This paper extends previous research, which proposed a methodology based on the following hypothesis: dealing with the problem of predicting the next-day USD/BRL exchange rate daily trend, the existence of calendar effects allows us to improve trained voting-based ensemble models without model retraining. Despite the evidence of good results, in the present work, we propose adding naïve models to the originally proposed methodology because naïve models would also potentially benefit from the calendar effect becoming a benchmark to consider. The experiments confirmed that naïve models are not just challenging benchmarks but also models that can be included in the process to improve existing voting-ensemble models. On average, adding the naïve models to the original solution generated an increase higher than 100% in the value of the primary metric adopted for performance measurement. Constantly overwhelmed by more complex solutions, we can take these outcomes as a reminder not to neglect simplicity.
Complete tool for submission, evaluation and publication of proceedings for scientific events.
