A Comprehensive System for Selecting and Evaluating DFA Model Parameters Chris Madsen, ASA, CFA, MAAA American Re-Insurance Company CAS DFA Forum, Chicago July 19th-20th, 1999 Chris Madsen, ASA, CFA, MAAA American Re-Insurance Company CAS DFA Forum, Chicago July 19th-20th, 1999
Discussion Overview Overview of a Integrated Risk Management System Focus on an Economic Model Calibration examples Optimization issues Conclusions Overview of a Integrated Risk Management System Focus on an Economic Model Calibration examples Optimization issues Conclusions
Model Structure
M2 Growth V2 Growth Inflation* GDP Growth* Interest Rates* (Forward, Spot, Yield) Equity Earnings Yield Equity Earnings Growth Asset Model * Currency Link (not currently modeled) Economic Model
SimulationDefining the r/i structure Modeling the portfolio Gross loss Net loss Ceded loss Retained premiums Ceded premiums Loss Simulation with DFA Loss data Premiums Customer requirements Limits Prices
What Makes A Good Scenario Generator? Logically defensible Economic theory Historical data Risk across time Logically defensible Economic theory Historical data Risk across time
Plausible Paths No negative interest rates Historical data does not necessarily equate expected value of statistics (trend sensitive) - rather, build distributions of statistic and ensure history is well-represented. No negative interest rates Historical data does not necessarily equate expected value of statistics (trend sensitive) - rather, build distributions of statistic and ensure history is well-represented.
Types of Models Strategic Long-term planning resource allocation (capital, business mix, asset mix, retro covers) Pricing Risk-neutral (replication) Does often generate unreasonable simulations (all returns = risk free rate) Strategic Long-term planning resource allocation (capital, business mix, asset mix, retro covers) Pricing Risk-neutral (replication) Does often generate unreasonable simulations (all returns = risk free rate)
Economic Model Long interest rates dl t = a l ( l - l t ) dt + l t l dZ l Short interest rates dr t = a r ( r - r t ) dt + r t r dZ r Long interest rates dl t = a l ( l - l t ) dt + l t l dZ l Short interest rates dr t = a r ( r - r t ) dt + r t r dZ r
Setting Targets Basic statistics (arithmetic mean, compound mean, st.dev., percentiles, min. & max., serial) Plausibility criteria (Becker - yield curve characteristics) Basic statistics (arithmetic mean, compound mean, st.dev., percentiles, min. & max., serial) Plausibility criteria (Becker - yield curve characteristics)
Target Example History Simulation
Calibration Example #1 Regressing on ‘74-’98, we get {A, B, C}={0.015, 1.3, } R 2 =58% 90% parameter confidence D=1.06 (ln(residual/mean)) Regressing on ‘74-’98, we get {A, B, C}={0.015, 1.3, } R 2 =58% 90% parameter confidence D=1.06 (ln(residual/mean))
Calibration Example #1 {A, B, C, D} = {0.75, 0.5, -0.04, 1.05} The two are quite similar though at first glance… Weight shifted from 30 Year Rate to Inflation Mean reversion up Volatility down slightly {A, B, C, D} = {0.75, 0.5, -0.04, 1.05} The two are quite similar though at first glance… Weight shifted from 30 Year Rate to Inflation Mean reversion up Volatility down slightly
Calibration Example #2
Conclusions Regression is a good starting point but may miss key statistics Key statistics may miss fundamental relationships Optimization is a valuable parameterization tool and enables us to monitor key statistics as well as fundamental relationships Regression is a good starting point but may miss key statistics Key statistics may miss fundamental relationships Optimization is a valuable parameterization tool and enables us to monitor key statistics as well as fundamental relationships