™ Gaussian Process Volatility Model Neural Information Yue Wu, Jose´ Miguel Hernandez-Lobato,´ and Zoubin Ghahramani Processing Systems Engineering Department, University of Cambridge, United Kingdom. Foundation 1. Introduction 6. Learning with Particle Filters 10. Predicted Volatility Surface Motivation: The prediction of time-changing variances is important when modeling We develop a new algorithm called Regularized Auxiliary Chain Particle filter Log Variance Surface for AUDUSD financial time series. Relevant applications range from the estimation of financial risk to (RAPCF) based on [4]. RAPCF has similar performance to the state-of-the-art method 0.35 portfolio construction and derivative pricing. Particle Gibbs with Ancestor Sampling (PGAS) [5], but RAPCF is much faster. Problem: Standard econometric models, such as GARCH, are limited as they assume fixed, usually linear, functional relationships for the evolution of the variance. 0.3 Solution: We model the unknown functional relationship in our new GP-Vol model using 4 Gaussian Processes. We perform fast online inference by adapting particle filters. 0.25 t 2 2. Characteristics of Financial Volatility 0 0.2 hhhhhh 10 4 output, v Financial returns exhibit: h −2 hh 0.15 Time-dependent standard deviation I −4 or volatility. 2 5 0.1 I Volatility clustering. 0 4 0 0 2 I Asymmetric effect on volatility from h 0 positive and negative returns. −5 −2 hhh input, x input, v t−1 t−1 −10h Jul07 Jan10 Surface generated by plotting the mean predicted outputs vt against a grid of inputs for vt−1 and xt−1. Cross section v vs v t t−1 Cross section v vs x 3. Standard Models 7. Experiment Setup t t−1 GARCH is the the best known and most popular volatility model [1]: Data: 4 q p 2 2 2 X 2 X 2 I 50 time series, consisting of 20 daily FX and 30 daily Equity returns. 3 xt ∼ N (0; σt ) and σt = α0 + αjxt−j + βiσt−i : (1) I Each time series contains 780 observations, from Jan 2008 - Jan 2011. j=1 i=1 2 I Each time series was normalized to have zero mean and unit standard deviation. GARCH can model a) time-dependence and b) volatility clustering, but does not account 1.5 1 for the asymmetric effect of positive and negative returns on volatility. GARCH variants, Model comparison: t 0 such as EGARCH [2] and GJR-GARCH [3], do account for the asymmetric effect, but t v The models compared were GP-Vol, GARCH(1,1), EGARCH(1,1) and GJR-GARCH(1,1). v they are all still limited by assuming a linear transition function for the volatility. I 1 −1 I Each model receives an initial time series of length 100. I The models are trained on the training data and a one-step forward prediction is made. −2 4. Gaussian Process Volatility Model 0.5 I Then the training data is augmented with a new observation and the process is repeated. −3 Our contribution is the new GP-Vol model: I For GP-Vol, an RAPCF with N = 200 and λ = :95 was used −4 2 2 2 xt ∼ N (0; σt ) and vt := log(σt ) = f (vt−1; xt−1) + t ; where t ∼ N (0; σn) : (2) 0 −5 We place a Gaussian Process (GP) prior on the the transition function f . 8. Comparison of Model Predictive Performance −6 −4 −2 0 2 4 6 −6 −4 −2 0 2 4 6 v x Advantages of GP-Vol: We perform a multiple comparison test, where all the methods are ranked according to t−1 t−1 their performance on the 50 time series or tasks. The average ranks are then tested for I GP-Vol Models the unknown transition function in a non-parametric manner. Left, predicted vt ± 2 s.d. for inputs (0; xt−1). Right, predicted vt ± 2 s.d. for inputs statistical significant differences. GP-Vol is the best model with significant evidence. I GP-Vol reduces the risk of overfitting by following a full Bayesian approach. (0; xt−1). Nemenyi Test 5. GP-Vol Graphical Model 11. References CD [1] T. Bollerslev. Generalized autoregressive conditional heteroskedasticity. GJR GARCH Journal of econometrics, 31(3):307–327, 1986. GP−VOL EGARCH I GARCH, EGARCH and GJR-GARCH are [2] D.B. Nelson.Conditional heteroskedasticity in asset returns: A new approach. all Hidden Markov models (HMM). The Econometrica, 59(2):347–370, 1991. dynamic variances are the hidden states. 1 2 3 4 [3] L.R. Glosten, R. Jagannathan, and D.E. Runkle. On the relation between the expected value and the volatility of the nominal excess return on stocks. GP-Vol is a Gaussian Process state space The Journal of Finance, 48(5):1779–1801, 1993. I 9. RAPCF vs. PGAS model (GP-SSM), a generalization of HMM [4] J. Liu and M. West. Combined parameter and state estimation in simulation-based filtering. Institute of Statistics and Decision Sciences, Duke University, 1999. in which the transition function is unknown RAPCF has similar predictive performance as the Method Configuration Avg. Time and represented by a GP. state-of-the-art method PGAS (see paper), but RAPCF N = 200 6 min. [5] F. Lindsten, M. Jordan, and T. Schon.¨ Ancestor Sampling for Particle Gibbs. RAPCF but is much faster. PGAS N = 10, M = 100 732 min. In Advances in Neural Information Processing Systems 25, pages 2600–2608, 2012. http://jhml.org/ [email protected].