Workshop: Predictive Analytics to Understand and Control Flight Risk Data science for deeper insights and more accurate predictions Peter Louch Founder and CEO peter.louch@ (917) 443-4572
Agenda Introductions Background: Predictive Analytics in Workforce Planning A Pragmatic Model for Predictive Analytics 2
Using Predictive Analytics in Workforce Planning Predictive Supply Analytics coupled with Demand Planning provides a blue print for Talent Management activities Business & Market Input: Strategies, Priorities, Market Environment, Maturity/ Growth State of Business, Budgets & Constraints Workforce Data: Demographics, Performance, Potential, Movement, Generational Shifts, Marketplace Talent Availability Talent Demand Staffing needs/timing by job role Productivity targets Scenarios: Impact of programs and changes (e.g., technology, customer demand) Cost of workforce Talent Supply Turnover risk Retirement risk Mobility projections Talent availability (industry surplus/shortage) Outcomes Guide Talent Management Talent Gap Prioritized Current State & Future State! Recruiting and Onboarding! Training and Development! Redeployment! Knowledge Transfer and Exits 3
Getting to Effective WFP What The Journey Looks like PERFORMANCE ILLUSTRATIVE " WFP Topic Poor Fair Effective Highly Effective Demand Planning Each manager decides on their need for talent. Managers have low industry experience Managers with high industry experience decide on their demand. Lead managers through a demand decision tree based on a job role criticality model Use demand drivers: e.g., oil & gas production workers per well path by project stage Internal Supply Analysis Don t factor in turnover Project forward turnover based on historical or industry rates Regression analysis to establish turnover, retirement and movement risk by job role Data mining, predictive analytics to establish risk of specific individuals External Supply Analysis Ignore market shortages when planning Use directional but stale data sources like BLS Avoid external supply analysis if there is not a tangible solution to offer the business Potentially, use big data solutions that mine actual market demand, labor supply Gap Analysis and Action Planning Use the WFP process for education / discussion but don t change behaviors based on the gap Use the gap to formulate hiring plans. Track performance against plan Use the gap to formulate a buildbuy-lease plan (i.e., recruit, development and contract out). Optimize where information available. Based on predictive analytics, respond with changes in hiring and talent management strategies and/or mitigate individual employee risks. 4
Challenges/Barriers to Use of Predictive Analytics Challenges/Barriers Basic confusion about retirement eligibility vs. retirement likelihood Traditional methods are poor at predicting the individuals who turn over and the timing Managers have gut feelings and anecdotes and don t want to listen to data Management won t act until they ve already felt the pain. Workforce analytics/planning function held responsible for data integrity issues (i.e., basic blocking/tackling issues that have low impact on forecasts by high impact on credibility) Change in business environment for utilities (e.g., lower customer demand, cost controls) have led to new employer-employee relationship with more turnover Potential Solutions Assume nothing, start every conversation showing the difference Use machine-learning statistics that allow for complex multi-variate analysis Get buy-in from senior leaders to help drive workforce agenda (better process for all) People buy insurance if they identify with the persona. Share stories of similar entities to estimate future pain, and then agree on ROI of avoiding it Create SLA with the customers of predictive analytics. Own the solutions consulting to solve data integrity issues and not the problems. Machine learning statistics allow sampling smaller data sets and less history to more quickly respond to changes in turnover drivers 5
Case Example: Challenges/Barriers to Use of Predictive Analytics Retirement at North American Utilities * Source: EUHRMG April 2013 Survey, conducted by Vemo Little connection between Retirement Eligibility and actual retirements Huge upside for using machine-learning statistics to create a much more accurate retirement forecast that the business can rely upon in talent planning 6
Case Example (cont d): Improvement opportunity using predictive analytics 163% Overshoot!" 326% Overshoot 67% Undershoot!"#$%&'"()"%*$)'$+'",&*-('%".%&'" /0"1+(.2",&,/&+%" 7
Case Example (cont d): Improvement opportunity using predictive analytics!"!"#$%&'"()"%*$)'$+'",&*-('%".%&'" /0"1+(.2",&,/&+%" 8
Case Example (cont d): Five-Year ROI from Predictive Modeling 9 Extra Hires 10 12 Extra Hires Extra Hires 9 10 Extra Hires Extra Hires 9
Case Example (cont d): Five-Year ROI from Predictive Modeling 9 Extra Hires 10 12 Extra Hires Extra Hires 9 10 Extra Hires Extra Hires 10
Case Example (cont d): Five-Year ROI from Predictive Modeling 2 1 2 0-1 11
Case Example (cont d): Five-Year ROI from Predictive Modeling ROI for company with 156 employees in target demographic: Total Extra Hires Without Model: 50 Total Extra Hires With Model: 4 Total Extra Hires Avoided = 46 Cost per Extra Hire X 100,000 ROI = $ 4,600,000 12
Predictive Analytics Model Pragmatic Five-Step Model $ Don t create impossible barriers to realizing benefits of Data Science Reference: Are You Recruiting A Data Scientist, Or Unicorn? Information Week, November 2013 The 5 step model shown above, proven through numerous engagements, breaks down into five interconnected workforce and predictive analytics activities that yield Data Science 13
Input $ Part 1: Data Acquisition Data Type Estimated Usage HR Action Data Standard Organizational Data 100% Employee Profile Data Standard Employee Data 100% Performance/Talent Management Data Supplemental Employee Data 60% Industry Data Benchmarking 30% Market Data Big Data 10% 14
$ Part 2: Data Transformation Input Throughput Differentiated Output (Attrition Example) HR Action Data Employee Profile Data Performance/Talent Management Data Industry Data Market Data Validation and Model Refinements Transformed Data (Data Warehouse of All Workforce Actions and Attributes) Refined Transformed Data (Additional Data Transformation to Optimize Data for Statistical Analysis) Traditional Statistical Models to Determine Correlation of Factors to Workforce Events (e.g., Attrition) Machine Forecast to Establish Risk of Workforce Events (e.g., Attrition) Segment entire workforce or specific populations (i.e., high value workforce) into risk bins (i.e., high-low or highmedium-low) Identify and micro-target populations for interventions Make policy/structural changes and measure improvement in turnover Institute changes in hiring practices and measure improvement in turnover In short, make dataguided management decisions 15
$ Part 3: Data Analysis Main Objectives of Data Analysis Key Drivers Analysis Identify the drivers of key outcomes (turnover, performance, retirement) Design interventions to improve outcomes Forecasting Pricing Estimate the flight risk for each employee Divide employees into risk bins Forecast other outcomes of interest, such as performance, engagement or retirement Data-driven Counterfactuals Model the likely impact of hypothetical events on outcomes (ie, how would a salary freeze affect turnover? ) Re-estimate forecast after simulated interventions 16
Multivariate Regression: The 40-40-20 Principle $ By including a wide range of predictors in the same model, we can identify the relative importance of various drivers Several of our past turnover studies conform to the 40-40-20 principle: 20% External Economic Conditions - Local Job Market - Stock Market - Other 40% Employer Actions / Pricing Pricing Workplace Conditions: - Promotions and Raises - Spans and Support Ratio - Work Environment - Salary Inequality - Manager Ratio 40% Employee Demographics - Tenure and Age - Job Type and Salary Range - Commute Time 17
$ Case Example of Driver Analysis Customer Business Unit X has 3 rd quartile turnover compared to peer group (i.e., not good but not horrible) Vemo Predictive Analytics used for turnover driver analysis, forecast, modeling interventions Scope Business unit has active regular employee headcount of ~4500 employees Excludes temporary and supplemental workers 18
Flight Risk by Tenure by Employee Level $ # Managers have typical pattern of flight risk diminishing as tenure increases # Professionals (individual contributor peers to Managers) have atypical pattern of increasing risk through Year 4, then sharp decrease in flight risk 19
$ Flight Risk by Salary by Grade Rank # Grade Ranks are 1 through 7, Director, Executive # For Grade Rank 1 and 2, very strong negative association between salary and flight risk # For Grade Rank 2-4, negative association between salary and flight risk # For Grade Rank 5-7, virtually no relationship between salary and flight risk Key The chart shows how flight risk varies across the typical salary range for each Grade Rank. The horizontal distance covered by each line corresponds to the typical salary range for that Grade Rank. 20
Flight Risk by Salary Inequality of Work Group $ # For low earners, a high level of salary inequality greatly increases flight risk # For middle of pay scale earners, salary inequality modestly increases flight risk # For high earners, salary inequality has no impact on flight risk # Structural reform to create work groups where similar workers are in each VP s group could have high retention impact, albeit with significant change Salary Inequality within Work Group (Department Owner 1) is a measure of overall distribution of pay within a workforce. A high level of salary inequality indicates that high earners make much more than the average level, and low earners make much less. In these graphs, moving from left to right the X-axis indicates that the level of salary inequality is increasing. 21
Flight Risk by Amount of Raise by Employee Type $ # Impact of raises varies significantly by Employee Type o Additional raises have highest impact for Employees, particularly for Employees receiving standard and sub-standard raise amounts o Additional raises have moderate impact for Professionals o Additional raises have little impact for Managers Workforce Type Employee Manager Professional Potential Intervention of 1% additional raise when current raise is 0-2% when current raise is 2-5% when current raise is 0-9% when current raise is 0-5% when current raise is 5-9% Impact High Impact Moderate Impact Low Impact Moderate Impact Diminishing Returns 22
$ Flight Risk by Recent Promotion # A promotion has a strong negative impact on flight risk in short term, however, the impact appears to diminish quickly over time o o Employees who have had promotion in last year are 85% less likely to turnover relative to employees who have not received promotion Employees who have had a promotion that occurred over 18 months have essentially the same flight risk as those who did not receive a promotion 23
Flight Risk by Lateral Transfer by Generation $ # Impact of lateral transfers varies dramatically by Age Generation o For Gen Y, there is a moderate impact it decreases flight risk by 22% o For Gen X, there is no impact o For Baby Boomers, there is strong adverse impact it increases flight risk by about 300% Key: A Lateral Transfer is a true job change without a promotion or demotion. For predictive modeling, it is typically important to ignore HRMS action codes and use business logic for promotions, demotions and lateral transfers. 24
$ Employee Turnover Forecast The forecasting model assigns employees to one of three configurable bins: 1) Highest Flight Risk 2) Moderate Flight Risk 3) Lowest Flight Risk Lowest Risk Bin Moderate Risk Bin Highest Risk Bin 25
Six-Month Forecast Summary Compared to Actuals $ # Chart shows predicted rolling 6-month turnover rates by Department Owner 3 Work Units with > 100 employees # Compares to historical 6-month forecasts vs. actuals which show good performance at aggregate level of forecast 26
Manipulating a Forecast to Explore Potential Interventions $ Forecasting models can be used to simulate the effect of a particular intervention on aggregate turnover rates The forecast is based in part on a certain set of assumptions (ie, the annual raise policy will remain at it s current level) By altering these assumptions (ie, by going into the data and assigning a 2% increase Pricing in the annual raise for each employee), we can simulate the likely impact of an intervention on turnover or other outcome variable 27
$ Simulating an Increase in Raises 2% 4% 0% 6% Assumed Annual Raise Amount 28
$ Simulating an Increase in Raises 2% 4% 0% 6% Assumed Annual Raise Amount 29
$ Simulating an Increase in Raises 2% 4% 0% 6% Assumed Annual Raise Amount 30
$ Simulating an Increase in Raises 2% 4% 0% 6% Assumed Annual Raise Amount 31
Predictive Analytics Model Pragmatic Five-Step Model $ Contact: Peter Louch Founder and CEO peter.louch@ 32
Appendix: About Vemo About Vemo Founded in late 2005 by leading experts in workforce planning Our customer base is comprised of leading organizations who implement highly effective, sustainable enterprise workforce planning and analytics programs Subscription Software Automated and End User Workforce Planning Slice & Dice Workforce Analytics for Analysts Push Reporting and Scorecards for End Users Predictive Analytics Benchmarking Consortia Vemo Differentiators State-of-the-art data model that can work with any data source/client business rules. Unlimited scenario planning, using both strategic and operational demand drivers. Unique analytics product that allows non-expert line, operations and HR users to conduct deep analysis through on-demand charts/grids. Internal supply forecasting integrated into tool. Predictive analytics to improve talent management. Consultative approach Dedicated to customer success. Sample Vemo Customers 33