Inference Engines for Streaming Datasets

The problem of forecasting streaming datasets, particularly the financial time series, has been largely explored in the past, but we believe the advancement of technologies such as the Internet of Things, which will connect an exponentially increasing number of sensors and devices, endowed with limited computational resources, yet capable of producing enormous amounts of sampled data, and the progressively higher social need to deploy intelligent systems, will make the prediction of time series a core industrial issue in the next future. Consequently, we also believe that investigating efficient models for accurate and reliable forecasting can be considered an urgent area of research. We explore several non parametric univariate time series forecasting approaches, mainly involving the use of pattern recognition and digital signal preprocessing modules to be coupled with established neural regressors, with the ambition to design hybrid machine learning frameworks candidate to be applied to a variety of application fields. Results show how the proposed models can outperform the benchmarking methods, suggesting good forecasting accuracy when applied to one of the most recent and less known financial time series: the Bitcoin-US Dollar hourly close spot rates.