Building a Model for Time Series Forecasting using AutoML Methods

The field of machine learning includes various algorithms that use data to train models that can afterwards deal with new previously unseen data. Selecting the most appropriate for a specific data algorithm and manually tuning its hyperparameters could be tedious and time-consuming. Methods and tools for the automation of this task form a field of Automated Machine Learning (AutoML). Most AutoML tools solve the problem of building models for classification and regression tasks based on data attributes without considering changes in a parameter over time. However, several systems already exist that automate the search of models for the time series forecasting problem. The peculiarity of the task of time series forecasting is that, in this case, it is necessary to consider the relationship between measurements and time, not just the diversity and other statistical characteristics of the samples. In this work, we explore several AutoML systems that capable to work with time series. The process of automated search of a model for time series forecasting is considered using the AutoGluon system and dataset containing data about temperature changes in different cities of the world. The effects of different limits and presets on the search of model is shown.