Business Forecasting

ECON 8310 · University of Nebraska at Omaha

Preface

This is the book form of ECON 8310: Business Forecasting at the University of Nebraska at Omaha. It is not a summary of the course and not a replacement for it — it is the same material in a format you can read, search and quote from.

Every chapter is one lecture, and carries that lecture’s title exactly. Within a chapter, each section is a slide from the deck: the table, figure or definition the class saw, followed by what was said about it.

How this book is made

Nothing here is written twice. Each chapter is generated from the course’s own files — the Quarto mirror of the Beamer deck, and the recording script written for that lecture — by scripts/build_book.py. The consequence matters more than the mechanism: a number, a table or a figure in this book cannot disagree with the deck it came from, because it is never copied. It is read from the deck on every build.

That property is not decoration, and assembling this book is what proved it. Three lectures turned out to have quietly lost whole sections from their web versions — one of them down from five sections to two — which the parity check of the day could not see because it compared slide counts rather than structure. An audit running at the same time found reported figures that no committed code reproduced, a diagnostics table whose headline claim the reproducible numbers contradicted, and a result presented as a finding that was smaller than the noise around it.

Some of those had been true when they were written and had stopped being true. Others were never quite true. A book assembled by hand would have inherited every one of them, and as a fourth copy of every number it would have started drifting on its own the day it was finished.

What the course is about

One dataset, measured honestly, method after method.

Every model in this book is fitted to the same panel of thirty Walmart store-category series and scored on the same held-out block, so the numbers can be put beside one another. The result is not the one anyone predicts in week one: a 46-coefficient linear model finishes at the top, ahead of four neural architectures, and the two with the most attention behind them finish fifth and sixth.

The chapters explain why, from the data rather than from the reputation of the methods. That is the actual subject: not which model wins, but how to find out, and how to tell when a confident number has not earned its confidence.