Transcript JAMES
Data Quality: Opportunities, Data, and Examples 1 Data Woes We are agents of CHANGE! The Kübler-Ross grief cycle …roller-coaster ride of activity and passivity as the person wriggles and turns in their desperate efforts to avoid the change. 2 3 Better and More Data – Level of analysis • Take a quick look at what/why use data • Linking data from disparate and third party sources – Explore data types – Typical issues & Tricks • • • • Cross validation and sourcing Reverse Look-up GIS layering Backfill from text correlated to codes – Information from operations • Text analytics 4 General Organizational Overview An information business focused on risk taking. Make. Sell. Serve. Sales and Distribution Underwriting Risk Selection and Pricing Portfolio Management Premium Adequacy Billing and Collections Management 5 Producer Segmentation Market Planning Revenue Forecasting Cross sell and Up sell Retention and Profitability Claims Payment Accuracy Claim Collaboration > Fraud Detection > Subrogation > Risk Transfer > 3rd Party Deductible > Reinsurance Recoverable Same Problems – Different Lines of Business • Personal – Auto, HO, Umbrella • Small Commercial – BOP, CPP • Middle Market Commercial – CPP w/GL, CP, Crime, CIM, • • • • • 6 B&M, WC, Auto Large Commercial Accounts Commercial Auto Workers Comp Umbrella/Excess Specialty Lines – D&O, EPL, E&O, Farm, FI Data Types and Forms Structured data Semi-structured data Unstructured data Text Spatial Pictographic Graphic Voice Video 7 Multiple Data Systems which must be pulled together for analysis. Great opportunity for cross-validation and sourcing Vendors/Partners Archive, Legacy Systems Current System Claim Medical Data - Bill Review - PPO - Case Management - Paradigm Data External Data Policy Multiple Underwriting Systems ACTIONS Multiple States Billing Systems Finance Systems CRM Systems, other data • Identify Data Systems • Get right data from right systems • Overcome internal Organizational Barriers • Bridge to legacy systems and archived data • Augment to create rich data mining environment • Expect the need to negotiate for resources 8 Some typical external data sources and vendors Dun & Bradstreet Experian Bureau of Labor and Statistics Market Stance AM Best Equifax US Census Claritas Melissa Data ISO GIS vendors U&C Data sets Code Sets for ICD-s and CPT’s … 9 Data Glitches – historical and on-going Systemic changes to data not process related – Changes in data layout / data types – Changes in scale / format – Temporary reversion to defaults – Missing and default values – Gaps in time series 10 Process Reasons for poor data entry 11 Defining Issues-sample Source Data 1-Define Issues 12 Constants Definition Mismatches Filler Containing Data Inconsistent Cases Inconsistent Data Types Inconsistent Null Rules Invalid Keys Invalid Values Miscellaneous Missing Values Orphans Out of Range Pattern Exceptions Potential Constants Potential Defaults Potential Duplicates Potential Invalids Potential RedundantValues Potential Unused Fields Rule Exceptions Unused Fields MORE ISSUES… Mapping across sources: Same Fact, Different Terms Data Element Concept Name: Country Identifiers Context: Definition: Unique ID: 5769 Conceptual Domain: Maintenance Org.: Steward: Classification: Registration Authority: Others Algeria Belgium China Denmark Egypt France ... Zimbabwe Data Elements Name: Context: Definition: Unique ID: 4572 Value Domain: Maintenance Org. Steward: Classification: Registration Authority: Others 13 Algeria L`Algérie DZ DZA 012 Belgium Belgique BE BEL 056 China Chine CN CHN 156 Denmark Danemark DK DNK 208 Egypt Egypte EG EGY 818 France La France FR FRA 250 ... ... ... ... ... Zimbabwe Zimbabwe ZW ZWE 716 ISO 3166 French Name ISO 3166 2-Alpha Code ISO 3166 3-Alpha Code ISO 3166 3-Numeric Code ISO 3166 English Name Data Filling • • • • • 14 Manual Statistical Imputation Temporal Spatial Spatial-temporal Geographic Hierarchy 15 Deriving Data = Power 16 Totals: Household Income Trends: Rate of Medical Bill Increases Ratios: Claims/Premium, Target/Median Friction: Level of inconvenience, ratio of rental to damage Sequences: Lawyer-Doctor, Auto-Life Policy Circumstances: Minimal Impact Severe Trauma Temporal: Loss shortly after adding collision Spatial: Distance to Service, proximity of stakeholders Logged: Progress Notes, Diaries, Who did it, When, “Why” Deriving Data = Power (Cont’d) 17 Behavioral: Deviation from past usage, spike buying Experience Profiles: Vendor, Doctor, Premium Audit Channel: How applied, How reported, Service Chain Legal Jurisdiction: Venue Disposition, Rules Demographics: Working, Weekly wage, lost income Firmographics: Industry Class Code Vs Injuries Claimed Inflation: Wage, Medical, Goods, Auto, COLA Gov’t Statistics: Crime Rate, Employment, Traffic Other Stats: Rents, Occupancy, Zoning, Mgd Care “Search” versus “Discover” Structured Data Unstructured Data (Text) 18 Search (goal-oriented) Discover (opportunistic) Data Retrieval Data Mining Information Retrieval Text Mining Searching Input Value [Jim] Jimmy Jim James Word Replacement Lists Transformed Input Value [JAMES] 19 JAMES JAMES JAMES Returns “Similar Matches” All Records Found: Jimmy Jim James Motivation for Text Mining • • Approximately 90% of the world’s data is held in unstructured formats (source: Oracle Corporation) Information intensive business processes demand that we transcend from simple document retrieval to “knowledge” discovery. 10% 90% 20 Structured Numerical or Coded Information Unstructured or Semi-structured Information Convergence of Disciplines Example 21 Techniques for attacking text data: Rules-based Statistical Text Analysis and Clustering Linguistic and Semantic Clustering Support Vector Machines Pattern Matching or other statistical algorithms Neural Networks Combination of methods from above Text is like a data iceberg 22 Claims processing – Progress notes and Diaries Service •Medical Management Staff •Special Investigation Unit •NICB •Vendor Management •Consulting Engineers •Hearing Representative •Structured Settlement Unit •Recovery Staff •Legal Staff 23 CLAIMS ADJUSTER •Home Office Staff •Field Office Claim Staff •Insured Risk Manager •Agent or Broker •Diary forward – “call Dr Jones next week” •Business Rule – large loss review •System Reminder – update case reserves •Correspondence Tracking – legal letter sent Semantic processing: Named Entity Extraction • Identify and type language features • Examples: • People names • Company names • Geographic location names • Dates • Monetary amount • Phone #, zipcodes, SSN, FEIN • Others… (domain specific) 24 Feedback to UW Forklift Hits Ladder Ladder in Doorway Forklift Couldn’t Stop Or No Barrier Signs No Policy 25 Forklift Brakes Defective Cooking Oil on Floor Forklift Going Too Fast Brake Maintenance Delayed Housekeeping Inadequate Speed Limits Not Enforced Lack of Personnel No Enforcement No Enforcement Data Quality: Opportunities, Data, and Examples 26