On geometric ergodicity of CHARME models
Summary (1 min read)
1 Introduction
- In particular, the important property of geometric ergodicity is obtained under some conditions.
- In practice, it is often not realistic to assume that the observed process has the same trend function m and the same volatility function σ at each time instant.
- In particular, the case p = 1 is interesting on its own.
2 First conditions for geometric ergodicity of CHARME
- The authors focus on their CHARME model (1.2) and make the following assumptions A. 1.
- The process {Qt} with values on {1, · · · ,K} is a first order strictly stationary Markov chain which is irreducible and aperiodic with probability distribution (π1, · · · , πK) and transition probability matrix A = (aij)1≤i,j≤K .
4 Concluding remarks
- The authors have considered Conditional Heteroscedastic Autoregressive Mixture of Experts models, a form of hidden Markov model with nonlinear autoregressive-ARCH components, which can be useful in, e.g. financial econometrics.
- Notice that their conditions for the mk and σk are easier to verify than some kind of sub linearity conditions or Lyapounov exponents conditions.
- Subsequent work based on their results can lead to sufficient conditions for the existence of moments for Xt or for the existence of limit theorems for Xt.
- This practical aspect is the subject of forthcoming publication.
Did you find this useful? Give us your feedback
Citations
163 citations
54 citations
Cites background or methods from "On geometric ergodicity of CHARME m..."
...However, here we shall focus on time series whose correlation structure changes slowly over time (early work on time-varying time series include Priestley (1965), Subba Rao (1970) and Hallin (1984))....
[...]
...…including, under certain assumptions on the innovations, the vector AR models (see Pham and Tran (1985)) and other Markov models which are irreducible (cf. (Feigin & Tweedie, 1985), Mokkadem (1990), Meyn and Tweedie (1993), Bousamma (1998), Franke, Stockis, and Tadjuidje-Kamgaing (2010))....
[...]
...However, here we shall focus on time series whose correlation structure changes slowly over time (early work on time-varying time series include Priestley (1965), Subba Rao (1970) and Hallin (1984)). As in nonparametric regression and other work on nonparametric statistics we use the rescaling device to develop the asymptotic theory. The same rescaling device has been used for example in nonparametric time series by Robinson (1989) and by Dahlhaus (1997) in his definition of local stationarity....
[...]
36 citations
Cites background or methods from "On geometric ergodicity of CHARME m..."
...Stockis et al. (2010) use the time series in eqn (4) as building blocks in a regime-switching model, the so called CHARME-models, in the context of financial time series....
[...]
...The idea behind this is that the test would only be applied in situations were this is the case and the approximation by a neural network based process is reasonable....
[...]
...Since Model (4) is a special case with only one regime, the following result is a straightforward application of Theorem 4 of Stockis et al. (2010)....
[...]
...In view of possible financial applications, an indirect example of model (1) is given by the nonlinear ARCH processes, see, for example, Stockis et al. (2010), an extension of the b-ARCH model introduced by Guégan and Diebolt (1994) which include the celebrated ARCH model defined by Engle (1982)....
[...]
...AMS subject classification 2000: 62G10, 62M45, 62G08...
[...]
35 citations
31 citations
References
5,931 citations
"On geometric ergodicity of CHARME m..." refers background or methods in this paper
...On proving Theorems 1 and 2, we have used the drift criterion of Meyn and Tweedie (1993)....
[...]
...Proof:As in the case p = 1 we are going to prove that the conditions of Theorem 15.0.1,(iii) of Meyn and Tweedie [ 7 ] pp 354 355 are satisfied....
[...]
...Proof: We are going to prove that the conditions of Theorem 15.0.1,(iii) of Meyn and Tweedie [ 7 ], pp 354 355, are satisfied....
[...]
4,338 citations
Additional excerpts
...[12], or Jiang and Tanner [10]....
[...]
3,636 citations
"On geometric ergodicity of CHARME m..." refers background in this paper
...This is, e.g. indicated by estimates of the tail index of the marginal distribution of certain asset returns (compare, e.g. Embrechts et al., 1997)....
[...]
...Tong, 1983) or qualitative threshold ARCH models (Gouriéroux and Montfort, 1992). The latter are in particular of interest, as they also contain a switching mechanisms which, however, is solely driven by the current states of the time series under consideration. By introducing the hidden Markov chain Qt, we additionally allow for sudden changes of the state of the process driven by external forces. For many applications, Qt will change its value only rarely, i.e. the observed process follows the same regime for quite some time before a change occurs. An example would, e.g. be time series of asset returns which change their behaviour if the market moves from a phase with increasing to one with decreasing trend or from one with generally high to one with low volatility. Tadjuidje Kamgaing (2005) gives some illustrations of such models to financial data in the context of asset management and risk analysis where the state functions are estimated by neural networks. Another noneconomic example would be EEG time series from sleeping people who during one night’s sleep move through different stationary phases where, compared with the timescales, the changes are almost sudden and can be modelled as jumps (compare Liehr et al., 1999). We remark that some of the single phases of a CHARME model may also correspond to explosive regimes provided that this states are visited rarely enough. This feature is well known for several parametric mixture models, e.g. for the flexible coefficient GARCH model of Medeiros and Veiga (2008). Independent of the type of estimates considered, a crucial condition for developing a theory of estimation and testing in the setting of CHARME is the existence of a stochastic process satisfying eqn (2) which is geometrically ergodic....
[...]
...Tong, 1983) or qualitative threshold ARCH models (Gouriéroux and Montfort, 1992). The latter are in particular of interest, as they also contain a switching mechanisms which, however, is solely driven by the current states of the time series under consideration. By introducing the hidden Markov chain Qt, we additionally allow for sudden changes of the state of the process driven by external forces. For many applications, Qt will change its value only rarely, i.e. the observed process follows the same regime for quite some time before a change occurs. An example would, e.g. be time series of asset returns which change their behaviour if the market moves from a phase with increasing to one with decreasing trend or from one with generally high to one with low volatility. Tadjuidje Kamgaing (2005) gives some illustrations of such models to financial data in the context of asset management and risk analysis where the state functions are estimated by neural networks....
[...]
3,414 citations
"On geometric ergodicity of CHARME m..." refers background in this paper
...So, to apply the drift criterion, we need to find a function g(f) > 1, b > 0 and M > 0 such that E gðftÞ ft 1 ¼ sl x g sl x g sl x b for kft 1k > M Let gðftÞ ¼ 1þ X2t : Then, E gðftÞ ft 1 ¼ sl x g sl x g sl x ¼ Pkðm2kðxÞ þ r2kðxÞÞEðStk j St 1 ¼ slÞ x2 1þ x2 P kðm2kðxÞ þ r2kðxÞÞalk x2 1 and the conclusion is obtained by Assumption 8. h Some financial time series are extremely heavy tailed, and even the adequacy of models assuming a finite variance is doubtful....
[...]
...In a similar manner, consistency of the parameter estimates in the pure nonparametric ARCH mixture model Xt ¼ XK k¼1 Stk rkðXt 1; . . . ; Xt p; hkÞ t can be shown, where t now are i.i.d. random variables with mean 0 and variance 1....
[...]
...2010, 31 141–152 1 4 4 E gðftÞ ft 1 ¼ sl x g sl x g sl x PkðjmkðxÞja þ rakðxÞEj tjaÞEðjStkja j St 1 ¼ slÞ jxja 1þ jxja P k alkðjmkðxÞj a þ rakðxÞEj tj aÞ jxja 1 and we conclude the proof by using Assumption 10. h REMARK 1....
[...]
...Stk is unknown; so, a direct way of getting the estimates is not possible....
[...]
...For illustration of how the previous results may be applied, we consider a special case of the model defined in equation (2), Xt ¼ XK k¼1 StkmkðXt 1; . . . ; Xt p; hkÞ þ r t ð5Þ where Stk is defined as previously, r>0 is fixed, and, for x 2 Rp mkðx; hkÞ ¼ m0k þ XH h¼1 mhkwð< ahk; x > þbhkÞ; k ¼ 1; . . . ; K ; ð6Þ is the output function of a one-layer feedforward neural network with Hk hidden neurons, <Æ> is the scalar product on R p and w the activation function, e.g. the logistic function....
[...]
Related Papers (5)
Frequently Asked Questions (11)
Q2. What is the main purpose of CHARME?
CHARME is quite useful for modeling time series data which are piecewise stationary such that their dynamics switch sometimes from one state to another.
Q3. what is the t if e is a unit vector with the kth?
Then A1 contains at least one integer between 1 and K and it is enough to prove that there exists t such thatP (( St+1 Xt+1 ) ∈ {e} ×A2 |S1 = sl, X1 = x ) > 0with e a unit vector with the the kth component equal 1 and sl a unit vector with the lth component equal 1.
Q4. what is the l-th component of x?
Let x = (xt, · · · , xt−p+1)′ be a vector of real numbers and sl a K-dimensional unit vector with the l-th component equal 1 and considerE(g(ζt+1) | (Xt, · · · , Xt−p+1)′ = x, St = sl)= 1 + K∑k=1alk(m2k(x) + σ 2 k(x)) + bp−1x 2 t + · · ·+ b1x2t−p+2 (3.1)Now, let us focus onK∑ k=1 alk(m2k(x) + σ 2 k(x))=
Q5. What is the proof of the central limit theorem?
If the process {Xt} is also strictly stationary it is well known that this implies that {Xt} is absolutely regular with geometric decreasing rate, which gives a very useful condition for deriving limit theorems like the central limit theorem.
Q6. What is the simplest way to prove that t is irreducible?
0and doing so iteratively, the authors obtainP (( St+1 Xt+1 ) ∈ {e} ×A2 |S1 = sl, X1 = x ) =K∑ j,··· ,jt−1 alj1 · · · ajt−1kbj,··· ,jt−1(x)which is strictly greater than 0 for some t because of the irreducibility of {Qt} and the fact that bj,··· ,jt−1(x) > 0.• Analogously it can easily be seen that {ζt} is aperiodic.•
Q7. What is the main topic of this paper?
In this paper the authors are analyzing the so-called Conditional Heteroscedastic Autoregressive Mixture of Experts, henceforth CHARME, models.
Q8. What is the simplest way to explain the process Qt?
{St = (St1, · · · , StK)′} inherits the properties of {Qt}.A. 2 Let Gt−1 = σ{Xr, r ≤ t−1} be the σ-algebra generated by {Xr, r ≤ t−1} and Gt−1 any event in Gt−1.
Q9. What is the proof of the theorem 4?
Geometric ergodicity of ζt can clearly be obtained even if some of the underlying dynamics taken on their own are not geometric ergodic or even stationary, provided the probability to go from a stable dynamic to a non stable dynamic is low enough and the probability to move from a non stable dynamic to a stable dynamic is large enough.
Q10. what is the simplest way to prove the existence of a petite set?
In particular in their case where {φt} is a Markov chain, it is enough to prove the existence of a petite set C, a function g ≥ 1 and constants 0 < r < 1 and B > 0 such thatE(g(φt) |φt−1 = x) < rg(x), x 6∈ C E(g(φt) |φt−1 = x) < B, x ∈ C.
Q11. What is the main difference between the two types of CHARME models?
Independently of the type of estimates considered, a crucial condition for developing a theory for estimation and testing in the setting of CHARME is the existence of a stochastic process satisfying (1.2) which is geometric ergodic.