ESTE 2023

Name: ESTE 2023
Start: 2023-07-30-0300:00
End: 2023-08-02-0300:00
Location: Hotel Torres da Cachoeira

20ª Escola de Séries Temporais e Econometria

Certificates Click here to download

From 30th July to 2th August Every day from 00h00 to 00h00

Florianópolis, SC Hotel Torres da Cachoeira

About the Event

The 20th Time Series and Econometrics Meeting (20th ESTE) will take place at the Hotel Torres da Cachoeira, in Florianópolis (SC), Brazil, from July 30 to August 2. ESTE has both national and international scopes and is organized every two years, gathering important Brazilian researchers as well as a selected number of international experts in the areas of Time Series and Econometrics.

Schedule of Oral Communications can be viewed here.

Schedule of Poster Sessions can be viewed here.

Scientific Comittee

Esther Ruiz Ortega (Universidad Carlos III de Madrid)
Ana B. Galvão (University of Warwick, UK)
Pedro Alberto Morettin (USP)
Valdério Anselmo Reisen (UFES)
Eduardo Fonseca Mendes (EESP-FGV)
Guilherme Valle Moura (UFSC)

Organizing Comittee

Guilherme Valle Moura (UFSC)
Aluísio Pinheiro (Unicamp)
Flávio Ziegelmann (UFRGS)
João F. Caldeira (UFSC)
Andrea Konrath (UFSC)
Pedro Chaim (UFSC)

Speakers

Ana Galvão
David Stoffer
Hedibert Freitas Lopes
Sílvia Gonçalves
Wilfredo Palma
Yukun Liu
André Portela Santos
Ricardo Giglio
Luiz Renato Lima
Carlos Brunet Martins-Filho
Paolo Santucci de Magistris
Amin Shams
Silvia Regina Lopes
Pedro Valls
Luiz Hotta
Pedro Morettin
Aluísio Pinheiro
Flávio Ziegelmann
Carlos César Trucíos Maza

Schedule

13h30 - Ricardo Giglio Data science and time series models: cross-validation and Python Darts library (Ricardo Giglio, Chief Data Officer at 3778.care) Short course +

Place: Auditório Victor Meirelles

Ricardo Giglio

The first part of this short course provides an introduction to the relationship between data science and time series analysis. With a practical and applied focus, the second part discusses particularities of cross-validation in the context of time series as well as its importance in building trust on the part of users - crucial for the adoption of this type of technology in practice. The last part of the short course presents the Python Darts library, which, with its many functionalities, deserves to be called the "Swiss army knife" for time series. In addition to presenting the basic concepts of using the tool, the short course shows how to easily generate visualizations, diagnoses, cross-validation and the use of a large collection of models available in the library with low programming effort.

Part 1) Introduction to Data Science and Time Series (30 min)
- Main problems and methods
- Case 1: Demand forecast in the context of Production Planning and Control
Part 2) Cross-validation in time series (40 min)
- Possibilities of experiment definitions (training, testing and validation)
- Assessment metrics aligned to the business
- Case 2: Claims forecast in health insurance plans
Part 3) Hands On: The Darts Package (50 min)
- Python and packages in 5 minutes
- Visualization and diagnostics (seasonality, trend, autocorrelation)
- Complex cross-validation without major programming efforts
- Univariate models (SARIMAX, Holt-Winters)
- Multivariate models (VARIMA, Kalman)
- Global models
- Past and future covariates

Slides for this tutorial can be downloaded here.

16h00 Oral Communication I Oral presentation +

Place: Auditório Victor Meirelles

Presentation of selected works submitted to XX ESTE.

08h30 Opening Opening +

Place: Auditório Victor Meirelles

Opening

09h00 - Sílvia Gonçalves Keynote Speech Lecture +

Place: Auditório Victor Meirelles

Impulse Response Functions for Nonlinear Models

10h00 Coffee Break Coffee break +

Place: Lobby

Coffee break.

10h30 - Aluísio Pinheiro, Michel Montoril, Pedro Morettin Wavelets Thematic Symposium +

Place: Auditório Victor Meirelles

Weighted U–statistics for absolutely regular processes with applications to time series and other dependent sequences

Aluisio S. Pinheiro, Rodney Fonseca

U-statistics is a classical tool in nonparametric inference, commonly used to establish large sample properties of various estimators. We focus on weighted U-statistics
for absolutely regular processes and discuss different aspects of this type of estimator, such as data dependency level, type of weights, and kernel degeneracy. We prove a
central limit theorem for weighted U-statistics with non-degenerate kernels, and discuss connections of this problem with V-statistics and the case where kernels are degenerate. An application illustrates how the proposed methods can be used in related problems in the literature.

Bayesian estimation of mixture regression by wavelets

Michel H. Montoril, Flávia C. Motta

In this work we consider a mixing problem of two Gaussian components, where the weight of the mixture has a dynamic behavior (for example, it varies over time). We propose a Bayesian method to jointly estimate the component parameters and the dynamic mixture weights. The key idea of this method is to apply a transformation to the data to deal with a regression problem, where dynamic mixture weights represent the regression function. Estimates are obtained based on MCMC samples of the posterior parameters. For this task, an efficient algorithm based ona Gibbs sampler is proposed. We observed a good performance of the method through Monte Carlo simulation studies. Furthermore, a real dataset application using an array Comparative Genomic Hybridization (aCGH) data illustrates our approach.

Clustering and classification by wavelets

Pedro A. Morettin, Chang Chiann, João R. Sato and Brani Vidakovic.

In this work we consider several wavelet-based procedures for clustering and classication purposes. In some situations, the time domain approach may not lead to clear classication or discrimination. When we move to the wavelet domain, the multiresolution analysis leads to look at data in several levels of resolution (or scales) and then the separation may become better. Among the wavelet-based procedures, we mention:

(a) Multifractal Spectra (MFS) and associated descriptors.
(b) DWT-CEM procedure: discrete wavelet transform combined with classication expectation maximization algorithm.
(c) DWT-Schur measures: discrete wavelet transform followed by the use of some Schur monotone measure.
(d) Wavelet-based Bayesian discriminant function.

10h30 - Carlos César Trucíos Maza Thematic Session on High Dimensional Econometrics Thematic Symposium +

Place: Sala Zininho

Forecasting Value-at-Risk and Expected Shortfall in Large Portfolios
Carlos Trucíos, Marc Hallin

Beyond their importance from the regulatory policy point of view, risk measures play an important role in risk management, portfolio allocation, capital level requirements, trading systems, and hedging strategies. However, due to the curse of dimensionality, their estimation and forecast in large portfolios is a difficult task. To overcome these problems, we propose a new procedure based on residual-based bootstrap, the general dynamic factor model and robust volatility models. The new procedure is applied in US stocks and the backtesting results indicate that the new proposal outperforms several existing alternatives.

Global variable selection for quantile regression

Eduardo Horta, Tais Bellini, Gabriela Cybis

Quantile regression models the conditional quantile function of the response variable Y given covariates X, providing robust estimators for the entire conditional distribution and handling outliers effectively. However, variable selection becomes complex when specifying the functional form for each quantile level tau, using techniques like LASSO or adaLASSO. In this study, we propose a method for global variable selection and coefficient estimation in linear quantile regression with flexible $beta and group adaLASSO penalization. We conduct a Monte Carlo study comparing six estimators across diverse scenarios, observing that the selection of lambda significantly impacts model selection and coefficient estimation. Traditional LASSO tends to include the true model more frequently but lacks shrinkage and covariate removal, whereas grouped approaches effectively eliminate less relevant coefficients.

HIGH-DIMENSIONAL LINEAR PROCESSES WITH DEPENDENT INNOVATIONS

Eduardo Mendes

We develop concentration inequalities for the $\ell_\infty$ norm of vector linear processes on mixingale sequences with sub-Weibull tails. These inequalities make use of the Beveridge-Nelson decomposition, which reduces the problem to concentration for the sup-norm of a vector-mixingale or its weighted sum. Using this decomposition, we derive a concentration bound for the maximum entrywise norm of autocovariance matrices of linear processes. These results are useful for estimation bounds for high-dimensional vector autoregressive processes estimated using $\ell_1$ regularization, high-dimensional Gaussian bootstrap for time series, and long-run covariance matrix estimation.

12h00 Lunch Lunch +

Place: Restaurant

Lunch

13h30 - David Stoffer Keynote Speech Lecture +

Place: Auditório Victor Meirelles

Detection of Narrowband Frequency Changes in Time Series

14h30 - Carlos Brunet Martins-Filho Keynote Speech Lecture +

Place: Auditório Victor Meirelles

Optimal nonparametric estimation of distribution functions, convergence rates for Fourier inversion theorems and applications

15h30 Coffee Break Coffee break +

Place: Lobby

Coffee break.

16h00 - Silvia Regina Lopes Session in Honor of Sílvia Lopes Ceremony Honoris Causa +

Place: Auditório Victor Meirelles

Session in Honor of Sílvia Lopes for her important contributions to the area of time series.

17h30 Oral Communication II Oral presentation +

Place: Auditório Victor Meirelles

Presentation of selected works submitted to XX ESTE.

17h30 Oral Communication III Oral presentation +

Place: Sala Zininho

Presentation of selected works submitted to XX ESTE.

19h00 Poster Session I Posters Dinâmicos +

Place: Lobby

Selected poster presentations.

08h30 - André Portela Santos Interpretability of Machine Learning Models (André Portela Santos, CUNEF University) Short course +

Place: Auditório Victor Meirelles

André Portela Santos

While machine learning models represent the frontier of predictive analytics, the inherent lack of transparency is often an obstacle to their application. This short course will review interpretability methods that allow you to open the "black box" of machine learning models with an emphasis on agnostic methods of interpretation. Throughout the course we will use a database containing 126 macroeconomic variables to illustrate the implementation of interpretability methods.

1. Introduction to methods of interpretability of ML models
2. Model-intrinsic interpretability methods
3. Model-agnostic interpretability methods
- 3.1. Partial dependence
- 3.2. Permutation
- 3.3. Surrogate models
- 3.4. Shapley values
4. Application: product growth forecast using regression trees

Materials for this short course are available here.

10h00 Coffee Break Coffee break +

Place: Lobby

Coffee break.

10h30 - David Stoffer, Flávio Ziegelmann, Wilfredo Palma Thematic Session on Nonlinear Models Thematic Symposium +

Place: Auditório Victor Meirelles

Unbiased estimation of near-root autoregressive models

Wilfredo Palma, Susana Eyheramendy and Felipe Elorrieta

In many practical cases, the commonly used methods for parameter estimation in the context of time series data such as the maximum likelihood estimator (MLE) and the
minimum square error estimator (MSE) exhibit biases. In particular, this problem arises when the parameters are close to the unit root or the process has additive noise. In
order to improve these parameter estimation procedures, this talk discusses the development of a simulation-extrapolation (SIMEX) methodology in the context of
time series analysis. We consider a standard class of autoregressive processes AR(p) observed at regular times. Further, we study this problem in a class of irregularly
autoregressive processes (IAR). Additionally, we analyze the effects of additive noise in the processes. The performance of the proposed SIMEX algorithm is investigated by
means of extensive simulations and real-life data illustrations. These results show that the proposed methodology substantially improves the quality of the estimates from
MLE and MSE methods, in terms of both bias reduction and estimation precision, in the context of both the regular/irregular and additive noise problems.

Efficient Fitting of Stochastic Volatility Models

David S. Stoffer

The stochastic volatility model is a popular tool for modeling the volatility of assets. The model is a nonlinear and non-Gaussian state space model. Most approaches use intensive numerical techniques that present challenges not seen in general. Bayesian approaches use Markov chain Monte Carlo (MCMC), but convergence and mixing problems plague MCMC algorithms used for the model. We present an approach that ameliorates the slow convergence and mixing problems when fitting stochastic volatility models. The approach accelerates the convergence by exploiting the geometry of one of the targets. We demonstrate the method on various numerical examples.

Forecasting Tail Risk for Energy Markets via Dynamic GAS Vine Copulas

Flávio A. Ziegelmann

Our main goal is to forecast tail risk for the energy commodities market. To address this
challenge, we build a naive portfolio and propose a dynamic D-vine copula modelling
strategy. This model allows us to capture the complex dependence structures of energy
commodity returns, while also accommodating for their specific characteristics, such as
asymmetries and heavy tails. To incorporate the time-varying dependence structure, we
use a generalized autoregressive score (GAS) model as the updating mechanism for the
copula parameters. Our results suggest that our approach outperforms others in terms of
average loss.

10h30 - Hudson Torrent, Paolo Santucci de Magistris, Vinícius Albani Thematic Session on Financial Econometrics Thematic Symposium +

Place: Sala Zininho

Realized Iliquidity

Paolo Santucci de Magistris, Demetrio Lacava, Angelo Ranaldo

Realized illiquidity is the ratio between realized volatility and trading volume refining the popular price impact measure proposed by Amihud (2002). We provide its theoretical foundation in which both price volatility and market liquidity follow stochastic processes in continuous time. We prove that the realized illiquidity is a precise measurement of the inverse of integrated liquidity over periods of unit length (e.g., a day). A comprehensive econometric analysis highlights the main distributional and dynamic properties of the realized illiquidity, including jumps, clustering, and leverage effects, and demonstrate that they help explain the time series of stock and currency returns.

Predicting Electricity Prices in the Brazilian Market by Stochastic Differential Equations

Vinícius Albani

We propose a model based on stochastic differential equations to predict forward electricity prices in the Brazilian market. The model accounts for usual stylized facts for commodities, such as mean-reversion and jumps. Since in the Southeast market, the primary power resource is hydroelectric, the model also depends on the level of water reservoirs. The model parameters are estimated from observed end-of-day forward contract prices by nonlinear regression with ridge penalty.

Let the machines speak: A comparison of variable selection methods for portfolio choice problems

Hudson Torrent, André A. P. Santos, Guilherme V. Moura

This paper investigates the performance of alternative variable selection methods from the machine learning literature when applied to portfolio choice problems. We parameterize portfolio weights as a function of a large pool of firm-level characteristics as well as their second-order and cross-product (interaction) terms, yielding a total of 4,610 predictors. We find that parametric portfolios formed with the variables selected with the lasso and L2-boosting methods outperform all other methods in terms of portfolio risk. However, the L2-boosting is clearly superior when portfolios are evaluated in terms of risk-adjusted returns. All methods indicate that interaction terms are dominant among selected predictors, which highlights the importance of accounting for non-linearities in the relation between stock characteristics and optimal allocations. Moreover, the number of selected variables i) varies across market conditions and ii) vastly exceeds the number of variables used in popular factor models, suggesting that ad-hoc sparsity can be detrimental to portfolio performance.

12h00 Lunch Lunch +

Place: Restaurant

Lunch

13h30 - Ana Galvão Keynote Speech Lecture +

Place: Auditório Victor Meirelles

Probabilistic Event Forecasting in Macroeconomics.

14h30 - Hedibert Freitas Lopes Keynote Speech Panel +

Place: Auditório Victor Meirelles

What events matter for exchange rate volatility?

Co-author: Igor Martins.

Abstract: This paper identifies and quantifies the effect of macroeconomics events of multiple countries on exchange rate volatility using high frequency currency returns while accounting for persistent stochastic volatility effects and seasonal components capturing time of the day patterns. Due to the hundreds of macroeconomic announcements and its lags, we rely on sparsity based methods to select relevant events for the model. We contribute to the literature in four ways: First, we identify the macroeconomic events that drive currency volatility, estimate their effect, connect them to macroeconomic fundamentals and show how they can be linked to lower frequency currency returns using a model averaging argument. Second, we find a connection between intraday seasonality, trading volume and opening hours of majors markets across the globe and provide a simple labor-based argument for the pattern found. Third, we show that inclusion of macroeconomic events and seasonal components are key for forecasting exchange rate volatility. Fourth, applying our proposed model for multiple currencies alongside a dynamic copula yields a Sharpe ratio 3.5 times higher than using standard SV and GARCH models.

15h30 Coffee Break Coffee break +

Place: Lobby

Coffee break.

16h00 - Pedro Valls Session in Honor of Pedro Valls Ceremony Honoris Causa +

Place: Auditório Victor Meirelles

Session in Honor of Pedro Valls for his important contributions in to area of time series econometrics.

17h30 Oral Communication IV Oral presentation +

Place: Auditório Victor Meirelles

Presentation of selected works submitted to XX ESTE.

17h30 Oral Communication V Oral presentation +

Place: Sala Zininho

Presentation of selected works submitted to XX ESTE.

19h00 Poster Session II Posters Dinâmicos +

Place: Lobby

Selected poster presentations.

08h30 - André Portela Santos Interpretability of Machine Learning Models (André Portela Santos, CUNEF University) Short course +

Place: Auditório Victor Meirelles

André Portela Santos

1. Introduction to methods of interpretability of ML models
2. Model-intrinsic interpretability methods
3. Model-agnostic interpretability methods
- 3.1. Partial dependence
- 3.2. Permutation
- 3.3. Surrogate models
- 3.4. Shapley values
4. Application: product growth forecast using regression trees

Materials for this short course are available here.

10h00 Coffee Break Coffee break +

Place: Lobby

Coffee break.

10h30 - Amin Shams, Paolo Santucci de Magistris, Yukun Liu Thematic Session: Econometrics of Digital Assets Thematic Symposium +