Abstract: In the realm of scientific inquiry, forecasting emerges as both a captivating domain of exploration and a pivotal methodology employed across industrial, governmental, and academic spheres. Illustrative examples encompass demand and inventory planning, electrical load prognostication, fiscal allocation for capital projects, pricing dynamics, machinery's predictive upkeep, macroeconomic projections, and supply chain predictions, among others.
The quintessence of forecasting often hinges on time series data, a form of data that is omnipresent, transcending the confines of data science and artificial intelligence disciplines. To elucidate, one might consider the monthly tabulations of initial unemployment claims, granular stock price fluctuations, corporate sales metrics on a weekly basis, daily pedometer readings from wearables, sensor-derived machinery performance metrics, and pivotal business function indicators, just to name a few.
Yet, it is imperative to recognize that time series data is distinct from cross-sectional data, primarily due to its inherent temporal dependence. This sequential linkage offers a conduit to extrapolate forthcoming data points in the series. In certain scenarios, amalgamating multiple time series can harness the interplay of temporal dynamics, thereby enhancing the forecasting precision. Predominantly, the analytical arsenal for time series data is enriched by methodologies stemming from statistical and machine learning paradigms. Recent advancements in time series analysis have also embraced the prowess of deep learning methodologies. Consequently, for any discerning data scientist, it is indispensable to be adept with time series statistical and machine learning algorithms.
This tutorial, spanning a duration of three hours, is designed to serve as a primer for novices in the domain of time series analysis. It endeavors to elucidate both the foundational statistical models and the “modern” techniques pertinent to time series forecasting. The tutorial will be bifurcated into two sections, each extending for approximately 1.5 hours with a 15-minute break in between.
The tutorial begins with several illustrative examples to motivate the significance of time series analysis, followed by formulating a time series forecasting problem to set the groundwork for the ensuing discussions. Prior to delving into statistical and machine learning modeling, it is crucial for attendees to acquaint themselves with the terminologies, rudimentary concepts, core principles, and essential data analysis techniques used in time series analysis forecasting. Subsequently, we introduce a cornerstone time series forecasting methodology that has proven its mettle over time and whose tenets have been integrated into modern forecasting frameworks. We shall elucidate these concepts further with empirical case studies. By the section’s conclusion, participants should have a comprehensive grasp of the prevalent terminologies, concepts, and methodologies in time series forecasting, enabling them to implement a tried-and-true statistical time series model.
Section 2 advances our understanding, building on foundational terminologies, principles, and methods of time series analysis. It delves into additional concepts and techniques from the deep learning domain that have recently garnered attention within the time series analysis community. This section offers a bird's-eye view of several contemporary forecasting frameworks, elucidating the underlying methodologies with intuitive explanations complemented by mathematical formulations. The relevance of these frameworks is highlighted through empirical examples, followed by a discussion on their technical implementation.
It's pivotal to understand that, while this tutorial is designed for beginners in time series analysis, the incorporation of mathematical expressions and terminology is essential. Time series analysis, at its core, is deeply rooted in mathematics, which serves as the fundamental language for the techniques we explore.
This tutorial is tailored for individuals unfamiliar with statistical and machine learning approaches to time series modeling, aiming to grasp the fundamentals of time series analysis. This encompasses data scientists, data engineers, software developers, and others who may not have a background in statistics and econometrics or extensive experience with statistical time series modeling.
Jeffrey serves as the Chief Data and A.I. Officer at Fanatics Collectibles, having transitioned from Amazon Music where he held the title of Global Head of Data Science, Analytics, and Engineering. Before Amazon, he was the VP of Data Science, Data Engineering, and Platform Engineering at Walmart Labs. Jeffrey's extensive experience also includes being the Chief Data Scientist at AllianceBernstein, managing assets over $550 billion, and leadership roles at Silicon Valley Data Science, Charles Schwab Corporation, and KPMG. An academic at heart, he has taught economics, econometrics, statistics, and machine learning at institutions like UC Berkeley, Cornell, NYU, University of Pennsylvania, and Virginia Tech. He holds a Ph.D. and M.A. in Economics from the University of Pennsylvania and an Executive MBA and a B.S. in Mathematics and Economics from UCLA.
Bio: Jeffrey Yau is currently Chief Data & A.I. Officer at Fanatics Collectibles. Most recently, he served as Global Head of Data Science, Analytics & Engineering at Amazon Music where he oversaw multiple teams who developed both insights-packed analytics and end-to-end statistical and machine learning systems. Prior to Amazon, Jeffrey worked at WalmartLabs as the VP of Data Science & Engineering where he led the team responsible for powering Walmart store mobile apps and the entire store finance system. Further, his team created end-to-end machine learning systems for key business initiatives and had a multi-billion dollar impact annually on Walmart U.S.
Over the years, he has held various senior level positions in quantitative finance at global investment management firm AllianceBernstein, consulting firm Data Science at Silicon Valley Data Science, multinational financial services company Charles Schwab Corporation, and the world's leading professional services firm KPMG. He began his career as a tenure-track Assistant Professor of Economics at Virginia Tech, and he was an adjunct professor at UC Berkeley, Cornell, and NYU, teaching machine learning and advanced statistical modeling for finance and business.