WHICH MODEL WORKS BEST FOR YOU?

Transcription

1 1 WHICH MODEL WORKS BEST FOR YOU? RMA s Annual Risk Management Conference Dallas, Texas November 14, 2016 Brooks Brady SVP Credit Risk Analytics and ALLL Zions Bancorporation

2 2 DISCLAIMER The views in this presentation are my own and do not necessarily reflect the view of Zions Bancorporation or its subsidiaries This presentation does not represent an endorsement of any product or service marketed at this conference The views in this presentation are not advice on regulatory or accounting issues

3 3 ABOUT ZIONS BANCORP AND BROOKS Zions Bancorp 7 local brands in the West and Southwest United States $60 billion in assets (SIFI) >70% of assets in Utah, Texas, and Coastal California About Brooks Credit risk modeling, data, and analytics ALLL, risk grading, and portfolio modeling Currently: SVP Credit Analytics at Zions Bancorp Formerly: Standard & Poor s, American Express, KPMG Masters Finance, NYU; BA Mathematic, Univ. of Utah

4 4 Standalone WHICH VIDEO GAME IS BEST? Integrated

5 5 WHICH CREDIT RISK MODEL IS BEST? Basel Committee for Banking Supervision say the best model is one that is integrated: Guidance on credit risk and accounting for expected credit losses, December The Committee expects banks to leverage and integrate common processes [that] encourage consistency in the measurement, management and reporting of credit risk and ECL. Forward-looking information, including economic forecasts and related credit risk factors used for ECL estimates, should be consistent with inputs to other relevant estimates within the financial statements, budgets, strategic and capital plans. The use of common processes, systems, tools and data strengthens, to the maximum extent possible, the consistency of the resulting estimates and minimises disincentives to following sound credit risk practices for all purposes.

6 6 INTEGRATE PRIOR TO MODEL BUILD Begin with the end in mind not just for the model you own, but for the entire organization What good is a model that is internally conceptually sound but breaks downstream processes? Start by understanding how the outputs of your model will be used outside of your domain, and make sure they integrate

7 7 INTEGRATE MODEL OBJECTIVES Stand-alone model objectives often differ from integrated model objectives example risk grade model changes: Standalone Model estimates an average 12- month default rate Large shifts in grades due to model change are acceptable Integrated objectives Model estimates a grade Zions Credit Vision: We lend to businesses [that can] weather adverse economic conditions. differs from average default rate DFAST and ALLL models estimate default rates conditional on economic conditions and for longer or shorter periods Regulatory grade definitions do not reference PD percentage Grade shifts due to model change should be minimized and offset DFAST and ALLL model development is dependent on grade consistency through time Portfolio and borrower hold limits are based on grades Grade migration affects the Fed s CCAR models

8 8 HOW DID ZIONS MODEL OIL & GAS? $3.2 billion oil & gas outstanding as of 12/31/2014 Zions needed to estimate oil & gas portfolio losses for investor disclosures, CCAR, and planning purposes When oil prices started to drop in late 2014, Models did not directly correlate losses with oil prices Zions had little internal historical data upon which to build a model, except for a brief oil price drop in 2008/9

9 9 HOW DID ZIONS MODEL OIL & GAS? 1. CCAR model, then based on industry employment, later updated to include oil price 2. Peak historical loss rate on oil & gas 3. Recent 12-month default and loss given default rates 4. Peak historical oil & gas Classified rate multiplied by historical Classified loss severity rate 5. Peak historical loss rate on Zions construction loans 6. Historical average loss factors by grade, applied to various loan downgrade sensitivities/simulations a. Loan-by-loan manual sensitivity analysis b. What if all oil & gas loans were downgraded to Special Mention

10 10 MODELS IN A RISK MANAGEMENT PROCESS RMA s Annual Risk Management Conference Dallas, Texas Jacob Seljan Senior Vice President of Credit Risk Analytics November 14, 2016

11 11 DISCLAIMER The views in this presentation are my own and do not necessarily reflect the view of U.S. Bank or its subsidiaries This presentation does not represent an endorsement of any product or service marketed at this conference The views in this presentation are not advice on regulatory or accounting issues

12 ABOUT U.S. BANK AND JACOB U.S. Bank Commercial bank ranking 5th largest Period-end assets $438 billion Period-end deposits $318 billion Period-end loans $269 billion Customers 18.6 million Bank branches 3,122 ATMs 4,923 Year founded 1863 At quarter end, June 30, 2016 Jacob Seljan Senior Vice President of Credit Risk Analytics CCAR/DFAST Credit Loss Forecast Models, Basel Advanced Credit Risk Models, Economic Scenarios At U.S. Bank since 2003, previously CGI, Green Tree Financial MBA University of Minnesota, BS Economics, Lewis & Clark College 12

13 MODEL DEVELOPMENT PROCESS Models used to satisfy regulations to quantify and manage risk require a well defined, structured process with input from multiple stakeholders in order to be useful Too much input from statisticians reduces business intuition Too much input from business owners reduces statistical validity Eight Step Process Design Model 2. Aggregate Data 3. Select Variables and Models 4. Test Models 5. Document Models 6. Validate Models Independently 7. Implement Models 8. Execute/Monitor Models The Best Model Formula Best Model = F(0.5*BI+ 0.5*SI) Where: BI = Business Input, SI = Statistical Input

14 MODEL DEVELOPMENT PROCESS 14 Select 1. Design 2. Aggregate 3. Variables & 4. Model Data Models Test Models Project Kick Off Rationale for project Inventory industry practices Proposed model type(s) Portfolio Segmentation Analysis Review, Challenge, and Approval Data Source collection, compilation, aggregation Data quality control check and reconciliation Data preparation and variable derivation Data assumptions and limitations Review, Challenge, and Approval Single factor analysis: missing treatment, derived variables, significance, correlation tests Business input and intuition Statistical selection process Variable coverage Review, Challenge, and Approval Assumptions and Limitations Tests: Goodness of fit, insample/out-of-sample, stability, backtesting, sensitivity tests Review, Challenge, and Approval 5. Document Models 6. Independent Validation 7. Implement Models 8. Execute & Monitor Models Draft initial documentation Management Review and Feedback Finalize document Review, Challenge, and Approval Receive final document from steps 1-5, related meeting minutes and code Validation Q/A Validation Issues Report, ratings, recommendations Review, Challenge, and Approval Determine/document implementation schedule Design data / process flows Complete implementation testing (IT, UAT), document results Migrate to production environment Review, Challenge, and Approval Plan and schedule Model Execution Model results review (internal and committee) Model monitoring results Benchmark and overlay discussion (based on model limitations) Results documentation Review, Challenge, and Approval

15 MODEL EXECUTION: CASE STUDY 15 The Ask What is the potential incremental credit expense of a severe and protracted decline in oil prices over 9 quarters? What is the impact on the Company s Corporate Energy Portfolio and knock on effect to Commercial Real Estate and Consumer portfolios? This occurs in conjunction with a severe recession (e.g. Supervisory Severely Adverse Macroeconomic Scenario) Unemployment doubles from 5% to 10% GDP Contacts 6% HPI declines 25% and CRE PPI declines 30% S&P 500 declines 50% Low interest rates

16 MODEL EXECUTION: DEFINE SCENARIO Step 1: Define the Scenario Oil experiences a significant price shock due to excess supply and a significant decline in demand due to the recession. As a result, crude oil prices decline 50% from $50 to $25 per barrel by the end of the first quarter of the scenario and remains depressed. The fundamental shift in oil prices has a pronounced and immediate adverse affect on the Company s borrowers whose income are dependent on oil revenue. Scope: Corporate, Commercial Real Estate, and Consumer Commercial Real Estate property values are adversely impacted in states more heavily dependent on price of crude oil (based on oil as % of state GDP) Consumer exposures are also more adversely impacted in these states (Mortgages, Auto, Credit Cards) 16 Question: what to do with borrowers that are positively affected by the fundamental shift in crude oil prices? Question: where does the scenario stop? Credit, Operational, Market, Revenue?

17 MODEL EXECUTION: KNOW YOUR MODELS Step 2: Know your models Corporate Loss forecasts are based on a PD, LGD, EAD framework, oil prices are not in the models, but related risk drivers are PD models include S&P Energy Sector Equity Index as risk driver 17 LGD models are based on a mix of collateral (working capital, depreciable assets, unsecured) EAD models are based on utilization and risk rating CRE Mortgage & Construction Loss forecasts are based on a PD, LGD, and EAD framework, oil prices are not in the models, but property location is PD models include drivers that reflect property values and property state information LGD models include drivers that reflect property values and property state information EAD models are based on utilization and risk rating (for construction) Consumer (Mortgage, Auto, Cards) Loss forecasts are based on a PD, LGD and EAD framework, again oil and fuel prices are not in the models, but borrower location is PD models include state-level unemployment rate LGD models include location and property value information (real estate secured) EAD models based on utilization (for lines)

18 MODEL EXECUTION: ADJUST FORECAST Step 3: Adjust your modeled loss forecasts Corporate 18 PD assume a fundamental shift in the risk rating of your borrowers based on qualitative factors (e.g. 1-2 notch downgrade), since the S&P Energy Sector Index is in the model, compare its relationship to oil prices (90% correlated), stress it accordingly (e.g. 50%). LGD assume the underlying collateral declines 50%, thus recoveries decline from 60% to 30%. EAD assume a full draw on borrowers that default. CRE Mortgage & Construction PD assume fundamental shift in risk rating of borrowers in oil dependent states and stress property values in these locations beyond that assumed in the stress scenario LGD assume property values are more adversely affected in states more oil dependent EAD assume a full draw on defaults (Construction) Consumer (Mortgage, Auto, Cards) PD place additional stress on unemployment in states more oil dependent LGD place additional stress on property values (for mortgages) in states more oil dependent EAD no adjustment

19 MODEL EXECUTION: REVIEW RESULTS Step 4: Execute Scenario and Review Results Estimate incremental loss by: Portfolio; Assumption; and Quarter. Evaluate results for reasonableness Scenario is designed to be severe, but if results are not intuitive, reassess assumptions It is okay to question the results Revise assumptions as appropriate Document results State assumptions and limitations 19