[ABE-L] Workshops FGV EMAp em Novembro

[Rodrigo Targino]

Prezados colegas,
Gostaria de convidá-los para dois eventos que serão realizados na Escola de
Matemática Aplicada (EMAp) da Fundação Getulio Vargas (FGV), no Rio de
Janeiro.

Dia 03/nov/25 (segunda-feira): Workshop em Modelagem Estocástica de Eventos
Climáticos
Dia 05/nov/25 (quarta-feira): Workshop em Estatística, Séries Temporais e
Finanças Quantitativas
Local (dos dois eventos): Fundação Getulio Vargas, Praia de Botafogo, 190,
5o andar, Auditório 537
Horário (dos dois eventos): 13.30 - 18.00h

Os eventos ocorrerão de forma presencial (sem transmissão online ou
gravação) e estão sendo organizados pela FGV EMAp e o PPG de Estatística do
IME-UFRGS, com o apoio dos dois departamentos e do CNPq através do Projeto
Universal “Modelagem Econométrica de Problemas Complexos: Séries Temporais
Funcionais, Modelos Não Lineares, Aprendizado Estatístico e Modelos de Alta
Dimensão”.

Os interessados devem enviar um email para eventos.emap em fgv.br para
realizar sua inscrição.

Os títulos e resumos dos trabalhos estão disponíveis abaixo.

Atenciosamente,
Rodrigo Targino

======================================================================

Workshop em Modelagem Estocástica de Eventos Climáticos (3/nov)

Leonardo Voltarelli (IMPA)

Precipitation nowcasting of satellite data using physically-aligned neural
networks

Accurate short-term precipitation forecasts remain concentrated in
radar-rich regions, limiting operational value in places most exposed to
climate extremes. We present TUPANN (Transferable and Universal
Physics-Aligned Nowcasting Network), a satellite-only model trained on
GOES-16 RRQPE. Unlike most deep learning models for nowcasting, TUPANN
separates the forecast into physically meaningful components: a variational
encoder–decoder infers motion and intensity fields from recent imagery
under optical-flow supervision, a lead-time-conditioned MaxViT evolves the
latent state, and a differentiable advection operator reconstructs future
frames. We investigate TUPANN's performance across four distinct climates
(Rio de Janeiro, Manaus, Miami, La Paz) at 30–180-min lead times using
CSI/HSS over 4–64 mm h^{-1} thresholds and both IMERG and GOES-16 data. Our
model is benchmarked against leading models based on optical flow
(PySTEPS), deep learning (Earthformer and CasCast) and a combination of the
two (NowcastNet). Overall, TUPANN delivers the best or second-best skill in
most settings, with pronounced gains at higher thresholds; training on
multiple cities further improves performance, while cross-city tests show
limited degradation and occasional gains for rare heavy-rain regimes.
GOES-16's near 5-min latency supports our model's real-time use. Beyond
skill, TUPANN exposes smooth motion fields that align with numerical
optical flow, improving interpretability for forecasters. These results
indicate that physically aligned learning can provide nowcasts that are
skillful, transferable and global.

—--------------------------------------------------------------------------------------------------------

Klaus Boesch (IFSul)

Modelling and Forecasting the Dynamics of Spatial Surfaces via Dynamic
FPCA: Application to Daily Temperature Fields in Southern Brazil

This work introduces a framework for modeling and forecasting spatial
surface time series using a dynamic extension of Functional Principal
Component Analysis (FPCA). We generalize the Bathia-Yao-Ziegelmann Dynamic
FPCA to cases where observations are spatial surfaces. Simulation studies
confirm the method's ability to accurately recover underlying dynamic
structures and identify the true latent dimension. An empirical analysis of
daily mean, maximum, and minimum temperature surfaces in Rio Grande do Sul,
Brazil, demonstrates the practical utility of the framework. Our findings
show that the optimal predictive dimensionality varies with the forecast
horizon and the specific temperature variable. The analysis yields
interpretable spatial patterns, captures their temporal dynamics, and
provides reliable forecasts, offering a valuable tool for climate science,
environmental studies, and other fields such as quantitative finance.

—--------------------------------------------------------------------------------------------------------

Reinaldo Marques (IRB-Re)
Spatial Extreme Events: a climate risk assessment in Brazil

Extreme weather events have become a major challenge for insurers' pricing
and catastrophe areas. In this study, we employ daily data (1961–2024) on
0.1°×0.1° grids for seven climate variables, from which more than 50
climate indices were computed. Spatial clusters were delineated within
Brazilian river basins using Machine Learning clustering algorithm. For
each cluster, representative extreme climate indices of Heat Waves (WSDI),
Cold Waves (CSDI), Heavy Rainfall (RX1D), and Droughts (CDD) were modeled
using the Generalized Extreme Value (GEV) distribution, enabling
characterization of distribution, variability, and frequency, as well as
estimation of return levels for 2-, 5-, 10-, and 50-year periods under
future scenarios. Additionally, extremes were spatially quantified using
the area-perimeter ratio based on excursion sets, assessing their extent
and displacement across the national territory. This study highlights the
high climate variability resulting from Brazil’s extensive territorial
heterogeneity, providing essential support for preventive planning in the
insurance sector and for the formulation of policies and products adapted
to different levels of exposure to climate-related disasters.

—--------------------------------------------------------------------------------------------------------

Livia Cereja (FGV EMAp)

Automatic Climate Events Categorization with Masked Siamese Networks

Understanding and representing complex climate variability is essential for
both scientific analysis and predictive modeling. However, identifying
meaningful climate regimes from raw variables is challenging, as they
exhibit high noise and nonlinear dependencies. In this work, we explore the
use of Masked Siamese Networks to discretize climate time series into
semantically rich clusters. Focusing on daily minimum and maximum
temperature, we show that the resulting representations: (i) yield clusters
that reflect meaningful climate states under our modeling assumptions,
offering a simplified representation for downstream use; (ii) enable
sampling and analysis of specific climate scenarios; and (iii) exhibit
statistical associations with El Niño events, underscoring their scientific
relevance. Our findings highlight the potential of self-supervised
discretization as a tool for climate data analysis and open avenues for
incorporating richer climate indicators in future work.

======================================================================

Workshop em Estatística, Séries Temporais e Finanças Quantitativas (05/Nov)

Eduardo Horta (UFRGS)

Product Disintegrations of Markov Chains: EVT and an application to climate
data

For a given sequence of random variables, Borsato et al. (2024, Statistics
& Probability Letters - DOI: https://doi.org/10.1016/j.spl.2024.110056)
introduce the concept of a product disintegration, which is a latent
sequence of random probability measures upon which conditioning makes the
original sequence independent, and such that a fixed point property holds
for the conditioning operator. Drawing from these authors, we show
constructively that any discrete-time Markov chain on a countable state
space admits, under mild conditions, a non-trivial product disintegration
which is also a Markov chain. In an EVT framework, we derive the “quenched”
limit for the conditional distribution of the maximum of such chains, and
obtain bounds for the “annealed” limit. Finally, we present a brief
application to precipitation data.

—--------------------------------------------------------------------------------------------------------

Flavio Ziegelmann (UFRGS)

Improving Copula-GARCH Risk Forecasting Learning from Factor Functional
Time Series

In modern days, the accurate prediction and forecasting of risk measures,
such as Value at Risk and Expected Shortfall, is an essential task for
asset market managers. When calculating risk measures, an essential step,
for most approaches, is to estimate the probability density function of
asset returns. A daily sequence of intraday return densities of p assets,
denoted by Y_t, t=1,…,n, can be seen as a p-dimensional functional time
series. If p is large (Y_t is high dimensional), then one has to perform a
two-way dimension reduction: in the high dimensional vector and in the
infinite dimensional curves. Here we propose combining a Functional Factor
Model with a univariate Dynamic Functional Principal Components Analysis as
a two way dimension reduction approach, which allied to a copula model
feeds the error term of a high-frequency ARMA-GARCH model aiming to
forecast future daily risk measures.

—--------------------------------------------------------------------------------------------------------

Eduardo Mendes (FGV EESP)

Estimation Risk in Conditional Expectiles

We establish the consistency and asymptotic normality of a two-step
estimator of conditional expectiles in the context of location-scale
models. We first estimate the parameters of the conditional mean and
variance by quasi-maximum likelihood and then compute the unconditional
expectile of the innovations using the empirical quantiles of the
standardized residuals. We show how replacing true innovations with
standardized residuals affects the asymptotic variance of the expectile
estimator. In addition, we also obtain asymptotic-valid bootstrap-based
confidence intervals. Finally, our empirical analysis reveals that
conditional expectiles are very interesting alternatives to assess tail
risk in cryptomarkets, relative to traditional quantile-based risk
measures, such as value at risk and expected shortfall.

—--------------------------------------------------------------------------------------------------------

Rodrigo Targino (FGV EMAp)

Risk-Budgeted Mean Variance Portfolios

We introduce the Risk-Budgeted Mean-Variance (RBMV) portfolio, a novel
framework that connects the classical Markowitz mean-variance problem and
the risk budgeting approach. By modifying the risk budgeting optimization
problem to include constraints on expected returns and volatility, RBMV
offers a disciplined way to manage the trade-off between risk concentration
and return maximization. The investor gains a lever to adjust how close the
portfolio sits to either framework, depending on her preferences. We show
that the optimization problem that defines the RBMV portfolio is convex,
efficiently computable, and typically delivers competitive returns with
reduced risk concentration in the context of long-only portfolios. We
illustrate our methodology using daily equity returns from the U.S. and
show that our methodology efficiently controls the volatility of returns
while also delivering Sharpe ratios that are consistently higher than the
traditional mean-variance approach.
-------------- Próxima Parte ----------
Um anexo em HTML foi limpo...
URL: <http://lists.ime.usp.br/pipermail/abe/attachments/20251030/4d161ef5/attachment-0001.htm>