Business Forecasting

The Emerging Role of Artificial Intelligence and Machine Learning

Michael Gilliland, Len Tashman, Udo Sglavo

• Editorial: Wiley
• Año: 2021
• ISBN:
  9781119782582
• Edición: 1
• Categoría: Commerce
• Categoría: Projections et prévisions

Discover the role of machine learning and artificial intelligence in business forecasting from some of the brightest minds in the field

In Business Forecasting: The Emerging Role of Artificial Intelligence and Machine Learning accomplished authors Michael Gilliland, Len Tashman, and Udo Sglavo deliver relevant and timely insights from some of the most important and influential authors in the field of forecasting. You'll learn about the role played by machine learning and AI in the forecasting process and discover brand-new research, case studies, and thoughtful discussions covering an array of practical topics. The book offers multiple perspectives on issues like monitoring forecast performance, forecasting process, communication and accountability for forecasts, and the use of big data in forecasting.

You will find:

• Discussions on deep learning in forecasting, including current trends and challenges
• Explorations of neural network-based forecasting strategies
• A treatment of the future of artificial intelligence in business forecasting
• Analyses of forecasting methods, including modeling, selection, and monitoring

In addition to the Foreword by renowned researchers Spyros Makridakis and Fotios Petropoulos, the book also includes 16 "opinion/editorial" Afterwords by a diverse range of top academics, consultants, vendors, and industry practitioners, each providing their own unique vision of the issues, current state, and future direction of business forecasting.

Perfect for financial controllers, chief financial officers, business analysts, forecast analysts, and demand planners, Business Forecasting will also earn a place in the libraries of other executives and managers who seek a one-stop resource to help them critically assess and improve their own organization's forecasting efforts.

CHAPTER 1
Artificial Intelligence and Machine Learning in Forecasting

It is five years since publication of our initial collection, Business Forecasting: Practical Problems and Solutions in 2015. Since that time the forecasting landscape has undergone a major transformation, and is now dominated by the explosion of interest in the role of artificial intelligence (AI) and machine learning (ML).

These five years have been a very exciting time of experimentation – applying existing AI/ML methods to time-series problems, and research into creating entirely new or hybrid methods. Research endeavors such as the M4 (2018) and M5 (2020) Forecasting Competitions provide important data to help evaluate the key questions we need to ask:

• Will AI/ML fundamentally change the way we do forecasting?
• Will AI/ML fundamentally improve our forecasting performance (both accuracy, and our understanding of forecast uncertainty)?
• Can AI/ML address the psychological and process issues that so greatly impact the real-life practice of forecasting?

This last question is not the least important. The value of better forecasting is delivered through better decision making (and the resulting better outcomes). So solving the statistical side of forecasting, alone, does not solve the business forecasting problem.

This chapter begins with three somewhat technical discussions of ML and deep learning, including neural networks, with examples of their application in online retail and energy. These three, along with Kolassa's critical assessment of deep and machine learning, are meant to provide objective tutorials on the use of these methods in forecasting, without hiding the critical issues users will face.

The next two pieces (and their accompanying commentaries) revolve around a forward look at the impact of artificial intelligence by Spyros Makridakis, originally published in a five-part series in Foresight: The International Journal of Applied Forecasting. Here, we get competing perspectives on what AI will be able to deliver to affect our lives.

The next four pieces delve into other areas impacted by AI/ML, including implications for supply chain and the forecasting process. We know, for example, that the review and manual override of computer-generated forecasts can be a considerable drain on management resources, with overrides often resulting in a degradation of forecast accuracy. Chase and Baker illustrate alternative ways ML can augment the role of a demand planner by identifying which forecasts are likely to benefit from adjustment, and which should be left alone. This “ML-assisted demand planning” can both save the planner time and also result in more accurate forecasts.

This chapter ends with a recap of findings from the M4 competition – the key takeaways for practitioners. We encourage readers to find more complete M4 coverage in a special issue of the International Journal of Forecasting 36(1), January–March 2020, which includes 35 articles on results, analysis, and commentary.

* * *

1.1 DEEP LEARNING FOR FORECASTING*

Tim Januschowski, Jan Gasthaus, Yuyang Wang, Syama Sundar Rangapuram, and Laurent Callot

---

While the term deep learning (DL) wasn't widely used until the 2010s, the techniques it refers to have been in development since the 1950s, namely artificial neural networks (NN or ANN for short). DL has scored major successes in image recognition, natural language processing (e.g., machine translation and speech recognition), and autonomous agents such as Google Deep Mind's AlphaGo. It is often used as a synonym for artificial intelligence (AI), by which name it has received extensive press coverage.

Deep learning has the potential to make forecasting systems both simpler and more robust while improving forecast accuracy relative to “classical” approaches. This first of two installments from Tim Januschowski et al. presents a tutorial on the basics of DL with illustrations of how it has been applied for forecasting Amazon product sales and other variables.

---

---

“Fears about the implication of the ‘black box’ are misplaced: compared to human intelligence, AI is actually the more transparent. Unlike the human mind, AI can be interrogated and interpreted. Human intelligence is and has always been the real ‘black box.’”

– Vijay Pande, New York Times, January 25, 2018

---

Introduction

The forecasting case one would likely associate with Amazon is that of the demand for products available for purchase on its websites. The products are organized in a catalog that includes a wealth of metadata, such as product descriptions, images, and customer reviews.

The demand for many of these products is highly seasonal or driven by external events or internal decisions (e.g., price changes or promotions), and new products are continually added to the catalog. In addition, many products have sparse and intermittent demands, adding to the forecasting challenge.

Many of these problems are especially challenging for traditional forecasting methods such as ARIMA, exponential smoothing (ES), and linear models, and some even more advanced modeling techniques (Seeger, Salinas, and Flunkert 2016).

Limitations of the Classical Methods

Due to diverse data characteristics, successfully addressing the entire forecasting landscape with a single method is unrealistic. Models that operate on individual time series, such as ARIMA and ES, perform well in situations where the data exhibit clear, regular patterns and a behavior compatible with the structural assumptions of the model. The number of real-world forecasting problems that match this ideal case is limited. Rather, one is often faced with situations where individual time series do not provide enough information to identify predictable dynamics such as seasonality or the influence of causal factors. Common reasons for this are inadequate data history (i.e. new products), intermittency, and factors with weak signals.

The classical approach focuses on the time series that handle orderly data well, and to perform data preprocessing steps when needing to deal with irregularities. Preprocessing includes seasonal adjustments, Box-Cox transformations, and corrections for causal effects. The chain of preprocessing and modeling procedures to deal with the variety of data characteristics is ...