This study presents a set of methodological proposals aimed at improving forecast reconciliation in the context of Hierarchical Time Series. The main
objective is to present original solutions to the theme, seeking to obtain more
accurate forecasts than those obtained by independent models for the different
levels of the hierarchy. The studies were conducted using real data, showing
the potentiality of application of the methods developed in different scenarios,
in which the time series are structured in a hierarchical fashion. This thesis
is composed of a set of essays that explore forecast reconciliation from the
perspective of a regression model, which gives foundations to optimal reconciliation. The first contribution addresses the problem of forecast reconciliation
from the perspective of robust estimators. The proposal presents an original
contribution applied to data from labor force surveys in Brazil, presenting a
set of solutions that can drive efficient public policies. In this case, the reconciled forecasts obtained through robust estimators provided consistent gains
in terms of accuracy when compared to methods that represent the state-of-the-art on forecast reconciliation in hierarchical time series. The second contribution deals with the problem of optimal reconciliation applied to energy
consumption time series in Brazil. We present an alternative proposal, less
sensitive to outlying forecasts at the reconciliation stage. The results obtained
in this second study show considerable improvements in standard evaluation
metrics with regard to the new forecasts. The third proposal seeks to offer
robust covariance structures for forecasting errors, which expands the set of
strategies presented in the literature. The main contribution is to incorporate
robust covariance estimates into the MinT (Minimum Trace) reconciliation
approach, which minimizes reconciliation errors, offering an estimator with
minimum variance.
|