinations: A) those that were not suitable for prediction due to having too few data points, B) inconsistent data that required machine learning prediction with additional data, and C) consistent data that predictive analytics could be applied to. Advoqt and the client chose to focus on the consistent data eligible for predictive analytics. Advoqt had identified the problem domain and behavior to have predictable historical performance including holiday sales spikes that could be modeled using the Facebook Prophet time series prediction engine. The Prophet engine was created by Facebook’s data science team specifically for solving time series problems automatically and has the highest accuracy performance amongst automatic prediction engines.
Our team identified that some of the historical data was susceptible to short-term changes that needed to be modeled using a prediction algorithm. To solve for this scenario, a double sliding-window algorithm was added to form a prediction ensemble with the Prophet engine. This increased the accuracy response of the forecasting system in case of random short-term sales spikes that may be caused by low-frequency events such as extreme weather, traffic or sporting events.
After the prediction model was built, Advoqt created a distributed rolling forecast system to contain it in production. The client required the forecasting system to create daily forecasts for the following 2 weeks, with a prediction time interval of no longer than 2 hours. To meet the execution speed criteria, Advoqt used a distributed computing technique to split the historical data into smaller datasets. These datasets were processed in parallel by several instances of the prediction model instead of processing all historical data as a single large dataset with a single prediction model. This met the execution performance criteria of predicting 2 weeks of sales for 1000 store/product combinations using the past 365 days of sales, delivery and inventory data.
A key facet of the solution to the client’s inventory forecasting problem was the ability to automatically add products and stores to the prediction model. A fully automatic solution was necessary to meet the client’s requirements of robotic process automation. Based on the knowledge of the data consistency and target accuracy, Advoqt created an automatic onboarding engine that cleaned the historical data of new product/store combinations and determined if they were eligible for prediction within the client’s accuracy parameters. Once that was accomplished, the client only needed to continue adding the daily sales data for the new product to the database and no further onboarding action was required.
To learn more about how Advoqt can help you implement ML in your organization please visit our website at www.Advoqt.com or call (617) 600-8161.
Explore Other Resources from Advoqt Technology Group
The blog created by Rapid7 concerns security orchestration and automation tools. The blog gives a definition of what these tools do. A SOAR implementation would begin with defining and understanding the security issues being faced by the organization and thinking...read more
Our fifth article is from Buyer’s Guide and is an article by Karen Scarfone of Scarfone Cybersecurity. This article gets into specifics concerning the top security information and event management (SIEM) systems. The tools collect security log data from many sources,...read more
Our fourth article is a blog by Rostam Dinyari, a strategic cloud engineer, and concerns how an organization needs to gather and prepare data for machine learning deployment. A list of guidelines is presented. The first phase in data collection is to define the types...read more