Although time series forecasting is the part of broad family of predictive analytics and supervised learnings methods, it's also quite different from most commonly known Machine Learning techniques. Historically, times series were forecasted mainly by simple univariate methods like Exponential Smoothing or ARIMA models. Nowadays, the abundance of data and methods encourage us to use more advanced multivariate methods known very well from machine learning domain. However, forecasting time series differs in a few significant aspects. Firstly, times series data differ from cross-section data, which is used usually in Machine Learning applications, due to non-stationary and non-IID (independent and identically distributed) properties of data generating process. Secondly, while Machine Learning problems are usually concerned with the classification problems, times series forecasting is focused more on regression problems, with the increasing importance of not only point estimates themselves, but also the distribution and confidence intervals of forecasted phenomena. Thirdly, seasonality is the another inherent aspect of times series, rarely encountered in cross-section and machine learning problems. Therefore, modern forecasting of time series requires data scientists to take advantage of two analytical worlds, i.e. (1) classical univariate time series toolbox accounting for peculiarities of times series data and (2) Machine-Learning-based models. Being fluent only in one toolbox will result in leaving money on the table. The aim of talk is to provide you with the overview of time series forecasting toolbox and snippets of R code enabling you to tackle some typical challenges faced while modelling times series data. Secondly, the talk will highlight some recent advances made in times series research.
A couple of words about our presenter:
Mateusz Zawisza is a Senior Analytics Advisor to McKinsey & Company, where he's responsible for: 1) designing and implementing analytical use cases, 2) coaching & providing trainings on analytics and 3) supporting & advising on recruitment of analytical talent. He has served Polish & international clients from sectors of: retail, telecommunication, financial services & production, where he levered analytics to support day-to-day decisions made within functional areas of: Sales & Marketing, Operations and Risk. He's a PhD candidate & lecturer at Warsaw School of Economics as well as the co-author of analytics book titled: "Receptury w R". Mateusz is especially interested in the translation of analytical calculations into actual decisions.