Retail And Consumer Goods Use Case - .microsoft

Azure Industries ExperiencesRetail and consumer goodsuse case:Inventory optimization through SKUassortment machine learningLorem ipsumMariya Zorotovich

Contents01Introduction02Solution overview03Accelerating SKU optimization with Neal Analytics04Next steps 2018 Microsoft Corporation. All rights reserved.This document is provided “as is.” Information and viewsexpressed in this document, including URL and other internetwebsite references, may change without notice. You bear therisk of using it.This document does not provide you with any legal rights toany intellectual property in any Microsoft product. You maycopy and use this document for your internal, referencepurposes.

01Intro A mother with two children in tow (a consumer) walksthrough a favorite market and scans a shelf for aparticular brand of cereal. But she finds that theparticular item (SKU) is out. Rather than wait or visitanother store, she settles for a competitor’soffering—which is a loss for the brand. This is anout-of-stock scenario, and consumer goods brandswant to avoid it because of the loss. One mitigationfor this is the idea of SKU assortment optimization.That is, using advanced analytics to more closelymonitor, predict, and provide a course of action foran out-of-stock scenario.The focus on the omnichannel experience andaccelerated last-mile fulfillment has cast a light on thevalue of a well-executed inventory plan to ensureproduct is available and in stock to purchase. Thereare many facets to the art and science of merchandising and inventory management that span fromproduct development, sourcing through buyplanning, assortment, allocation and movement.Product availability is paramount to any brand’ssuccess, specifically having product stocked onshelves with the right combination of complimentaryproducts. This maximizes sales, delivers a greatservice experience and drives brand loyalty.How being out-of-stock affects businessGiven the amount of product choices available to aconsumer, there is nothing more frustrating when youare looking for a particular brand of product and it isout-of-stock, and unavailable when you need it most.According to EKR, when faced with an out-of-stockscenario, only 15 percent of shoppers will delaypurchase; there is more likelihood of shifting to adifferent brand. You lose the sale in that moment, andwhat’s worse is that if your product is consistentlyout-of-stock or difficult to access, you open the doorfor your competitor to win over your customer withtheir product.

01[continued] Barriers to solutionsOne might think keeping shelves stocked with availableproduct might be an easy task to undertake; however,this is not the case given the scale and often number ofSKUs to be managed. Here are some key reasons stockouts are experienced: Imprecise demand forecast accuracy: Planassortments for sales channels need accuratedemand forecasts. Methodologies tend to beleveraging only historical data and traditional dataoutput with manual manipulation and littlevisualizations to drive insight and predictiveindicators. Today’s demand forecast accuracystands at 60 percent .1 Inaccurate data management: There are threecategories that account for inaccurate data relatedto stock outs: retailers and brands not systematically capturing data related to stock outs, information that is tracked but not shared across theenterprise, and systems are siloed and unable toshare data systematically.Vague perpetual inventory (PI) data: PI dataleverages POS estimation methods to determinewhen an item is out-of-stock. The accuracy rate ofthis data is below 50 percent and providesguidance at a store/outlet level, not at a shelf level.Inappropriate shelf-space allocation: Mostcompanies do not use unstructured location-baseddata sources to identify products which are proneto be out-of-stock. Increasing flexible fulfillmentoptions and shelf-pace for fast moving items addspressure on inventory accuracy.Low planogram/floor set compliance: Highplanogram compliance for fast-selling categories ismuch needed for optimum retail execution andsell-through. But for low categories, there is a highchance of stock outs given the lack of visibility andbreadth of portfolio.People, process and theft problems: Inventorymanagement, particularly in-store is still a hightouch process requiring human intervention.One of the most pressing issues for solvingout-of-stocks is the decreasing marginSKU assortment optimizationIn this use case, we focus on a consumer goods’problem of SKU assortment, optimizing availableSKUs, and tailoring assortments by store to maximizerevenue. We solve for this by developing algorithmsfor SKU assortment optimization. This is an approachthat uses advanced analytics on big data to createmitigation plans. The partner we focus on, NealAnalytics, uses Azure to create powerful visualizationsthat support actionable decision making throughenhanced insights.Questions often asked related to solvingout-of-stocks: Which SKUs are performing best in a givenmarket or store? Which SKUs should be allocated to a givenmarket or store based on their performance? Which SKUs are low performing and should bereplaced by higher performing SKUs? What other insights can we derive about ourconsumer and market segments?Shelf ownershipThe problem of keeping stocks shelved belongs toeither the retailer, or to the consumer brand. Bothcare about keeping stocks shelved. Ultimately, shelfownership determines who manages SKU assortmentoptimization.Scenario: food and beverageFor this use case, we focus on ownership by aconsumer goods brand related to food and beverage.A category manager is the primary decision maker,supported and consulted by brand management andproduct planning. If the retailer owns the shelf, the1EKN and Microsoft, Plugging out-of-stock gaps in consumer goods.

01[continued] merchant/category manager may be the primary ownerin partnership with the buy planner, who managesdemand and financials, and an allocator (in some retailformats), who supports the distribution and flow ofinventory.It is also important to note the type of merchandisecategory—hardlines vs. softlines—which may dictate thelevel at which you will manage assortment; SKU or Stylerespectively. For example, food and beverage categorytends to have many SKUs that are rather persistent overseason and time. On the other hand, apparel may havemany SKUs to reflect the number of colors and sizecombinations available. Therefore, style/color level datais more applicable to managing assortment.

01[continued]“When we started with this project, we weresearching for new and improved ways toserve our clients and consumers whilebooting profitability. We needed to better usethe data we already had and gain a morecomprehensive understanding of salesvariations and correlations between multiplevariables.”Lizeth Refugio Salas, Revenue Growth Management Chief, Arca Continental“The first project was about marketing. Butthere are other areas we want to pursue, likeproduction, logistics, and warehousing. Wecan use Azure ML to generate answers foreach individual area and get combinedanswers for the entire company.”Ruben Dario Torres Matinez, IT Manager,Arca Continental

02Solution overviewStrategyvisualizations in an easy to deploy, repeatable mannerthat require less custom code to develop.The solution must be able to handle millions of SKUsand outlets (customers) by segmenting sales datainto peer groups that enable detailed comparisons.The outcome of the solution is the ability tomaximize sales at every outlet or store by tuningproduct assortments using advanced analytics andvisualization tools to enable insights. Improvedassortment optimization is estimated to result in a5-10 percent sales increase. Insights result in: Understanding SKU portfolio performance andmanage low performers Optimized distribution of SKUs to reduceout-of-stocks Understanding how new SKUs supportshort-term and long-term strategies Creating repeatable, scalable, and actionableinsights with data already availableFor the current scenario, Neal Analytics works with acustomer to identify a set of data sources particular tothe food and beverage business. Capture starts withcollecting sales history, customer demographic, andSKU data data across a geography of stores. Thesubsequent data process involves standardization,aggregation, and storage of data to be used formodel training and feature creation. Once the modelis trained and the data pipeline is ready for production, the components are read to operationalize.Once the solution is in production, the data used toscore the model and the model predictions arecaptured and stored for dashboarding and reporting.Users can then use PowerBI to create custom visualizations. Through visualizations, users gain insightsthat help them to plan SKU optimization.Let’s look at the components of an assortmentoptimization solution (Figure 1).To address this use case, Azure offers severalout-of-the-box services that can be used to build ascalable analytics solution, that is cost effective andcomprehensive. These services provide industryleading functionalities needed for compute, connectivity, storage, machine learning, reporting and dataCaptureDatapreprocessingCustomersource dataProcessCustomer source data can be defined as: Sales—transaction, sale price, quantity, transaction location, etc.ModelAggregate salesdataTrain machinelearning modelStore dataLoad historical andproduction sale dataJoin andprepare datafor machinelearningExport resultsOperationalizeStore dataBuildreportingFrontlinedashboardsFigure 1: Assortment Optimization SolutionExport data tovisualization toolBusinessdashboards

02Building the solution Outlet (customer)—geography hierarchy, outletidentifiers, date, time, register, etc. Product dimension—product hierarchy, SKU,product name, price, cost, etc.Data preprocessing prepares large datasets to becost-effectively parsed into manageable sets to beleveraged for machine learning models. This canoccur on-premises, or directly in the cloud dependingupon your requirements and preferences. If in thecloud, data would likely be stored in Azure BlobStorage to be consumed by Azure Databricks, orAzure Machine Learning Workbench. At this stage,data integration and aggregation are typically doneby time or by product group, and not by customer, todeliver as much insight and detail as possible. Oncethe datasets are curated, model features are developed using a variety of Platform-as-a-Service (PaaS)offerings such as HDI Spark or Azure Databricks. Toolsselection is often driven by preference and internalrequirements.Data is stored in Azure SQL Data Warehouse, andthen fed into Azure Machine Learning services, wheremodels are trained, and APIs are published. AzureMachine Learning deploys predictive models into avariety of environments. Models are stored, registeredand managed via the Azure cloud. Azure Data Factoryis a cloud data integration service. It lets you composedata storage, movement and processing services intoautomated data pipelines such as batch scoring andretraining of the models over time.Azure Machine Learning Environment has severaloptions. Azure Machine Learning Studio manages andmaintains models more easily through built-inautomation and a drag-and-drop user interface. Theinterface facilitates model development, managementand operationalization for any dataset up to 10 GB ofdata. For larger datasets, or model workloads thatrequire significant compute or scalability, a first-partyAzure service like Azure Machine Learning Workbenchor Databricks is recommended.Reporting and visualization layers will vary depend-Figure 2: Assortment Optimization Solution leveraging Azure Services