Overview

The last course I had to do was another seminar minister by professor Luis K. Hotta. Since it was in COVID-19 period this seminar was online. In the first part, we read and summarized three online presentations chosen by the professor. While in the last part, we elaborated two presentations. The former presentation was a class with topics for master degree level, and the latter was about an article. Besides that, we had to do a report on the classmates’ seminar.

Course assignments

In the first presentation my topic was quantile regression. Quantile regression was introduced by Koenker and Basset (1978). This model can be view as an extension on linear regression when the interest lies on estimate the impact of explanatory variable on the entire distribution of the response variable not only on the mean response. Particularly, the goal is estimate the \(\tau\)-quantile of \(Y\) given the covariates \(\mathbf{x}\), i.e., \[ Q_\tau(Y \mid \mathbf{x}) = \mathbf{x}_i\,\boldsymbol{\beta}(\tau) \] My complete presentation (in Portuguese) contains further details about quantile regression, estimation, inference (frequentist and Bayesian), and extensions.

The second presentation I chose the paper “FFORMA: Feature-based forecast model averaging” by Montero-Manso, P., Athanasopoulos, G., Hyndman, R. J. and Talagala, T. S. In this work, the authors proposed a method for computing weighted forecasts from different methods using time series features. The full presentation (in Portuguese) can be found here. My main criticism about the method is related to the computational effort, according to the authors the model spent 5 days in order to fit 100.000 of time series from M4 competition.