Abstract
Advanced computing and processing techniques have yielded abundant information for financial time series forecasting. It is, therefore, natural to ask for possible extensions of time series models to accommodate the wealth of information. In this article, we develop a new model for financial volatility estimation and forecasting by incorporating exogenous covariates in a semi-parametric log-GARCH model. With additional information, we gain an increased prediction power. We propose a quasi-maximum likelihood procedure via spline smoothing technique. Consistent estimators and asymptotic normality are obtained under mild regularity conditions. Simulation experiments provide strong evidence that corroborates the asymptotic theories. Additionally, an application to SPY index data demonstrates strong competitive advantage of our model comparing with GARCH(1,1) and log-GARCH(1,1) models.
Similar content being viewed by others
References
Andersen TG, Bollerslev T (1998) Answering the skeptics: yes, standard volatility models do provide accurate forecasts. Int Econ Rev 39(4):885–905
Baker SR, Bloom N, Davis SJ (2013) Measuring Economic Policy Uncertainty. Chicago Booth Research Paper 13–02, Stanford University, Department of Economics
Bollerslev T (1986) Generalized autoregressive conditional heteroskedasticity. J Econom 31(3):307–327
Bollerslev T, Engle RF, Nelson DB (1994) ARCH models. In: Handbook of Econometrics, vol 4, Elsevier
de Boor C (2001) A practical guide to splines. Springer, Berlin
Bühlmann P, McNeil AJ (2002) An algorithm for nonparametric GARCH modelling. Comput Stat Data Anal 40(4):665–683
Cai J (1994) A Markov model of unconditional variance in ARCH. J Bus Econ Stat 12(3):309–316
Carroll RJ, Härdle W, Mammen E (2002) Estimation in an additive model when the components are linked parametrically. Econom Theory 18(4):886–912
Chen NF, Roll R, Ross SA (1986) Economic forces and the stock market. J Bus 59(3):383–403
Ciner C (2001) Energy shocks and financial markets: nonlinear linkages. Stud Nonlinear Dyn E 5(3):203–212
Čížek P, Spokoiny V (2009) Varying coefficient GARCH models. In: Mikosch T, Krei JP, Davis RA, Andersen TG (eds) Handbook of financial time series. Springer, Berlin Heidelberg, pp 169–185
Cutler DM, Poterba JM, Summers LH (1991) Speculative dynamics. Rev Econ Stud 58(3):529–546
Engle RF (1982) Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation. Econometrica 50(4):987–1007
Fan J, Qi L, Xiu D (2013) Quasi maximum likelihood estimation of GARCH models with heavy-tailed likelihoods. Working Paper Series
Francq C, Zakoïan JM (2010) GARCH models: structure, statistical inference and financial applications. Wiley, Chichester
Francq C, Wintenberger O, Zakoïan JM (2013) Garch models without positivity constraints: exponential or log garch? J Econom 177(1):34–46
Gogineni S (2008) The stock market reaction to oil price changes. Division of Finance, Michael F Price College of Business, University of Oklahoma, Norman
Gray SF (1996) Modeling the conditional distribution of interest rates as a regime-switching process. J Financ Econ 42(1):27–62
Hamilton JD, Susmel R (1994) Autoregressive conditional heteroskedasticity and changes in regime. J Econom 64(1):307–333
Han H, Kristensen D (2014) Asymptotic theory of the qmle for garch-x models with stationary and non-stationary covariates. J Bus Econ Stat 32:416–429
Han H, Park JY (2008) Time series properties of arch processes with persistent covariates. J Econom 146:275–292
Han H, Park JY (2012) Arch/garch with persistent covariate: asymptotic theory of mle. J Econom 167:95–112
Hansen PR, Lunde A (2005) A forecast comparison of volatility models: does anything beat a GARCH (1, 1)? J Appl Econom 20(7):873–889
Hastie TJ, Tibshirani RJ (1990) Generalized additive models. Chapman and Hall, London
Huang JZ, Yang L (2004) Identification of non-linear additive autoregressive models. J R Stat Soc Ser B 66(2):463–477
Huang R, Masulis R, Stoll H (1996) Energy shocks and financial markets. J Futures Markets 16(1):1–27
Iglesias EM (2009) Finite sample theory of qmles in arch models with an exogenous variable in the conditional variance equation. Stud Nonlinear Dyn E 13(2):1–28
Jensen ST, Rahbek A (2004a) Asymptotic inference for nonstationary GARCH. Econom Theory 20(6):1203–1226
Jensen ST, Rahbek A (2004b) Asymptotic normality of the QMLE estimator of ARCH in the nonstationary case. Econometrica 72(2):641–646
Jones CM, Kaul G (1996) Oil and the stock markets. J Financ 51(2):463–491
Klaassen F (2002) Improving GARCH volatility forecasts with regime-switching GARCH. Emp Econ 27(2):363–394
Lee SW, Hansen BE (1994) Asymptotic theory for the GARCH (1, 1) quasi-maximum likelihood estimator. Econom Theory 10:29–52
Sadorsky P (1999) Oil price shocks and stock market activity. Energy Econ 21(5):449–469
Stone CJ (1985) Additive regression and other nonparametric models. Ann Stat 13(2):689–705
Wang L, Yang L (2007) Spline-backfitted kernel smoothing of nonlinear additive autoregression model. Ann Stat 35(6):2474–2503
Wang L, Liu X, Liang H, Carroll RJ (2011) Estimation and variable selection for generalized additive partial linear models. Ann Stat 39(4):1827–1851
Wang L, Feng C, Song Q, Yang L (2012) Efficient semiparametric GARCH modeling of financial volatility. Stat Sin 22:249–270
Yang L (2006) A semiparametric GARCH model for foreign exchange volatility. J Econom 130(2):365–384
Yang L, Härdle W, Nielsen J (1999) Nonparametric autoregression with multiplicative volatility and additive mean. J Time Ser Anal 20(5):579–604
Zhang S, Han H (2014) Semiparametric arch-x model for leverage effect and long memory in stock return volatility. J Econ Theory Econom 25:81–100
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
11749_2015_442_MOESM1_ESM.pdf
SUPPLEMENTARY MATERIAL Supplement to “Semiparametric Estimation and Forecasting for Exogenous Log-GARCH Models“ The supplementary material contains detailed proofs of Lemmas 2 to 5, 7 to 10 stated in Appendix. 155kb
Appendix: Proof of the theorems
Appendix: Proof of the theorems
Throughout the section, let \(\Vert \cdot \Vert \) be the Euclidean norm and supremum norm \(||\phi ||_{\infty }=\,\) sup\(_{x\in [0,1]}|\phi (x)|\). For any matrix \({\mathbf {A}}\), denote its spectral norm as \(\left\| {\mathbf {A}}\right\| _{2}\) \(=\) \(\mathrm{sup}_{\mathbf {x}\ne \mathbf{0}}\frac{ \left\| {\mathbf {Ax}}\right\| }{\left\| {\mathbf {x}}\right\| }\). For any vector \({\mathbf {V}}\), define \(\Vert {\mathbf {V}} \Vert _{\infty }=\,\) max\((\vert v_{1}\vert ,\ldots , \vert v_{n}\vert )\), and for any measurable function \(\phi \), define \(||\phi ||_{2}^{2}=E\phi ^2({\mathbf {X}}).\) Throughout the section, \(\kappa , \kappa _1\) and \(\kappa _2\) will be used as constants in a generic way.
Proof of Lemma 1
Proof
Recursively using the expression in (3), we have
Under Assumption (A2), the first term on the right-hand side converges almost surely since
Similarly, the second term converges almost surely under the same assumption. Therefore, \(\log \sigma _{t}^{2}\) converges almost surely. \({\mathbf {Z}}_{t}\) is stationary and ergodic, then \( \log \sigma _{t}^{2}\), as a function of an ergodic process is stationary and ergodic.
In the following we list Lemmas 2, 3 and 4 to verify the regularity conditions for maximum likelihood function in (5).
Lemma 2
Under Assumptions (A2) and (A4), \(E_{{\varvec{\gamma }}_{0}}\left\| \frac{\partial l_{t}({\varvec{\gamma }}_{0})}{\partial {\varvec{\gamma }}}\right\| <\infty \) , \(E_{{\varvec{\gamma }}_{0}}\left( \frac{\partial l_{t}({\varvec{\gamma }}_{0})}{ \partial {\varvec{\gamma }}}\right) =0, E_{{\varvec{\gamma }}_{0}}\left\| \frac{\partial ^{2}l_{t}({\varvec{\gamma }}_{0})}{\partial {\varvec{\gamma }} \partial {\varvec{\gamma }}}\right\| <\infty \).
The proof of 2 is provided in the supplementary material to this article. The proofs of Lemmas 3 to 5, 7 to 10 below are also provided in the supplementary material.
Lemma 3
Let \({\mathbf {J}}=E_{{\varvec{\gamma }}_{0}}\left( \frac{\partial ^{2}l_{t}( {\varvec{\gamma }}_{0})}{\partial {\varvec{\gamma }}\partial {\varvec{\gamma }}} \right) \), then \({\mathbf {J}}\) is invertible.
Lemma 4
Under Assumptions (A2) and (A4), we have
here \(V({\varvec{\gamma }} _{0})=\left\{ {\varvec{\gamma }}:\exists \ \xi > 0, \left\| {\varvec{\gamma }}-{\varvec{\gamma }} _{0}\right\| _{\infty }<\xi ,{\varvec{\gamma }}\in {\varTheta }_1 \right\} \).
Next we consider an approximation to the likelihood function in (5). Define \(\breve{l}_{t}=\log \breve{\sigma }_{t}^{2}+y_{t}^{2}\breve{\sigma }_{t}^{-2}\), where
Lemma 5
Under Assumptions (A2) and (A4),
-
1.
\(\lim _{T\rightarrow \infty }\sup _{{\varvec{\gamma }}\in {\varTheta }_{1}}1/T\sum _{t=1}^{\infty }\left| l_{t}({\varvec{\gamma }})-\breve{l}_{t}({\varvec{\gamma }})\right| =0,a.s\), and \(E_{{\varvec{\gamma }}_{0}}\left| l_{t}({\varvec{\gamma }}_{0})\right| <\infty .\)
-
2.
If there exists \(t \in {\mathcal {Z}}\) such that \(\log \sigma ^{2}_{t}({\varvec{\gamma }})=\log {\sigma }_{t}^{2}({\varvec{\gamma }}_{0}),a.s.\), then we have \({\varvec{\gamma }}={\varvec{\gamma }}_{0}\).
Following Theorem 7.1 in Francq and Zakoïan (2010) and applying Lemma 5, we immediately obtain the following lemma.
Lemma 6
Under Assumptions (A2) and (A4), as \(T\) goes to infinity, \( {\varvec{\breve{\gamma }}}\rightarrow {\varvec{\gamma }}_{0}, \ a.s.\)
Lemma 7
Under Assumptions (A2) and (A4),
Lemma 8
Under Assumptions (A2) and (A4), as \(T\rightarrow \infty \),
where \(\alpha \) is \(E_{{\varvec{\gamma }}_{0}}\left( 1-\epsilon _{t}^{2}\right) ^2\) and \(\mathbf {J} =E_{{\varvec{\gamma }}_{0}}\left( \frac{\partial ^{2}l_{1}( {\varvec{\gamma }}_{0})}{\partial {\varvec{\gamma }}\partial {\varvec{\gamma }}^{\mathrm{\tiny T}}} \right) \).
Now we consider the likelihood function with \(\eta \) approximated by \(\tilde{ \eta }\) in (10). Define \(\tilde{l}_{t}=\log \tilde{\sigma } _{t}^{2}+\frac{y_{t}^{2}}{\tilde{\sigma }_{t}^{2}}\), where \( \log \tilde{\sigma }_{t}^{2}=\frac{1-a^{t}}{1-a}c+\sum _{j=1}^{t}a^{j-1}\left( {\mathbf {Z}}_{t-j}^{1{\mathrm{\tiny T}}}{\varvec{\beta }}+\tilde{\eta }\left( {\mathbf {Z}}_{t-j}^{2}\right) \right) \log y_{t-j}^{2}. \) Then define
Lemma 9
Under Assumptions (A1)–(A5), as \(T\) goes to infinity, \(\tilde{{\varvec{\gamma }}}\rightarrow {\varvec{\gamma }}_{0},\ a.s.\)
Lemma 10
Under Assumptions (A1)–(A5),
Combining Lemmas 6, 7, 8, 9, 10 with Slutsky’s Lemma, we immediately have
Lemma 11
Under Assumptions (A1)–(A5),
where \(\alpha \) is \(E_{{\varvec{\gamma }}_{0}}\left( 1-\epsilon _{t}^{2}\right) ^2\) and \(\mathbf {J} =E_{{\varvec{\gamma }}_{0}}\left( \frac{\partial ^{2}l_{1}( {\varvec{\gamma }}_{0})}{\partial {\varvec{\gamma }}\partial {\varvec{\gamma }}^{\mathrm{\tiny T}}} \right) \).
Next, define \(\hat{l}_{t}=\log \hat{\sigma }_{t}^{2}+y^{2}\hat{\sigma }_{t}^{-2}\) with
Lemma 12
Under Assumptions (A1)–(A5),
Proof
Define \(\log \breve{\hat{\sigma }}^{2}_{t}({\varvec{\gamma }},{\varvec{\lambda }})= \frac{c}{1-a}+\sum _{j=1}^{\infty }a^{j-1}\left( {\mathbf {Z}}_{t-j}^{1} \mathbf {\beta }+\mathbf {B}({\mathbf {Z}}_{t-j}^{2}){\varvec{\lambda }}\right) \log y_{t-j}^{2},\) and \(\breve{\hat{l}}_{t}= \log \breve{\hat{\sigma }}^{2}_{t} + y^{2}_{t}e^{-\log \breve{\hat{\sigma }}^{2}_{t}}\). It is easy to show that
Similarly as in Lemma 3, we have \(\frac{\partial ^{2}\breve{\hat{l}}_{t} ({\varvec{\theta }})}{\partial {\varvec{\theta }}\partial {\varvec{\theta }}^{{\mathrm{\tiny T}}}}\) is invertible. Notice that
which is of \(o_{a.s}(1)\). We have,
which is invertible. Then \( \frac{\partial ^{2}\hat{\mathcal {L}}({\varvec{\theta }})}{\partial {\varvec{\theta }}\partial {\varvec{\theta }}^{{\mathrm{\tiny T}}}}\) is invertible as desired.
Recall that \({\varvec{\theta }}\!=\!\left( c,a,{\varvec{\beta }}^{\mathrm{\tiny T}},{\varvec{\lambda }}^{\mathrm{\tiny T}}\right) ^{\mathrm{\tiny T}}\!=\!\left( {\varvec{\gamma }}^{\mathrm{\tiny T}},{\varvec{\lambda }}^{\mathrm{\tiny T}}\right) ^{\mathrm{\tiny T}}, \hat{{\varvec{\theta }}}=\left( {\varvec{\hat{\gamma }}}^{{\mathrm{\tiny T}}}, \hat{{\varvec{\lambda }}}^{{\mathrm{\tiny T}}} \right) ^{{\mathrm{\tiny T}}} ={\text {argmin}}_{{\varvec{\theta }} \in {\varTheta }}\frac{1}{T}\) \( \sum _{t=1}^{T}\hat{l}_{t} \) with \({\varTheta }={\varTheta }_1\times {\varTheta }_2\) and \(\tilde{{\varvec{\theta }}}=\left( {\varvec{\tilde{\gamma }}}^{{\mathrm{\tiny T}}}, \tilde{{\varvec{\lambda }}}^{{\mathrm{\tiny T}}} \right) ^{{\mathrm{\tiny T}}} \) with \({\varvec{\tilde{\gamma }}}\) in (13) and \(\tilde{\eta }(\cdot )=\mathbf {B}\tilde{{\varvec{\lambda }}}\).
Lemma 13
Under Assumptions (A1)–(A5),
Proof
Let \(\hat{\mathcal {L}}_T = T^{-1}\sum _{t=1}^T \hat{l}_t\). By Taylor expansion,
where \({\varvec{\zeta }}=\mathbf {t}\hat{{\varvec{\theta }}}+\left( \mathbf {I}-\mathbf {t} \right) \tilde{\varvec{\theta }}\). Therefore, \( \hat{{\varvec{{\theta }}}}-\tilde{{\varvec{\theta }}}=-\left( \frac{\partial ^2 \hat{\mathcal {L}}_T({\varvec{\zeta }})}{\partial {\varvec{\theta }} \partial {\varvec{\theta }}^T}\right) ^{-1} \frac{\partial \hat{\mathcal {L}}_{T}(\tilde{{\varvec{\theta }}})}{\partial {\varvec{\theta }}} \).
First, \( \frac{\partial \hat{\mathcal {L}}_{T}(\tilde{{\varvec{\theta }}})}{\partial {\varvec{\theta }}} = \left\{ \left( \frac{\partial \hat{\mathcal {L}}_{T}(\tilde{{\varvec{\theta }}})}{\partial {\varvec{\gamma }}} \right) ^{^{{\mathrm{\tiny T}}}},\left( \frac{\partial \hat{\mathcal {L}}_{T} (\tilde{{\varvec{\theta }}})}{\partial {\varvec{\lambda }}}\right) ^{^{{\mathrm{\tiny T}}}}\right\} ^{^{{ \mathrm{\tiny T}}}}\), where
Simple computations give,
Under Assumption (A2), the preceding terms are all bounded. To be specific,
Since \(\left| 1-y_{t}^{2}\tilde{\sigma }_{t}^{-2} \right| \le \left| 1-{y_{t}^{2}}{{\sigma }_{t}^{-2}}\right| +{y_{t}^{2}}O(h^{p})\), we obtain
Combining (14), (15) and (16), we have \(\left\| \frac{\partial \hat{\mathcal {L}}_{T}(\tilde{{\varvec{\theta }}})}{\partial {\varvec{\gamma }}}\right\| _{\infty }=O\left\{ \frac{1}{T(1-\delta )^{3}}+\frac{h^{p}}{(1-\delta )^{2}}\right\} =O(T^{-1}+h^{p}).\) We also have
Thus, \(\left\| \frac{\partial \hat{\mathcal {L}}_{T}(\tilde{{\varvec{\theta }}})}{\partial {\varvec{\theta }} }\right\| _{\infty }=O(T^{-1}+h^{p}).\)
Second, let
According to Lemma 12, we have \(V_{T}^{-1}=O(1), a.s.\), then
Proof of Theorem 1
We have
According to Lemma 13, \(\left\| \hat{\eta }-\tilde{\eta }\right\| _{2}=O_{p}\left\{ N^{1/2}\left( h^{p}+T^{-1}\right) \right\} \). Therefore,
which is of \(O_{p}\left\{ N^{1/2}\left( h^{p}+T^{-1}\right) \right\} \).
For the additive components, by Lemma 1 of Stone (1985), \(\left\| \hat{\eta }_{s_{2}}-\eta _{s_{2},0}\right\| _{2,s_2}=O_{p}\left\{ N^{1/2}\left( h^{p}+T^{-1}\right) \right\} \), for \(1\le s_{2}\le d_{2}\). And, similar to Lemma A.8 in Wang and Yang (2007), we obtain the same rate for empirical norms, i.e., \( \left\| \hat{\eta }_{s_{2}}-\eta _{s_{2},0}\right\| _{n,s_2}=O_{p}\left\{ N^{1/2}\left( h^{p}+T^{-1}\right) \right\} \), for \( 1\le s_{2}\le d_{2}.\)
Proof of Theorem 2
“The” result follows immediately from Lemmas 11, 13 and Slutsky’s Lemma.
Rights and permissions
About this article
Cite this article
Chen, M., Song, Q. Semi-parametric estimation and forecasting for exogenous log-GARCH models. TEST 25, 93–112 (2016). https://doi.org/10.1007/s11749-015-0442-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11749-015-0442-6
Keywords
- Financial volatility
- Log-GARCH
- Exogenous variable
- Semi-parametric regression
- Spline
- Quasi-likelihood estimation