Balancing the Accuracy and Stability of Forecasts
It is expected that forecasts will change when they are updated with new information, such as the latest sales. But frequent, and especially drastic, forecast changes can create havoc in the planning process. While forecast accuracy grabs most of the attention, researchers have given less thought to the topic of forecast stability. In this paper, Christoph Bergmeir presents the relevant takeaways from research he and his co-authors recently published in the International Journal of Forecasting. Bergmeir discusses two complementary notions of stability – vertical and horizontal – and studies how they interact with accuracy.
Special Feature: Energy Forecasting
- An Introduction to Solar Forecasting
Dazhi Yang provides a comprehensive introduction to solar forecasting, a critical tool for integrating solar power into the electrical grid. He asserts that accurate forecasting relies fundamentally on physical models like numerical weather prediction and satellite image advection to predict cloud and atmospheric conditions, rather than purely data-driven machine learning approaches. The solar forecasting process is framed around five essential aspects: generating irradiance forecasts, converting them to power output, statistical postprocessing, rigorous verification, and aligning with grid regulations.
- Forecasting the Monthly Peak Day
Accurate forecasts of monthly peak load days are critical for power companies to manage costs, improve reliability, and support load control programs. Shreyashi Shukla, Tao Hong, Claude Martin, and Shahab Afshar compare three methods to predict whether tomorrow will be the monthly peak day. The first two traditional methods rely on monthly peak load forecasts, while the proposed method, a local search algorithm, identifies peak days without forecasting the magnitudes of monthly peaks.
Long-Tail Demand Forecasting in E-Commerce
Forecasting e-commerce demand is challenging with traditional methods because most products sell sporadically and exhibit sparse, intermittent demand. Even though machine learning models can capture complex patterns with such data, their black-box nature makes it difficult for planners to understand how a given forecast was structured. To bridge this gap, Shane Catts, Xuetong Wang, Bhargav Jairam Shetgaonkar, and Shubhankar Ray have developed DRUM: a Decomposable Retail Unified Model for long-tail demand forecasting.
Book Review
Foresight Deputy Editor Stephan Kolassa reviews Marco Peixeiro’s new book Time Series Forecasting Using Foundation Models. He finds it an immensely practical, hands-on introduction to the topic that shows an impressive understanding of forecasting-specific challenges often overlooked in the machine learning community.
Essay
While data is now being collected from seemingly everywhere, Ira Sohn expresses concern about the quality of global economic data and its impact on forecasters. In this essay, Ira documents examples of political meddling and other challenges to official government data.
Opinion-Editorial
This Op-Ed is a concise list of rules for time series forecasting from Skander Hannachi. As a reminder of good practices, these 13 points should be tacked on the wall next to every working forecaster.
To receive this current issue, download here.
