📄 earth science

Verifying storyline attribution of extreme weather and climate events

This paper proposes and applies an objective, falsifiable methodology using historical meteorological analogues to verify storyline extreme event attribution results, finding that while machine learning-based estimates accurately capture global temperature influences on extreme heat in Southern and Central Europe, they are less effective for the Northern region.

Original authors: Noah Diffenbaugh, Jared Trok

Published 2026-09-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Noah Diffenbaugh, Jared Trok

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a heatwave sweeps across a continent, leaving cities sweltering and crops withering, a critical question arises: how much of this intensity is caused by the changing climate, and how much is simply the result of a particularly unlucky streak of weather? For decades, scientists have worked to answer this by comparing the world as it is today with a hypothetical world that never experienced human-induced warming. This field, known as extreme event attribution, tries to separate the signal of global warming from the noise of natural weather variability. However, verifying these answers has been notoriously difficult. Because we cannot travel back in time to see what would have happened if the same weather patterns occurred in a cooler past, researchers have lacked a way to independently test whether their stories about climate change are actually true. Without a method to check these claims against real-world observations, the field has relied heavily on the internal consistency of computer models rather than direct proof.

A new study by researchers at Stanford University offers a way forward, turning these theoretical stories into testable predictions. The team focused on a specific approach called "storyline attribution," which asks: if this exact weather pattern happened today, how much hotter would it be compared to if it happened in a cooler era? To verify this, the researchers did not look at the weather of a single day in isolation. Instead, they looked for "analogues"—pairs of events that shared very similar atmospheric patterns, like the shape of high-pressure systems hovering over Europe, but occurred in different decades. They took the most intense heatwaves from the recent past, between the years 2000 and 2024, and matched each one with a similar, though slightly less intense, weather pattern from the earlier part of the satellite era, between 1979 and 1999. By comparing these pairs, they created a natural experiment where the weather ingredients were nearly identical, but the background temperature of the planet was different.

The researchers then put a machine learning system to the test. This system had been trained to predict how hot a region would get based on the weather patterns and the current global temperature. The team asked the system to predict the temperature difference between the modern heatwave and its older analogue. They then compared these predictions to what actually happened in the historical records. The results were strikingly clear. For the regions covering Southern and Central Europe, the machine learning predictions matched the observed temperature differences with high accuracy. The system correctly identified that when similar weather patterns occur in a warmer world, the heat is significantly more intense. In fact, the correlation between what the model predicted and what the thermometers recorded was so strong that it suggested the model was capturing the true influence of global warming on these events.

However, the study also revealed where the method faces limits. For Northern Europe, the match between prediction and reality was much weaker when the researchers isolated the effect of global warming from the specific details of the weather patterns. While the model could still predict the general intensity of individual events, it struggled to consistently explain how much of that intensity was due to the warming planet versus the specific quirks of the local weather. This finding does not mean the model failed entirely, but it highlights that the influence of global warming is not uniform across every region and that some areas are more sensitive to the specific details of atmospheric circulation than others.

Crucially, the study found that the machine learning models tended to be conservative in their estimates. The models slightly underestimated the severity of the hottest events and slightly overestimated the severity of the older, cooler analogues. This means that when the models say a heatwave was made worse by climate change, the actual impact was likely even greater than the numbers suggested. The researchers also noted that the models did not fully account for regional factors like aerosols, which are tiny particles in the air that can influence heat. Despite these small gaps, the ability to test these predictions against independent historical data represents a major step forward. It moves the field from simply telling a plausible story about climate change to providing a framework where those stories can be proven or disproven by the evidence of the past.

The work confirms that the influence of global warming on extreme heat is not just a theoretical concept but a measurable reality that can be detected even when the weather patterns themselves remain similar. By pairing modern extremes with their historical look-alikes, the researchers demonstrated that the planet's rising baseline temperature acts as a multiplier, turning dangerous heat into life-threatening extremes. While the method is not perfect for every single region, it provides a robust tool for verifying the science of attribution. This approach allows scientists to move beyond asking if climate change is happening and toward answering exactly how much it is changing the weather we experience, offering a clearer, more reliable picture of the risks we face in a warming world.

Technical Summary: Verifying Storyline Attribution of Extreme Weather and Climate Events

Problem Statement
Extreme event attribution (EEA) has matured significantly, with methodologies for both probabilistic and storyline attribution becoming operational. While the underlying methodological elements of EEA—such as model fidelity, statistical paradigms, and inter-method comparisons—have been thoroughly evaluated, a critical gap remains: the independent verification of specific storyline attribution results. Storyline attribution quantifies thermodynamic contributions to event intensity conditional on specific meteorological ingredients. Unlike probabilistic attribution, which can be evaluated via out-of-sample threshold exceedance frequencies, storyline results are difficult to falsify because it is impossible to directly observe identical meteorological conditions in both the presence and absence of anthropogenic forcing. Consequently, there is a need for a framework that treats storyline EEA results as falsifiable predictions testable against independent observations.

Methodology
The authors propose a methodology to objectively test storyline EEA results by comparing recent extreme events (2000–2024) with "analogues" from earlier in the satellite era (1979–1999) that share similar meteorological patterns but occurred under different background climate conditions.

The framework utilizes machine learning (ML)-based EEA results previously published by Callahan et al., which employ convolutional neural networks (CNNs) trained on global climate model (GCM) ensembles. These CNNs predict daily regional-mean temperatures based on daily meteorological inputs (ERA5 reanalysis), calendar day, and annual global mean temperature (GMT).

The verification process involves four primary steps:

Event Identification: For each year from 2000 to 2024, the hottest 3-day summer event is identified for three IPCC regions: Northern Europe (Northern), Western & Central Europe (Central), and the Mediterranean (Southern).
Analogue Selection: For each recent event, the closest 3-day summer period from 1979–1999 is identified based on the pattern correlation of 500 hPa geopotential height anomalies. The search is restricted to June–August and allows for a calendar-day start date difference of ±28 days.
Difference Calculation: The study calculates the temperature difference ( $t$ ) between the recent event and its analogue. This is done for both the predicted values (using the CNN ensemble at actual GMT) and the observed values (using ERA5 reanalysis temperatures).
Isolation of Global Temperature Influence: To isolate the specific contribution of global warming from meteorological variability, the authors decompose the temperature difference ( $t$ ) into a meteorological component ( $t_{meteorology}$ ) and a global temperature component ( $t_{GMT}$ ). $t_{meteorology}$ is derived by predicting the event-analogue difference assuming a counterfactual GMT of 0°C. The residual ( $t_{GMT}$ ) represents the influence of global temperature after normalizing for meteorological differences.

The analysis code was developed and reviewed using a multi-agent AI workflow (Claude and Codex models) under human supervision to ensure rigor, with adversarial reviews conducted to identify and correct inconsistencies.

Key Results
The study applies this framework to the ML-based storyline attribution results for the three European regions:

Raw Correlation: There is a high correlation ( $r^2 = 0.851–0.892$ ) between predicted and observed event-analogue temperature differences ( $t$ ) across all three regions before normalization.
Normalized Correlation (Isolating GMT): After removing the effect of meteorological differences ( $t_{meteorology}$ $t_{m e t eor o l o g y}$ ) to isolate the influence of global temperature ( $t_{GMT}$ $t_{GM T}$ ):
- Southern and Central Regions: The correlation remains high ( $r^2 = 0.733$ and $r^2 = 0.617$ , respectively), indicating the EEA framework accurately captures the influence of global temperature on historical extreme heat events in these areas.
- Northern Region: The overall correlation is substantially reduced ( $r^2 = 0.259$ ), although agreement remains high for many individual events. This suggests the EEA results for this region do not fully capture the influence of global temperature, with raw differences being strongly driven by meteorological variability.
Bias Analysis: The predicted $t$ exhibits a persistent negative bias (underestimation of the difference). The authors note that the CNNs tend to slightly underpredict the severity of recent hottest events and overpredict the severity of analogues. This bias suggests that the ML-based attribution estimates are conservative. The relationship between $t$ and global temperature is also weaker in predictions than in observations, further supporting the conclusion of conservative estimates.

Significance and Claims
The paper claims to provide a pathway for verifying storyline EEA results within a framework of independent falsifiability. By pairing recent extreme events with historical analogues and controlling for meteorological differences, the authors demonstrate that ML-based EEA approaches can accurately predict the difference in event intensity when similar meteorological conditions occur under different background climates.

The authors emphasize that while their out-of-sample test does not definitively confirm the contribution of global warming to every individual event, it validates the methodology for evaluating storyline EEA results. The findings suggest that for the Southern and Central regions, the ML-based approach accurately captures the response of historical heat events to global warming. For the Northern region, the results highlight the necessity of carefully designing evaluations to isolate global temperature influences from meteorological noise. The study concludes that such verification is essential for establishing the evidentiary basis for claims of financial responsibility and for extending attribution to extreme event impacts.

More like this