Análise e previsão de séries temporais via Facebook Prophet

Abstract

Forecasting is an estimate or projection of what may happen in the future based on information and data available from the past and present. In other words, it is an attempt to make an informed guess about what is likely to occur in the future, using previous information and knowledge of the current context. Forecasting is applied in various areas such as cryptocurrencies, finance, weather, science, politics, and more. In the statistical approach, there are several ways to make forecasts. An example is the Exponential Smoothing method, known for its simplicity and ease of understanding, as its formulas are not overly complex. Additionally, we have Neural Networks, which are useful for capturing non-linear relationships present in the data, improving the accuracy of forecasts. The field of forecasting is constantly studied and improved by professionals who use it in their daily work. In 2017, two data scientists from Facebook developed an innovative technique called Facebook Prophet. This technique was created to address two issues found in traditional time series forecasting methods. The first concerns the inflexibility and fragility of completely automatic forecasting techniques to incorporate useful assumptions or heuristics. The second issue is related to the dependence on a skilled and competent data science analyst to use more robust tools. With the rise of cryptocurrencies in the current world, there is a growing interest among professionals in obtaining well-adjusted models for digital currencies. This enables them to position themselves appropriately regarding the issue of buying or selling these assets. In this work, three useful forecasting methods will be compared: Exponential Smoothing (Holt-Winters), Neural Networks, and Facebook Prophet. Some accuracy measures such as Mean Absolute Error (MAE) and Mean Squared Error (MSE) will be used for the final selection of the best model.