Zero-shot foundation models for forecast-based decision support: Accuracy, calibration and decision costs on port weather and CO₂ emissions to 2026
Tahmine dayalı karar desteği için sıfır atışlı temel modeller: 2026'ya kadar liman hava durumu ve CO₂ emisyonlarında doğruluk, kalibrasyon ve karar maliyetleri
This paper is under review at a peer-reviewed journal. The journal name and full text will be added after acceptance.
Abstract
Time series foundation models produce probabilistic forecasts without training and could simplify forecasting in decision support systems, but independent evidence on whether they improve decisions is scarce. This study evaluates four foundation models (Chronos-Bolt in two sizes, Chronos-2 and TimesFM 2.5) against five statistical methods on 470 monthly series: weather at 60 ports (wind, temperature, precipitation and solar irradiance, to September 2026) and fossil CO₂ emissions by country and sector (to August 2026). In a rolling-origin design with twelve-month horizons, 3,302 series–origin pairs are evaluated on point accuracy, quantile loss, interval calibration and the expected cost of newsvendor-type decisions, with rank-based tests, conformal recalibration and a robustness check on targets that postdate all model releases. The foundation models ranked first in both families. Chronos-2 had the lowest scaled quantile loss for port weather, 5% below exponential smoothing and 9% below climatology, and TimesFM 2.5 had the lowest loss for emissions, 9% below the seasonal naive method, which nevertheless gave the most accurate point forecasts. Gains over the best statistical method were small for wind, precipitation and solar irradiance, but the ranking persisted for post-release targets. Raw 80% intervals were miscalibrated for several methods, including Chronos-2 for weather (74% coverage); split conformal recalibration restored coverage to 78–82% for every method at almost no cost in sharpness. The best method depended on the cost asymmetry of the decision. For decision support systems, foundation models are a strong default when combined with conformal recalibration, simple benchmarks as monitors and data-quality screening.
Keywords
Time series foundation models; Probabilistic forecasting; Conformal prediction; Forecast evaluation; Decision costs; Port operations; Carbon emissions
Open code and data
Loading code…