1 / 29

The Continuous Limit of Markov Switching GARCH Carol Alexander & Emese Lazar, ICMA Centre

The Continuous Limit of Markov Switching GARCH Carol Alexander & Emese Lazar, ICMA Centre. Workshop on Statistical Modelling of Complex Systems M ünchen , 2006. Outline. Introduction Motivation Literature on the continuous limit of GARCH ( Nelson , Corradi)

hesper
Download Presentation

The Continuous Limit of Markov Switching GARCH Carol Alexander & Emese Lazar, ICMA Centre

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. The Continuous Limit ofMarkov Switching GARCHCarol Alexander & Emese Lazar, ICMA Centre Workshop on Statistical Modelling of Complex Systems München, 2006 Carol Alexander & Emese Lazar, ICMA Centre, 2006

  2. Outline • Introduction • Motivation • Literature on the continuous limit of GARCH (Nelson, Corradi) • Continuous time limit of weak GARCH • Weak MS GARCH • Continuous limit of weak MS GARCH • Conclusions Carol Alexander & Emese Lazar, ICMA Centre, 2006

  3. Introduction Normal GARCH(1,1)(Engle 1982, Bollerslev 1986): ht = the conditional variance of the residuals Carol Alexander & Emese Lazar, ICMA Centre, 2006

  4. Motivation for the continuous time limit of GARCH • Many advantages to link discrete time modelling to continuous time modelling (GARCH option pricing) • We are interested in a stronger relationship between the two types of models (DT and CT) – an equivalence: • The limit of the DT model should be the CT model • Discretizing the CT model should give the DT model • Time aggregation is important because: if a model is not aggregating in time → then a different model will be valid for each frequency → but discretizing the CT model will give the same model for each frequency → contradiction Carol Alexander & Emese Lazar, ICMA Centre, 2006

  5. Continuous limit of GARCH  freedom to set the assumptions! Notation: Δω, Δα, Δβfor parameters using returns with step-length Δ - the returns and the conditional variance (normalized): Carol Alexander & Emese Lazar, ICMA Centre, 2006

  6. Convergence theorems – which one? • Nelson (1990): Continuous-time limit: stochastic variance process with independent BM • Corradi (2000): Continuous-time limit: deterministic variance process Carol Alexander & Emese Lazar, ICMA Centre, 2006

  7. Diffusion in variance? YES: • The variance of the variance is nonzero in discrete time, it should be the same in continuous time • Results from simulations • More intuitive NO: • GARCH has only one source of uncertainty in discrete time • The Euler discretization of Nelson’s limit model does not return the original GARCH model (but Taylor’s ARV model) • Nelson’s model cannot be extended to multi-state GARCH • GARCH and GARCH diffusion have non-equivalent likelihood functions (Wang, 2002) • Other papers: Duan et al. (2005), Brown et al (2002) Carol Alexander & Emese Lazar, ICMA Centre, 2006

  8. Simulation results Carol Alexander & Emese Lazar, ICMA Centre, 2006

  9. Weak GARCH Drost and Nijman (1993): ht = the best linear predictor (BLP) of the squared residuals Weak GARCH is aggregating in time! Carol Alexander & Emese Lazar, ICMA Centre, 2006

  10. Simulated volatility smiles Volatility smiles generated by the weak GARCH, assuming θ = 0.05; ω = 0.0045; α = 0.1; μ(t) = 0; e(t) = 0; S0 = 100; V0 =0.09; T – t = 1; r = 0% 100 steps and 100,000 runs were used for the simulations (a) τ(t) = 0; η(t) = 3 (b) τ(t) = -1; η(t) = 5 (c) τ(t) = -1.5; η(t) = 10 Carol Alexander & Emese Lazar, ICMA Centre, 2006

  11. Weak GARCH – inverse relationship • Drost and Nijman (1993) derive the low frequency parameters (ω, α and β) from the high frequency parameters • We inverse this relationship by deriving the high frequency parameters from the low frequency ones (also in Drost and Werker, 1996) • Then we compute the limit of the parameters as the time step converges to zero Carol Alexander & Emese Lazar, ICMA Centre, 2006

  12. Weak GARCH – assumptions • There is no distributional assumption about the error process, having a general function for the unconditional kurtosis • Assumption 1: the kurtosis has a finite limit • Assumption 2: the BLP of the squared residuals converges at rate Δ to the conditional variance Carol Alexander & Emese Lazar, ICMA Centre, 2006

  13. Conditional moments The conditional moments (normalized) are: Carol Alexander & Emese Lazar, ICMA Centre, 2006

  14. Limit of weak GARCH –results Convergence rates for the parameters (proved): → also see Drost and Werker (1996) Continuous time model: Carol Alexander & Emese Lazar, ICMA Centre, 2006

  15. Discretization - idea (in progress) → Time aggregation property is kept! Carol Alexander & Emese Lazar, ICMA Centre, 2006

  16. Discretization - idea • The independent Brownian motion in the variance process is discretized as follows: • This is not normally distributed (chi-square instead), but it satisfies the basic properties required: Carol Alexander & Emese Lazar, ICMA Centre, 2006

  17. Multi-state GARCH • They generally offer a better fit than single-state models, they can model time varying skewness and kurtosis • NM GARCH is a simplified version of MS GARCH where the transition matrix has rank 1 • However, NM GARCH does not have a continuous time limit (it is not a Lévy process), because it does not satisfy the stochastic continuity condition (the number of variance switches during a fixed period of time converges to infinity when the time step converges to zero – this is because the conditional probability of a switch does not change when the frequency changes): Carol Alexander & Emese Lazar, ICMA Centre, 2006

  18. Normal mixture distribution • can model skewness & excess kurtosis; analytically tractable • Interpretation: • The normals give the different types of traders in the market • The normals give the different types of expectations in the market Carol Alexander & Emese Lazar, ICMA Centre, 2006

  19. Markov Switching 0.9 0.8 State 1 small step large step 0.66 0.1 0.4 0.2 0.2 State 2 0.33 0.8 0.6 Carol Alexander & Emese Lazar, ICMA Centre, 2006

  20. Markov switching GARCH • State indicator vector: • State probabilities = conditional expectations: • Transition matrix = help to form expectations: • We have: Carol Alexander & Emese Lazar, ICMA Centre, 2006

  21. MS GARCH: 3 types • Cai (1994), Hamilon & Susmel (1994): Realized variance - leads to a non-recombining tree for the volatility process - estimation problems • Gray (1996), Klaassen (2002): Weighted average - analytically not too tractable; not intuitive • Haas, Mittnik & Paolella (2004): Its own lagged value - tractable - the variance series allowed to be more free from each other Carol Alexander & Emese Lazar, ICMA Centre, 2006

  22. Realized volatility (simulated) Carol Alexander & Emese Lazar, ICMA Centre, 2006

  23. Weak MS GARCH Based on Haas, Mittnik & Paolella (2004) and Drost and Nijman (1993): Carol Alexander & Emese Lazar, ICMA Centre, 2006

  24. Weak MS GARCH • The GARCH component of the model is aggregating in time • Still, the model is not aggregating because multi-state models do not aggregate in time • Explanation: if for a time step Δ there are two states, then for a time step of length 2 Δ there would be 3 or 4 states • Another problem: Nelson’s assumptions cannot be generalized for this framework, so convergence rates similar to Corradi’s are used Carol Alexander & Emese Lazar, ICMA Centre, 2006

  25. Weak MS GARCH Convergence rates for the parameters (proved): To achieve convergence, we must assume: Transition rate matrix: Carol Alexander & Emese Lazar, ICMA Centre, 2006

  26. Limit of weak MS GARCH The limit model is: Carol Alexander & Emese Lazar, ICMA Centre, 2006

  27. Properties • The framework can be extended to include the leverage effect similar to the AGARCH model of Engle and Ng (1993) • The state uncertainty for an option can be hedged using another derivative (only in a two-state world) • The market is arbitrage free, but incomplete • The discretization, however, would not return the original model Carol Alexander & Emese Lazar, ICMA Centre, 2006

  28. Conclusions • Used the weak definition of GARCH to derive its ‘equivalent’ continuous time limit; the limit model has correlated brownians • Showed that there is no freedom to choose the convergence rates for the parameters; the weak definition implies the convergence rates • Defined the weak MS GARCH; for this only the GARCH component is aggregating in time and the relationship between the parameters for different frequencies is derived • Its continuous limit is very flexible, but it cannot (?) be discretized to return the original model • NM GARCH models do not have a continuous time limit Carol Alexander & Emese Lazar, ICMA Centre, 2006

  29. Further questions • Are the assumptions realistic? Can they be proved? • Applications  GARCH option pricing • What to do with the discretization problems? Carol Alexander & Emese Lazar, ICMA Centre, 2006

More Related