Trusted Experts in Business Analytics Predictive Assortment Planning March, 2015
Introductions VP of Advanced Analytics QueBIT Consulting, LLC IBM World Wide Predictive Analytics Lead Retail Director of Merchandising Science Advance Auto Parts Implemented Predictive Assortment Planning at 3 Automotive Parts Retailers (2 US/1 Canada) e: smutchler@quebit.com P: 540-524-8380
Outstanding Reputation for Analytics ROI Experts in analytics strategy, implementation and training 700+ successful implementations in finance, sales, marketing and operations 350+ analytics customers in all types of industries
Repeat Recipients of Excellence Awards IBM Business Analytics Partner Excellence Awards: 2013-14: Worldwide Overall Business Analytics Business Partner Excellence Award 2012: Business Analytics North America Business Partner Excellence Award 2011: Worldwide Financial Business Analytics Achievement Award 2010: Mid-Market Partner of the Year, and Sales Excellence
Agenda Unique Challenge of Automotive Parts Retail High Level Overview of Predictive Assortment Planning Benefits of Predictive Assortment Planning Techniques Organizational Integration
Unique Challenge of Automotive Parts Retail Long tail demand with high expectation of part availability and limited space 9,000+ make/model/year combinations at a single store 100 s of part types Commercial, retail and service channels often have competing requirements of assortments Stores [even in close proximity] can have completely different commercial customers and vehicle populations Clustered assortments are easier to manage but are too broad brush resulting in highly inaccurate assortments with high part duplication of low velocity parts (in-market) Store specific assortments result in higher assortment accuracy but require a very high number of assortment decisions 100 stores x 100,000 stockable SKUs = 10 million decisions!
High Level Overview of Predictive Assortment Planning Data driven Assortment changes are driven by predictive models trained on your data Models are validated using historical hold-out data Not biased by vendor goals Merchants/vendors have input into what goes into the model Scalable The solution can process [score] billions of records in a few hours Highly accurate Typical [first year] sell-through for predictive models on new inventory is 65-85% (varies by part category) compared to 40-60% standard approaches Space efficient / availability maximizing Maximizes part availability for key categories based on demand elasticity Ensures all the parts required to complete the job are in the assortment Optionally: Maximizes market availability by splitting long-tail demand across stores Low risk Simulate changes before deploying pilot in select stores
Benefits All 10-30% increase in assortment accuracy US Retailer 400 bp improvement in inventory turn 7% reduction in inventory levels 11% increase in sales Highest customer satisfaction scores in part availability in company history Scores increased every month 3,000+ hours of merchant time was re-allocated to vendor management, etc.
What will sell? - Propensity Modeling Train models to predict which parts will sell in the next 12 months using parts stocked for the past 12 months (and if they sold or not) The data used to evaluate the models is NOT the same as the data used to train the model.
How many vehicles will need the part? Using historical sales/lookup data we model part failure rates for each part type/make/vehicle type (e.g. Rotors Ford Truck) Rotors
How many vehicles will need the part? 1. Construct a trade area using radius (based on population) 2. Aggregate store/part VIO (using fitment) 3. Weight VIO by lifecycle 800 VIO -> 73 vehicles that will need the part next year at this store
Simple rules automate assortment changes X months stocked without sales = REMOVE Propensity to sell > Y% with incremental VIO coverage of Z = ADD ADD SKUs can be constrained to space freed + some % of growth Expected gross revenue for assortment changes = probability of sale * gross margin
Completing the job having all the parts We mine millions of transactions to find parts that are required to complete the job (e.g. complete brake jobs, water pumps/timing belts, tune-ups etc.) High lift associations between parts are then forced into new assortments (if having an oil pump screen drives more sales (lifts) pump sales significantly then we add the screen along with the pump) SKU 372 is purchased 534x more often if SKUs 540 and 705 are also stocked.
Forward positioning parts based on elasticity Customers will wait longer for certain part categories (in-elastic demand) Forward position (store stock) elastic part categories (e.g. brakes) Re-locate in-elastic part categories (e.g. radiators) to DC or hubs, etc.
Organizational integration Modeling process: We work closely with Merchandising to Identify potential sales predictors Validate model outputs (do ADDs/REMOVEs make sense?) Tweak assortment rationalization rules to meet market/company-wide objectives On-going basis: The system will generate company-wide assortment recommendations on a weekly or monthly basis (100% automated) Merchant organization receives reporting that helps them validate any changes Impact reporting at SKU/line/vendor level (add to X stores, removed from Y stores) Inventory management receives impact reporting to understand supply chain impact Impact reporting at SKU/DC level (added to X stores/dcs, removed from Y stores/dcs)
Teams/Roles 2 FTEs Modeling lead Preferably an internal FMP analyst that QueBIT will help train/mentor to build/tune/evaluate models Analytical leaning Good communication skills (will work closely with Merchandising, etc.) Data lead Strong knowledge of corporate data Strong SQL skills Reporting skills