Controversies in Zero-Shot Forecasting with Time Series Foundation Models and a Three-Layer Framework for Trustworthy Evaluation

Time series foundation models (TSFMs) aim to forecast new datasets without target-specific parameter updates. Assessing this capability requires scrutiny of how evaluation evidence is generated and interpreted. We present a semi-systematic review of 80 core papers on the evaluation of TSFMs and large language model (LLM)-based forecasting methods. The review comprises 68 papers from the initial search and 12 from a targeted supplementary search within the same January 2022 to August 2026 publication window. We organize the evidence around benchmark integrity, ranking robustness, and domain transferability. In the reviewed literature, empirical audits of overlap between TSFM pretraining corpora and test benchmarks remain limited. Protocol choices, including normalization, forecast horizons, information availability, and metric aggregation, can alter model comparisons. Domain studies report both competitive performance and limitations, depending on task formulation, the relevance of pretraining data, sampling granularity, and data quality. These findings warrant conditional interpretations of zero-shot results rather than general claims of success or failure. We propose a five-part checklist for trustworthy evaluation and identify research priorities in corpus auditing, statistical comparison, domain-specific calibration, broader empirical coverage, and model selection. Reliable deployment requires evidence relevant to the intended task and operating conditions.