Model Assessment and Adaptive Design Group (MAAD) Carlos M. Carvalho Summary Report 1 Model Assessment and Adaptive Design • General Goal: Enhance, explore and demonstrate the potential of particle based methods to represent sequential model uncertainty. • Specifics... I Explore the ability of SeqMC methods to effectively explore very large discrete model spaces. I Parallel computation issues... connections to Shotgun methods of Jones et. al (2005) and Hans, Dobra and West (2006) I Reinvent chapter 11 of West and Harrison (1997). Model monitoring and intervention. • Models and Applications... I Different specification of the economy... again, DSGE models in Macroeconomics. I Finance models... I Dynamic variable selection I ... 2 Summary of Discussion: Week 1 Three main topics were discussed: 1. Marginal Likelihood Computation via SMC I Many examples in Finance and Economics... I Astrophysics... 2. SMC Exploration of Large Model Spaces I Variable selection problems I DLM model specification 3. Adaptive Design and Control I Astrophysics context I Dynamic control and decision making 3 Five Months Later... 1. Marginal Likelihood Computation via SMC I Astrophysics (Tom Loredo et al.) I Non-parametric regression (Dunson, Das and Shi) 2. SMC Exploration of Large Model Spaces I Variable selection problems (Dunson, Das and Shi) I DLM model specification (Hao, Reeson and Carvalho) 3. Adaptive Design and Control I Dynamic Control and Sequential Decision Making (Argon, Rodriguez and Bain) I Adaptive Sampling (Liu, Li and Dunson) 4 Papers in preparation... 1. “Bayesian Distribution Regression via Augmented Particle Learning” (Dunson and Das, 09) 2. “Particle Stochastic Search for High-dimensional Variable Selection” (Shi and Dunson, 09) 2. “Sequential Learning in Dynamic Graphical Models” (Wang, Reeson and Carvalho, 09) 2. “Adaptive Sampling for Bayesian Variable Selection” (Liu, Li and Dunson, 09) 5 Sequential Learning in Dynamic Graphical Models Hao Wang Craig Reeson Carlos M. Carvalho Application Context High-dimensional asset allocation problems 9 “Structured” models for the time-varying covariance matrix of a large number of assets 9 Tractable models in very many dimensions 2 Gaussian Graphical Models 9 A graph G=(V,E) defines conditional independence relationships for a set of random variables 3 Hyper-Inverse Wishart 9 Conjugate prior (local) distribution for 9 Unique hyper-Markov law with consistent componentmarginals 4 Exchange Rate Example Explore distribution of weights Carvalho, Massam and West (2007) – Biometrika 6 Exchange Rate Example Smaller variability of portfolio weights 7 Portfolios and Graphs Stevens (1998) “… the amount of money invested in asset i depends on the ratio of the expected return that cannot be explained by the linear combination of assets to the nondiversifiable risk.” 8 DLMs and Graphs Conditional independence in multivariate time series models Common components defining each individual DLMs General, fully-conjugate framework 9 Extension to Graphs Closed-form sequential update 9 Forecasting 9 Filtering – retrospective analysis 10 Sequential Update Posterior at time t-1: Prior at time t: 11 Sequential Update One-step forecast: Posterior at time t: 12 Time-Varying Covariance 9 Treat the covariance matrix as a state in the model, allowing it to vary stochastically in time 9 Define an evolution equation -- transition distribution Stochastic map established through multiplicative beta shocks to the diagonal elements of the Cholesky decomposition of the precision 9 Generalizes Uhlig(1994) and extends Quintana(2003) to graphs 13 Time-Varying Covariance “Locally Smooth” – Dynamic Discounting Information loss – discount factor Prior has the same “location” of the previous posterior Implies a random walk in the transition equation of log-variances 14 Sequential Update Posterior at t-1: Prior at t: Posterior at t: Approximate EWMA for the harmonic mean 15 EX Example Revisited 16 EX Example Revisited δ = 0.99 δ = 0.90 17 Dynamic Portfolios (EX) Variance of Portfolio Weights 18 EX Example Revisited 9 Higher realized cumulative returns – structure helps in practical terms 9 Lower risk portfolios in terms of one-step ahead predicted variances 9 Lower volatility of the optimal portfolio weights – more stable portfolios 19 S&P 500 – Model Selection 9 346 stocks in the index from 1999 until 2004 Challenge: uncertainty about the graph. Efficient exploration of model space Metropolis Search Shotgun Stochastic Search Jones, Carvalho et. al. (2005) – Stat. Sci. 20 Marginal Likelihood for Graphs Computations facilitated by one-edge moves Small changes to allow for time dependence 22 S&P 500 Top graph with 29,181 edges 23 Modelling time-varying Gt Recursively conditional sufficient statistic for (Gt , Σt ): St = δSt−1 + yt yt0 Information set Dt = {y1 , · · · , yt } and update Gt as follows i Posterior at time t: p(Gt | Dt ) = p(Gt | St ) ∝ π(Gt )p(Y1:t |Gt , St ) ii Prior at time t + 1 : p(Gt+1 | Dt ) ∝ π(Gt+1 )p(Y1:t |Gt+1 , δSt ) iii Posterior at time t + 1 : p(Gt+1 | Dt+1 ) = p(Gt+1 | St+1 ) ∝ π(Gt+1 )p(Y1:t+1 |Gt+1 St+1 ) Wang, Reeson and Carvalho Sequential Graph DLM Sequential Learning of Gt Particle Learning (Carvalho et al., 2008) + Particle Stochastic Search (Shi & Dunson, 2009) Suppose at time t, most probable graphs (i) {Gt } ∼ p(Gt | Dt ) Re-sample: ∝ p(Y1:t+1 |Gt , St+1 ) Propagate: i Update estimation of edge inclusion probabilities of p(Gt+1 | Dt+1 ) ii Sample new particles using edge inclusion probabilities iii Repeat until “some” stop criteria is satisfied. Wang, Reeson and Carvalho Sequential Graph DLM Exchange Rate Data 1 2 0.8 4 1 2 0.8 4 0.6 6 0.6 6 0.4 8 0.4 8 0.2 10 2 4 6 8 10 0.2 10 0 2 4 6 8 10 1 2 0.8 4 0 1 2 0.8 4 0.6 6 0.6 6 0.4 8 0.4 8 0.2 10 2 4 6 8 10 0 0.2 10 2 4 6 Figure: Sequential Graph DLM Wang, Reeson and Carvalho 8 10 References n Carvalho and West (2007) “Dynamic Matrix-Variate Graphical Models” o Wang, Reeson and Carvalho (2009) “Sequential Learning in Dynamic Graphical Models” p Carvalho and Scott (2009) “Objective Bayesian Model Selection in Gaussian Graphical Models” q Reeson, Carvalho and West (2009) “Financial Time Series Graphical Modeling and Portfolio Analysis” r Quintana, Carvalho, Scott and Costigliola (2009)“Futures Markets, Bayesian Forecasting and Risk Modeling” 26