Automated Selection of ML/DL Techniques for Time Series Data

Abstract

Time-series data is widely used in a lot of domains, such as finance, and the analysis of it positively affects the development of society. It provides information on how data changes over time. However, analyzing and forecasting time-series data is challenging since it always comes with specific characteristics which may have an impact on the performance of models and researchers do not know before implementing models. At the same time, various models make it difficult to choose a suitable one for the datasets. In this thesis, a novel Automated Model Selection technique is proposed. We collect a wide range of models and time-series datasets and choose some of them to conduct a series of experiments to explore how different elements affect the performances of models. We make a thorough quantitative and qualitative analysis of the experimental results and based on this analysis we formulate several outcomes. We then develop the proposed technique based on these outcomes. This Automated Model Selection technique achieves the goal of selecting a suitable model for the input datasets automatically.