Venue: H10-31
Time: 16:00
Ilker Birbil (Sabanci University)
Tuesday 3 October
Parallel Incremental Optimization Algorithm for Solving Partially Separable Problems in Machine Learning
Abstract:
Consider a recommendation problem, where multiple firms are willing to cooperate to improve their rating predictions. However, the firms insists on finding a machine learning approach, which guarantees that their data remain in their own servers. To solve this problem, I will introduce our recently proposed approach HAMSI (Hessian Approximated Multiple Subsets Iteration). HAMSI is a provably convergent, second order incremental algorithm for solving large-scale partially separable optimization problems. The algorithm is based on a local quadratic approximation, and hence, allows incorporating curvature information to speed-up the convergence. HAMSI is inherently parallel and it scales nicely with the number of processors. I will conclude my talk with several implementation details and our numerical results on a set of matrix factorization problems.
Arun Chandrasekha (Stanford University)
12 October 2017 ****Cancelled****
Title: TBA
Abstract:
TBA
Michael Smith (Melbourne Business School)
24 October 2017
Time Series Copulas for Heteroskedastic Data
Abstract:
We propose parametric copulas that capture serial dependence in stationary heteroskedastic time series. We suggest copulas for first order Markov series, and then extend them to higher orders and multivariate series. We derive the copula of a volatility proxy, based on which we propose new measures of volatility dependence, including co-movement and spillover in multivariate series. In general, these depend upon the marginal distributions of the series. Using exchange rate returns, we show that the resulting copula models can capture their marginal distributions more accurately than univariate and multivariate GARCH models, and produce more accurate value at risk forecasts. Last, we outline an alternative approach to solving this problem based on extracting the “implicit” or inversion copula of existing parametric time series copula models.
Marc Hallin (Université libre de Bruxelles)
27 October 2017
Quantile Spectral Analysis for Locally Stationary Time Series
Abstract:
Classical spectral methods are subject to two fundamental limitations: they only can account for covariance-related serial dependencies, and they require second-order stationarity. Much attention has been devoted lately to {\em quantile} (copula-based) {\em spectral methods} that go beyond traditional covariance-based serial dependence features. At the same time, covariance-based methods relaxing stationarity into much weaker {\it local stationarity} conditions have been developed for a variety of time-series models. Here, we are combining those two approaches by proposing copula-based spectral methods for locally stationary processes. We therefore introduce a time-varying version of the copula spectra that have been recently proposed in the literature, along with a suitable local lag-window estimator. We propose a new definition of local {\it strict} stationarity that allows us to handle completely general non-linear processes without any moment assumptions, thus accommodating our copula-based concepts and methods. We establish a central limit theorem for the new estimators, and illustrate the power of the proposed methodology by means of a simulation study. Moreover, real-data applications demonstrate that the new approach detects important variations in serial dependence structures both across time and across quantiles. Such variations remain completely undetected, and are actually undetectable, via classical covariance-based spectral methods.
Based on joint work with Stefan Birr (Ruhr-Universit\" at Bochum), Holger Dette (Ruhr-Universit\" at Bochum), Tobias Kley (London School of Economics) and Stanislav Volgushev (University of Toronto).
(The main reference (JRSSB 2017 DOI: 10.1111/rssb.12231) is available online:
Julia Schaumburg (Vrije Universiteit Amsterdam)
16 November 2017
Time-Varying Vector Autoregressive Models with Structural Dynamic Factors
Abstract:
We suggest a simple methodology to estimate time-varying parameter vector autoregressive (VAR) models. In contrast to the widely used Bayesian approach, our approach is based on combining a dynamic factor model for the VAR coefficient matrices and a score-driven model for the time-varying variances. Our algorithm is robust and fast, while being easy to implement. In a small simulation study, we demonstrate the good performance of the method. Furthermore, using the empirical data set on U.S. macroeconomic and financial variables that is also used in Prieto et al. (2016), we show that our approach is promising in modeling time-varying macro-financial linkages.
Joint work with Paolo Gorgi and Siem Jan Koopman
Ana Galvao (Warwick University)
23 November 2017
Credit Conditions and the Effects of Economic Shocks: Amplification and Asymmetries
In this paper we address three empirical questions related to credit conditions. First, do they change the dynamic interactions of economic variables? Second, do they enlarge the effects of economic shocks? Third, do they generate asymmetries in the effects of economic shocks? To answer these questions, we introduce endogenous regime switching in the parameters of a large Multivariate Autoregressive Index (MAI) model, where all variables react to a set of observable common factors. We develop Bayesian estimation methods and show how to compute responses to common structural shocks. We find that credit conditions do act as a trigger variable for regime changes. Moreover, demand and supply shocks are amplified when they hit the economy during periods of credit stress. Finally, good shocks seem to have more positive effects during stress time, in particular on unemployment.
(with A. Carriero and M. Marcellino)
Arthur Lewbel (Boston Colleage)
30 November 2017
Keeping up with peers in India: A new social interactions model of perceived needs
Abstract:
We propose a new nonlinear model of social interactions. The model allows point identification of peer effects as a function of group means, even with group level fixed effects. The model is robust to measurement problems resulting from only observing a small number of members of each group, and therefore can be estimated using standard survey datasets. We apply our method to a national consumer expenditure survey dataset from India. We find that each additional rupee spent by one's peer group increases one's own perceived needs by roughly 0.5 rupees. This implies that if I and my peers each increase spending by 1 rupee, that has the same effect on my utility as if I alone increased spending by only 0.5 rupees. Our estimates have important policy implications, e.g., we show potentially considerable welfare gains from replacing government transfers of private goods with the provision of public goods.
George Kapetanios (Queen Mary University of London)
7 December 2017
Title:Topics in time varying coefficient models
Abstract:
This presentation discusses recent work on time varying coefficient models. The first parts discusses a test to distinguish between stationary and persistent volatility models. We find significant empirical evidence in favour of persistent volatility and against standard volatility models such as GARCH. The second part discusses a new method for estimating time varying models based on the principles underlying the Hodrick Prescott filter.
Emilio Porcu
12 December 2017 (12:00-13:00)
Title: TBA
Abstract: TBA
Andros Kourtellos (Cyprus University)
14 December 2017
Endogeneity in Semiparametric Threshold RegressionAbstract:
Abstract:
In this paper, we investigate semiparametric threshold regression models with endogenous threshold variables based on a nonparametric control function approach. Using a series approximation, we propose a two-step estimation method for the threshold parameter. For the regression coefficients, we consider least-squares estimation in the case of exogenous regressors and two-stage least-squares estimation in the case of endogenous regressors. We show that our estimators are consistent and derive their asymptotic distribution for weakly dependent data. Furthermore, we propose a test for the endogeneity of the threshold variable, which is valid regardless of whether the threshold effect is zero or not. Finally, we assess the performance of our methods using a Monte Carlo simulation.