SaPHAL Sales Prediction powered by HANA and Predictive Analytics 1
SaPHAL Sales Prediction Powered by HANA and Predictive Analytics 1 Introduction - SaPHAL Agenda 2 3 4 Business Case Pain Points & Solution Approach Architecture, Technical details and Key Features 5 The Algorithm 6 DEMO 2
Powered by SAP HANA Easy to use Front End SaPHAL Introduction Effect of Exogenous Variables on Sales Forecast Primary Sales Secondary Sales Stock coverage Competitor Data Vendor Fulfillment Weather Data Rapid Simulations & Scenario Analysis Flexibility & Ease of Use Product Highlights Workflow based Application Live Market Feeds & Instant Notification 3
Demand Variability Business Case Comparison Parameters OEM Supply Chain Aftermarket Supply Chain Nature of Demand Largely Stable, Easy to predict Highly Sporadic, Difficult to predict Required Response Standard, can be scheduled ASAP (same or next day) Product Portfolio Largely Homogeneous Always Heterogeneous No. of SKUs handled X 15 X - 20 X Inventory Turns 6-50 a year 1-4 a year Aim of Inventory Management Delivery Network Maximize Flow of resources, Reduce bottlenecks as well as Inventory carrying costs Depends on nature of product; multiple networks necessary Pre-position resources to support promised Customer Service Levels at optimum Inventory carrying cost Generally Single network, capable of delivering different services Performance Metric Fill Rate Service Level (Turnover time) / Product Availability High Low Forecasting Difficulty Supply Chain Type Responsive Efficient The root of the vicious cycle of having as well as in aftermarket business, is 4 High Low
Business Case Sales Ecosystem Sales Office 1 Sales Office n Main Dealer 1 1 st Trade level Main Dealer 2 Main Dealer n Purview of Demand Forecast includes only Primary sales Inward driven forecasting system 2 nd Trade level Secondary Sales & external factors affecting demand is not considered in current forecasting system Consumer Base Flow of products / Services Downstream Flow of Orders Upstream 5
Primary Sales Business Case Use case Description Unfulfilled Demand 10-15% Non-moving inventory 20% 20% of schemes don t meet the intent The Process Annual Business Plan Broken down to Monthly Sales Plan Inward looking Sales Plan Shortfalls + High Inventory Schemes OEM Sales Secondary Sales Past Scheme Data Weather Data Competition Data Periodic check of Achievability with predictive Algorithms 1. Scheme Proposal backed up scientific data 2. Simulation Based Models on current Parameters Stock Clearance Optimized Process 6
Pain Points and Solution Approach Sales Prediction tool built on Native HANA Application using Rich client SAPUI5/HTML5 Business Drivers/ Pain points Solution Planning model and focus is inward driven - Market & external factors ignored leading to mismatch between supply and demand Inability to understand business impact due to market dynamics and react quickly Product promotion through schemes are intuitive and not based on scientific analysis 7 Complex Algorithm processing at the backend, easy to use UI and Simulation Models on different Parameters Role based screens (Sales office head and Product Manager) to be able to publish and freeze predictions Solution Benefits Sales Prediction using scientific algorithms. Impact of diverse factors such as Rainfall, Secondary Sales, Stock coverage, Vendor fulfillment, Competitor data Simulate key factors using market intelligence, understand possible business scenarios and react quickly to unachievable plan Realistic sourcing and improved demand management leads to better Inventory management Ability to choose right schemes based on past experience
Pain Points and Solution Approach Solution Brief SaPHAL will Predict Sales for the future based on an Algorithm that derives a relationship and trends from past Sales with additional data points e.g. Seasonal data, Secondary Sales, Competitor Info, Vendor Fulfillment, Past Scheme data Features Easy to use frontend with all complex processing of algorithms and Data Mining at the background Simulations on top of system proposed values and be able to version them Summary information for each Sales Prediction Live Market feeds and top influencing factors affecting Sales Workflow in using Sales Predictions Sales Office Head Publishes -> Product Manager reviews and Freezes 8
Convert to High Fidelity Architecture, Technical Details & Key Features Server Solution Layout Primary Sales Data (SAP ECC6.0) Secondary Sales Data (RDBMS) Market Data (OEM + Competition Sales) Data Replication Data services 1. Choosing the best fit algorithm for Forecast and Predefined Simulations 2. Storing Algorithms inside HANA Engine for XS application Online Process of Algorithms for a cumulative data set ~ 25 Mio Records SaPHAL - UI Powered by XS Engine CRM Trade Promotion Unstructured Promotion Data Past 5 years Seasonal data 9
Architecture, Technical Details & Key Features Online Processing of Algorithms and storage as SPs in HANA SAP PA Architecture Source Data Layer Data Rep and Ext Modeling HANA Modeler and DB XS Engine XSJS Native (HANA App) SAPUI5 and HTML 5 SLT* BODS Primary Sales Data (SAP ECC6.0) Secondary Sales Data (MS SQL Server) Market Data (OEM + Competition Sales) CRM Trade Promotions Weather data Mail and Word Document Data No SLT, Primary Sales Data is extracted through BODS 10
Architecture, Technical Details & Key Features Source Data Layer Data Rep and Ext Target Layer ECC (Primary Sales) HANA Artifacts Secondary Sales Weather data Market data Application Tables Native HANA App (JavaScript + SAPUI5 + HTML5) SQL connection Excel, CSV connection Web Service connection Additional support SaPHAL Secondary Sales Weather data Web API Market data (OEM + Competition Sales) Web API 11
Architecture, Technical Details & Key Features Key Features Sales Predictions: Tried and tested algorithm - Vector Auto Regression that analytically foretells sales from exogenous variables. Rapid Simulations and Scenario Analysis : Rapid and relevant simulations on data points in spontaneous relation with the field and local knowledge of respective Sales Managers and Product Managers. 12
Architecture, Technical Details & Key Features Workflow based Application: Sales Office heads can Simulate, version and publish the predictions; Product Manager can review, fine tune prediction with further simulation before freezing for demand planning. Product Manager View Sales Manager View Storing Algorithms inside HANA engine for XS application Flexibility and Ease of Use : State-of-the-art UI developed in SAPUI5/HTML5 with Complex processing of Algorithms and data mining at the background. 13
The Algorithm Forecasting Model based on Vector Auto Regression (VAR) Model can capture the cross effects of exogenous variables (dependent variable at time t depends on different combinations of all independent variables at time t n) f(p) T = f(p1, P2, P3, P4) T-n Weather t Weather t-1 Weather t-2 Stock Coverage t Stock Coverage t-1 Stock Coverage t-2 Sales Forecast Secondary Sales t Secondary Sales t-1 Secondary Sales t-2 Competitor Data t Competitor Data t-1 Competitor Data t-2 Factor n t Factor n t-1 Factor n t-2 14
The Algorithm Data Requirements Exogenous Variable Historical Data Requirement (Yrs) Primary Sales 4 Secondary Sales 4 Stock Coverage 4 Rainfall 4 Competitor Data 3 Vendor Fulfillment 4 15
The Algorithm Assumptions Model assumes that data is stationary If data is not stationary, then transformations are done to convert data to stationarity After transformations, If data is not stationary, then model is not built Facts about the Model Variability of the forecasted values would increase with increase in forecast period Model parameters variance increases sharply with increase in data seasonality, so the model works ideally for large datasets (ideally 4 years or more with 6 variables) Prediction Models are built based on Quantities and Values. However, by default value model is chosen for better accuracy Each variable is evolved based on its own lags and lags of other exogenous variables Lag is determined using AIC information criterion. Causality is determined using Granger Causality test Stability of the model is checked by orthogonal decomposition (Eigen vectors) 16
17 DEMO