Advances in forecasting: Highlights from the 44th International Symposium on Forecasting 

For many data scientists and analysts, the first week of July was long blocked in the calendar. The 44th International Symposium on Forecasting (ISF), the leading event for forecasting professionals, was held in the picturesque city of Dijon. In this recap, we share our main takeaways.

Outstanding keynotes

The conference hosted several compelling keynotes, three of which stood out for us:

• Prof. Francis Diebold, renowned for co-authoring a widely used statistical test for comparing the predictive accuracy of forecasting models, delivered an intriguing talk on modeling Arctic sea ice levels. In his presentation, he estimated when we might expect an ice-free Arctic. Unfortunately, it’s sooner than we might have guessed.
• Prof. Sergei Guriev explored the economic impact of populists in power, highlighting how conflicts and sanctions can undermine economic growth.
• Prof. Mike West talked about Bayesian forecasting, demonstrating how Bayesian Predictive Decision Synthesis (BPDS) allows for the explicit integration of decision outcomes and raw predictive performance in model and forecast pooling, thereby incorporating a broader aspect of decision-focused forecasting.

Energy in the spotlight

The symposium featured 12 parallel thematic tracks from various domains, such as supply chain, finance, climate, and healthcare. Notably, two full parallel tracks were dedicated to the energy sector. These tracks covered price forecasting (real-time, day-ahead, and longer horizon), load forecasting, and the use of meteorological data.

Interestingly, researchers and practitioners in this field use similar features but vastly different techniques, ranging from simpler linear models to Markov Switching Regime Models and Deep Learning (including attention layers).

Insights into foundation models

Foundation models were, unsurprisingly, a hot topic. During ISF, authors of models such as Lag-Llama and Chronos described their approaches to the problem and what makes their models unique. Nixtla demonstrated TimeGPT’s scalability in forecasting billions of time series and showed a benchmark comparing the performance of current foundation models on common datasets.

Discussions followed on the future of foundation models for time series, including enhancements with contextual text information.

From model to business problem

The conference struck a balance between theoretical and practical content. While some talks focused on introducing new approaches or models aiming to become the next state-of-the-art, others showcased real-world applications.

The latter often highlighted how specific companies used existing models or approaches to solve their business problems. These sessions were particularly valuable, as out-of-the-box solutions rarely work straight away. Solving a business problem often requires a creative approach, rigorous data processing and cleaning, and sometimes using a less complex model to ensure interpretability for business stakeholders.

Highly recommended

It’s rare for so many forecasting professionals to be in one place at the same time. Thus, talks aside, the ISF is an opportunity to talk to like-minded peers and spark new ideas. You might even meet industry legends such as Prof. Rob J. Hyndman, the author of the #1 forecasting resource, “Forecasting: Principles and Practice.”

If you missed this excellent event, you can watch some keynotes and talks on the conference’s official YouTube channel.

For insights into forecasting at Dexter, keep exploring our knowledge hub with the related reads below this article.