Optimal Decentralized ALM Jules H. van Binsbergen, Stanford University Michael W. Brandt, Duke University and NBER Ralph S.J. Koijen, University of Chicago Rotman ICPM Forum, Toronto June 3-4 2008 © Michael W. Brandt, 2008 All rights reserved without exception Motivation Motivation Decentralized investment management • Institutions decentralize investment decisions along asset classes – Example CIO Fixed Income Portfolio Manager Equities Portfolio Manager – Why? Division of labor Specialization ) Generating value (i.e., lower transaction cost or positive alpha) within an asset class requires specialized skill © Michael W. Brandt 2008 3 Optimal Decentralized ALM Motivation Misalignment of objectives 1. Suboptimal diversification – Joint optimization by CIO over all assets dominates best combination of portfolios optimized by portfolio managers within asset-classes – Sharpe (1981) and Elton and Gruber (2004) 2. Different risk preferences – Portfolio managers take more or less risk than CIO desires – CIO does not generally know the managers’ risk preferences – van Binsbergen, Brandt, and Koijen (2008) 3. Mismatch risk between assets and liabilities – Portfolio managers do not consider liabilities in their optimization – Main motivation for this project © Michael W. Brandt 2008 4 Optimal Decentralized ALM Motivation Benchmarking • Performance benchmarks are commonly used to evaluate and compensate portfolio managers – Emphasis is on measuring effort or skill – Benchmarks are taken as exogenously given (e.g., cash or index) • We examine to what extent optimally designed benchmarks can alleviate the misalignment induced by decentralization • To be realistic, we focus on – Benchmarks that are tradable portfolios and can be matched by the portfolio managers (i.e., no cross-benchmarking) – Benchmarks that do not depend on unknown quantities – Unconditional benchmarks © Michael W. Brandt 2008 5 Optimal Decentralized ALM Motivation Objective of our study • Quantify in an intuitive way the economic cost of decentralization – How much active skill do delegated portfolio managers have to have in order to justify decentralization? • Show how to construct benchmarks that perfectly align objectives and achieve the same outcome as if the investment process was centralized and the CIO had the skills of the portfolio managers – Full benefits of diversification – Optimal mismatch risk – Optimal alpha overlay © Michael W. Brandt 2008 6 Optimal Decentralized ALM Motivation Objective of our study (cont) • Show how to operationalize our approach – Our optimally constructed benchmarks depend on the portfolio managers’ risk tolerances and active skill levels – Three possibilities Take an ex-ante stance on both sets of parameters Construct an empirical cross-sectional distribution and incorporate the resulting “parameter uncertainty” Limit the role of both sets of parameters through constraints Integrated and fully operational approach for decentralized liability driven investment (LDI) management © Michael W. Brandt 2008 7 Optimal Decentralized ALM Problem setup Problem setup Pension fund • CIO – Liabilities Exogenous with Treasury-like dynamics – Assets Centralized portfolio management (7 assets) – Fixed income indices (Aaa, Baa, and Treasuries) – Equity indices (Growth, Intermediate, Value) – Cash Decentralized portfolio management (3 assets) – Fixed income manager – Indices + orthogonal alpha technology – Equities manager – Indices + orthogonal alpha technology – Cash – Preferences = power utility over AT/LT © Michael W. Brandt 2008 9 Optimal Decentralized ALM Problem setup Portfolio managers • Fixed income manager (4 assets) – Indices (Aaa, Baa, Treasuries) – Independent alpha technology – No cash position – Preferences = power utility over A1T/B1T • Equities manager (4 assets) – Indices (Growth, Intermediate, Value) – Independent alpha technology – No cash position – Preferences = power utility over A2T/B2T • Two types of benchmarks (cash or optimally chosen) © Michael W. Brandt 2008 10 Optimal Decentralized ALM Cost of decentralization Cost of decentralization Decomposition Cost of Decentralization = Suboptimal Diversification + Asset/Liability Mismatch Alpha © Michael W. Brandt 2008 12 Optimal Decentralized ALM Cost of decentralization Suboptimal diversification • Cash benchmarks • No alpha technologies • Portfolio managers have relative risk aversion of 10 © Michael W. Brandt 2008 13 Optimal Decentralized ALM Cost of decentralization Centralized ALM • CIO’s optimal allocation to the 6 risky assets max SR weights liability hedging weights • Note – No liability hedging with ° = 1 (log utility) – Full liability immunization as ° ! 1 © Michael W. Brandt 2008 14 Optimal Decentralized ALM Cost of decentralization Centralized ALM (cont) © Michael W. Brandt 2008 15 Optimal Decentralized ALM Cost of decentralization Decentralized ALM with cash benchmarks • Both portfolio managers maximize their (absolute) SR with which includes their alpha technologies (technically ¤ ¤C) • CIO invests optimally in the 2 managed portfolios and cash © Michael W. Brandt 2008 16 Optimal Decentralized ALM Cost of decentralization Decentralized ALM with cash benchmarks (cont) © Michael W. Brandt 2008 17 Optimal Decentralized ALM Cost of decentralization Cost of decentralization • How much alpha do the portfolio managers have to add for the CIO to be indifferent between centralized and decentralized ALM? IR © Michael W. Brandt 2008 18 Optimal Decentralized ALM Optimal benchmarks for delegated ALM Optimal benchmarks for decentralized ALM Optimal benchmarks • Response of the portfolio managers to benchmark with weights ¯i • Note – Benchmarks are ineffective with ° = 1 (log utility) – Tracking error volatility ! 0 as ° ! 1 © Michael W. Brandt 2008 20 Optimal Decentralized ALM Optimal benchmarks for decentralized ALM Optimal benchmarks (cont) • Understanding how portfolio managers respond to benchmarks, the CIO’s optimal benchmark choice is where xiC is the CIO’s optimal allocation to the portfolio manager’s assets including the manager’s alpha technology • These benchmarks induce the first-best solution – Full benefits of diversification – Optimal mismatch risk – Optimal alpha overlay Optimal benchmarks achieve the same outcome as if the investment process was centralized and the CIO had the skills of the portfolio managers © Michael W. Brandt 2008 21 Optimal Decentralized ALM Optimal benchmarks for decentralized ALM Optimal benchmarks (cont) © Michael W. Brandt 2008 22 Optimal Decentralized ALM Optimal benchmarks for decentralized ALM Optimal benchmarks (cont) © Michael W. Brandt 2008 23 Optimal Decentralized ALM Practical implementation Practical implementation Unknown quantities • The optimal benchmarks depend on two unknown quantities – Portfolio managers’ risk tolerance – Portfolio managers’ active skill (IC) • Unknown quantities can be dealt with the same way as they usually are in portfolio choice problems – “Plug-in” = pick values and proceed as if they known – Bayesian = construct a subjective cross-sectional distribution of risk tolerance and active skill levels (be careful, they are likely highly correlated) and then integrate out the unknown quantities © Michael W. Brandt 2008 25 Optimal Decentralized ALM Practical implementation Empirical solution • Looking at past returns on active managers through the lens of a structural model of delegated portfolio management (like ours), we can learn a lot about managers’ risk preferences and skill ) Koijen (2008) • Intuition – Structural models predict how much beta exposure and active risk managers take on as a function of their risk aversion and skill – Beta exposure and active risk can be measured fairly accurately (especially when compared to historical alpha estimates) – These estimates can then be inverted to risk aversion and skill © Michael W. Brandt 2008 26 Optimal Decentralized ALM Practical implementation Empirical solution (cont) • E.g., cross-sectional distribution of relative risk aversion of U.S. mutual fund managers © Michael W. Brandt 2008 27 Optimal Decentralized ALM Extensions and conclusions Extension and conclusions Extensions • Long-only constraints • Risk constraints at the portfolio manager level • Alternative CIO preferences (van Binsbergen and Brandt, 2007) • Other suggestions? © Michael W. Brandt 2008 29 Optimal Decentralized ALM Extension and conclusions Conclusions • Three contributions – Quantify in an intuitive way the economic costs of decentralization – Show how to construct benchmarks that perfectly align objectives and achieve the same outcome as if the investment process was centralized and the CIO had the skills of the portfolio managers – Show how to operationalize our approach • Integrated and fully operational approach for decentralized ALM © Michael W. Brandt 2008 30 Optimal Decentralized ALM