Time series formalism : a systems approach

Abstract

Time series data has become a modern day phenomena: from stock market data to social media information, modern day data exists as a continuous flow of information indexed by timestamps. Using this data to gather contextual inference and make future predictions is vital to gaining an analytical edge. While there are specialized time series databases and libraries available that optimize for performance and scale, there is an absence of a unifying framework that standardizes interaction with time series data sets. We introduce a python-based time series formalism which provides a SQL style querying interface alongside a rich selection of time series prediction algorithms. Users can forecast data or impute missing entries using a specialized prediction query which employs learning models under the hood. The decoupled architecture of our framework allows it to be easily substituted with any SQL database. We show the functionality of our abstraction with a single machine implementation which will be a building block towards a scalable distributed platform for time series analysis.