Introduction

This working paper examines the evolution and internal consistency of climate scenario projections produced by the Network for Greening the Financial System (NGFS). While these scenarios are widely adopted as common benchmarks for analysing physical and transition risks, the NGFS itself has raised concerns that some applications may exceed the technical precision of the outputs. The study focuses on examining the internal evolution of these scenarios across successive releases and the resulting implications for interpretation over time. It clarifies that NGFS outputs are exploratory and illustrative rather than predictive forecasts, depending on specific policy assumptions and model optimisation frameworks.

A conceptual framework of uncertainty in NGFS scenarios

The research identifies three primary dimensions of uncertainty: scenario, model, and vintage. Scenario uncertainty arises from differing policy assumptions, such as carbon-price trajectories and mitigation timing. Model uncertainty reflects structural differences across Integrated Assessment Models (IAMs), including technology representation and sectoral detail. Vintage uncertainty, a central focus of this study, arises from revisions to the scenario framework itself over time. This dimension is further decomposed into two components: revisions to initial conditions and recalibration across successive releases. Initial-condition uncertainty captures dispersion in reported historical statistics used to initialise models, while recalibration uncertainty reflects updates to model parameters and assumptions within a given model-scenario pair.

Data and methodology

The study analyses NGFS scenario data from vintages v2 to v5, covering four releases, five common climate pathways, and three core IAMs: REMIND-MAgPIE, MESSAGEix-GLOBIOM, and GCAM. Vintage v1 is excluded due to inconsistent scenario narratives and changing variable definitions. The dataset includes 254 variables extracted from the NGFS-IIASA Data Explorer. To ensure comparability across diverse variables, the researchers used a common normalisation process, indexing paths to a common 2020 base year. This methodology allows path-based uncertainty to be examined independently from revisions to starting levels. Global steel production is used as a primary illustrative case to demonstrate how these uncertainty dimensions can be isolated in practice.

Results

Empirical analysis reveals that dispersion in reported initial conditions is a first-order feature of NGFS outputs. In the near-term horizon (2020-2025), initial-condition uncertainty exceeds scenario uncertainty for nearly 80% of variables and exceeds model uncertainty for over 40%. For approximately 20% of variables, this uncertainty exceeds scenario uncertainty by at least a factor of twenty. Recalibration uncertainty also emerges as a significant source of dispersion, exceeding scenario uncertainty for 85% of variables at the five-year horizon and nearly half of variables by 2050. This implies that annual updates to a single model can generate as much dispersion as changes in policy narratives or structural differences between models. For sectoral variables like steel and cement production, recalibration-related divergence often dominates revisions to initial conditions.

Discussion

The paper proposes several hypotheses for the observed vintage dispersion. On the initial-condition side, differences may reflect changes in underlying data sources, sectoral mappings, or variable definitions. On the recalibration side, dispersion may arise from structural updates within a model family, such as changes in technology representation or numerical implementation choices. Interpretability challenges are noted for systemically important variables; for example, the divergence between steel and cement production pathways in certain scenarios is difficult to reconcile with historical construction practices. These findings suggest that NGFS outputs should be interpreted as evolving projections rather than fixed reference trajectories. Financial institutions and regulators are advised to explicitly account for initial-condition and recalibration sensitivity in their climate stress testing and supervisory exercises.

Conclusion

The study concludes that uncertainty in NGFS scenarios is fundamentally multi-dimensional. Revisions to initial conditions and annual model updates play a quantitatively important role alongside scenario narratives and inter-model differences. This is particularly pronounced over near-term horizons where financial decisions are most sensitive. Robust interpretation and prudent risk assessment require users to recognise that changes in projected outcomes may reflect shifts in the underlying information set rather than changes in policy ambition. Incorporating uncertainty bands that summarise empirically observed dispersion is recommended to provide a more transparent complement to scenario narratives and help users assess the robustness of their financial risk analyses.