Electricity consumption forecast in Brazil
We used models with a large number of variables to predict electricity consumption in Brazil in the short and medium term. These models were able to capture trends and seasonal patterns effectively. Variable selection indicated that past consumption values were essential for very short-term forecasts, while climate and calendar variables are relevant for longer horizons. These findings have significant implications for the energy market.
Prediction of infection with covid-19 at the beginning of the pandemic in Brazil
We present an innovative strategy to estimate the essential epidemiological equations for predicting the evolution of Covid-19 in early 2020, taking into account the presence of underreporting of the disease, especially among the asymptomatic group.
With this new approach, we were able to overcome the challenges arising from underreporting, a crucial factor that significantly distorts epidemiological projections. By focusing particularly on the asymptomatic, whose infections often go unnoticed, we were able to provide a more comprehensive and accurate view of the spread of the virus.
Forecast of interest rate term strucuture
we predict the term structure of interest rates. We identified that model combination methods provide superior performance over individual models. In particular, the longer the forecast horizon, the greater the contribution of the combination of forecast techniques.
Development of an uncertainty index for the Brazilian economy
We developed an uncertainty index for the Brazilian economy using neural networks and a public news base. We show that this index is competitive with other indices available in the literature and that it is capable of predicting changes in the prices of financial assets.
Credit scoring models with machine learning
We explored a variety of statistical and computational, includinh machine learning, techniques, to assess default in different real--world credit datasets. We analyze credit scoring models and investigate default behavior using survival anbalysis, ensemble methods, support vector machines, etc. Results show that machine learning models usually provide better results in predicting default, in comparison with traditional techniques such as logistici regression and decision trees.