ECON 8310 · Business Forecasting
This page holds what the course is — what it teaches, what it assumes you already know, and which books it reads from. It is deliberately free of anything specific to one university, one term or one section: no dates, no grading scale, no room, no policies. Those live on Canvas for enrolled students.
Everything else on this site — the slides, the labs, the assignments, the book and the workbook — is the course material itself, and is the same in every offering.
The course develops practical forecasting skills for graduate students in business, economics and data science. It works through a broad suite of methods — from classical time series models to machine learning to Bayesian inference — applied to real business forecasting problems. Three things are emphasized throughout: implementation in Python, interpretation of model output for business decisions, and honest comparison of methods against each other and against a benchmark.
The first half covers traditional econometric forecasting; the second half turns to predictive analytics and machine learning.
On AI. This course is designed to fully integrate modern AI tools, in both teaching and learning. The instructor uses AI a lot, says so when he has, and asks you to do the same. The AI use policy sets out what that means in practice: any assistant is fine, every assignment carries a prompt budget and a prompt log, and the written interpretation is yours.
By the end of the course you will be able to:
(These outcomes were drafted with help from Claude.)
Familiarity with basic statistics — regression and hypothesis testing — and introductory programming.
No prior knowledge of machine learning or time series analysis is required. The course starts from the beginning on both.
All are free online. Nothing needs to be purchased.
The primary textbook exists in both an R and a Python version. The Python version is newer and has two extra chapters. You are highly encouraged to use Python — many basic forecasting tasks can be done in R if that is what you already know, but when the course reaches the more advanced techniques, Python is the only way through.
How the two editions line up. Chapters 1 through 13 carry the same numbers and titles in both, so for the first two-thirds of the course either edition works. Two differences matter:
That is the practical reason for the Python recommendation. Through Lecture 6 the choice of language is yours; from Lecture 7 onward the readings exist only in Python, and so does all of the course code.
FPP contains no tree-based methods and no regularization. Lectures 4 through 6 use:
ISLP is by four of the same statisticians behind the standard graduate reference in this area, it is written at this level, and every chapter ends with a Python lab. Chapter 8 covers decision trees, bagging, random forests and boosting together; Chapter 6 covers regularization. That is Lectures 4 through 6 almost exactly.
Neither book covers Bayesian methods. Lectures 10 through 12 use:
BMCP is written by PyMC core developers, and every example is PyMC and ArviZ code — the exact libraries this course uses. It is the closest thing to a purpose-built text for the last three lectures: Chapter 1 covers priors and prior predictive checks, Chapter 2 the MCMC diagnostics you will be asked to report, §4.5–4.6 pooling and hierarchical models, and §6.4.3 is Bayesian Structural Time Series by name.
A note on level. BMCP is mathematically heavier than FPP in places, and you are not responsible for the derivations. Read the assigned sections for the ideas and the code, run the examples, and treat the rest as reference. The slides define what you are accountable for; BMCP shows you the same ideas in working PyMC.
Per-week readings are in the weekly schedule on the course home page.
This course covers a wide range of methods — classical statistics, deep learning, Bayesian inference. Some weeks will be more familiar than others depending on your background, and which weeks those are differs from person to person. Work with each other so you can complement each other.
The methods build on each other. Falling behind in one week makes the next harder, and the effect compounds. Use office hours early rather than after you are stuck.
| Slides | One deck per lecture, published as interactive HTML rather than PDF — HTML carries real headings, alt text, table headers and readable maths |
| Labs | One executable lab per lecture except the synthesis, designed to be run alongside the deck. Each carries a suggested time — that is the classroom pace, not a target. Working through one on your own reasonably takes longer, and taking longer is not a sign you are behind |
| Assignments | Eleven submissions across seven assignments, each with its own prompt budget |
| The book | Every lecture as continuous prose — the slides plus what was said about them |
| The workbook | Every lab and assignment in one searchable volume, code shown rather than run |
The weekly schedule numbers the teaching weeks in order. It carries no calendar dates, because those belong to a particular term and section; enrolled students get the calendar, including any break weeks, from Canvas.