Valedictory | Twenty Years On: How Is Forecasting Faring?

After nearly 20 years in editorial roles on the Foresight staff, Paul Goodwin stepped down at the end of 2024. In this issue, Paul leaves us with a farewell address on the current state and future direction of forecasting.

Learnings from the VN1 Forecasting Competition

In 2024’s VN1 forecasting competition, familiar methods like LightGBM and ensembling performed well, yet surprisingly few participants beat the naïve (no-change) model. Competition organizer Nicolas Vandeput shares his key takeaways from the top performers.

Decision Modeling to Increase Forecast Usability

Forecasts are not inherently beneficial, but can provide value by improving organizational decision making. To ensure this happens, James Taylor argues for a formal, robust, and structured decision model to identify relevant forecasts and their features during decision making. His approach is intended to make forecasters understand how their forecasts impact decision making.

The Trade-Offs between Forecasting Performance and Computational Cost

Forecasting performance is typically evaluated by statistical measures of forecast error, ignoring the computational cost of producing the forecast. Yet costs in both computer time and environmental impact can be huge. Foresight Associate Editors Fotios Petropoulos and Evangelos Spiliotis consider the tradeoffs, and show how forecast computation time can be dramatically reduced without significant impact on forecast accuracy.

Two-Part Forecasting for Time-Shifted Metrics

In the hospitality sector along with some others, the timing of an event’s occurrence (e.g., a hotel stay) is distinct from the timing of its initiation (i.e., making the reservation). This complicates the act of forecasting, which must now span multiple time axes. To address this challenge, new Foresight contributors Harrison Katz, Erica Savage, and Kai Thomas Brusch describe a two-part forecasting methodology that treats the forecasting process as a time-shift operator.

Retrieval-Augmented Forecasting: Bridging Human Insight and Machine Precision

Retrieval-augmented generation (RAG) techniques have enhanced the capabilities of large language models. Building on these advancements, Ryan Fattini and Ryan Young introduce a novel application of retrieval-augmented forecasting. By integrating natural language processing, this facilitates a conversational approach that enables users to generate and refine forecasts without the need for deep technical knowledge.

Opinion Editorials

  • The Mythical Influence of Metric Asymmetry It has often been noted that the asymmetry in some performance metrics (including MAPE) might encourage forecasters to “game” the metric by purposely over- or underforecasting. But is this really happening? Patrick Bower doesn’t think so, and in his opinion-editorial he argues that other factors have much larger influences on forecaster behavior than metric asymmetry.
  • Systems Thinking to Address Sustainability Leo Sadovy advocates a systems thinking approach for forecasters seeking to assist in sustainability challenges.

 

To receive this issue, start your membership to the IIF today! Join Now