How Firms Adapt to Physical Climate Risk: New LLM-Based Evidence from 13,500 US Companies

Extreme weather events have caused over $2 trillion in damages in the United States alone over the past two decades, and both frequency and severity are expected to increase. Yet while researchers have invested heavily in measuring firms’ exposure to physical climate risks, much less has been known about what firms actually do to adapt. A new working paper by Tobias Schimanski (University of Zürich / ETH Zurich) changes that.

The paper, “Firm-level Climate Change Adaptation” (April 2026), introduces the first comprehensive, time-varying measure of corporate climate adaptation at scale—covering 128,860 Form 10-K filings from over 13,500 US public companies between 2003 and 2025. Using a Large Language Model-based classification pipeline, the author processes 255 million paragraphs to track how firms prepare, restructure, insure, and financially buffer themselves against physical climate hazards.

Why This Gap Matters

Research on physical climate risk has, until now, focused primarily on exposure: keyword and machine learning methods that scan earnings calls and disclosures to quantify how much firms discuss climate hazards (e.g., Sautner et al., 2023; Li et al., 2024), and geospatial methods that link firm locations to storm tracks or flood zones (Kruttli et al., 2025). These approaches tell us who is in harm’s way—but not what firms are doing to protect themselves.

This distinction matters for investors, regulators, and risk managers. Exposure and preparedness are fundamentally different dimensions of climate risk. A firm located in a hurricane corridor with hardened infrastructure faces a different risk profile than one with the same location and no preparation. Prior evidence from voluntary CDP disclosures hints that firms react to higher expected climate risks with more adaptation, but voluntary, self-selected samples limit what can be inferred about the broader corporate sector.

The Framework: Five Dimensions of Adaptation

Schimanski develops a three-step hierarchical classification framework grounded in international definitions from the IPCC, ISO, UNFCCC, and UNEP FI.

Step 1 establishes a broad, inclusive definition: climate adaptation encompasses any action or solution that reduces exposure or vulnerability to physical climate risks, or increases a firm’s resilience and adaptive capacity.

Step 2 distinguishes between:

• Adaptation actions — measures a firm takes to protect itself (adjustments to operations, assets, governance, financial buffers)
• Adaptation solutions — products, services or technologies a firm provides to help others adapt

Step 3 disaggregates adaptation actions into five specific categories:

This structure mirrors the progression from narrow, tangible preparedness (physical protection) to broader, more organizational forms (risk assessment)—capturing the full spectrum of how firms respond to climate threats.

The Classification Pipeline

To operationalize this framework at scale, Schimanski uses a teacher-student training paradigm:

1. Framework definitions are translated into structured classification prompts
2. A sample of 1,000 paragraphs is manually annotated by the author as a benchmark dataset
3. GPT-4.1 (the teacher model) is validated on this benchmark, confirming reliable classification performance (precision of 87% for the adaptation classifier)
4. GPT-4.1 labels a corpus of 6,000 paragraphs to create training data
5. Qwen2.5-7B-Instruct is fine-tuned as the open-weight student model—orders of magnitude cheaper and fully reproducible

The full classification pipeline applies four layers to each paragraph from a 10-K filing:

1. Pre-filter for physical climate risk relevance (using existing classifiers)
2. Adaptation classifier — does this paragraph discuss adaptation?
3. Action vs. solution — is the firm adapting itself or helping others adapt?
4. Category classifier — which of the five adaptation action types applies?

An additional classifier (Schimanski et al., 2026) links adaptation to specific extreme weather types: storms, floods, heatwaves, droughts, wildfires, and cold waves.

All fine-tuned models, training data, and the manually annotated test dataset are openly available at huggingface.co/climate-adaptation.

What the Data Shows

Coverage and Intensity

In any given year, 19.6% of firms mention adaptation to physical climate risks in their 10-K filing. Over the full 22-year period, 28.4% of firms mention it at least once.

Among adaptation actions, risk transfer is by far the most commonly discussed (mean: 0.192 paragraphs per 1,000), followed by risk assessment (0.068) and adaptive operations (0.033). Physical protection is discussed least (0.014 per 1,000)—yet, as we will see, it turns out to be the most financially relevant.

Adaptation disclosures rise noticeably after major extreme weather events. A marked increase followed Hurricane Katrina; further upticks align with periods of high cumulative NOAA damage estimates. Notably, there is a sharp rise in 2025—potentially reflecting a broader shift toward physical climate preparedness as transition policy emphasis declined.

Industry Patterns

The leading industries for adaptation actions are:

• Insurance carriers (highest mean adaptation score)
• Agricultural production and forestry
• Electric, gas, and sanitary services (utilities)
• Hotels and lodging

Industries adapt through different mechanisms: agriculture and forestry lead on physical protection, while insurance dominates risk transfer and financial reserves. Construction and engineering companies rank highly in providing adaptation solutions to others.

Adaptation Is Primarily Firm-Specific

A variance decomposition reveals that time effects explain less than 1% of adaptation variation. Industry effects account for 5%–18%, depending on the measure. The overwhelming share—79% to 93%—reflects heterogeneity at the firm level.

Moreover, around 58% of this firm-level variation is persistent over time, meaning some firms consistently maintain higher adaptation levels while others consistently lag. The dynamic component is largest for physical protection and adaptive operations—the categories most likely to respond to new events and information.

Measurement error analysis (following Hassan et al., 2019) estimates that only 4%–15% of total variation is attributable to noise. The adaptation measures capture economically meaningful differences.

Application 1: Does Pre-Event Adaptation Mitigate Hurricane Losses?

The first economic application links firm-level adaptation to cumulative abnormal stock returns (CARs) around US hurricane landfalls, using the establishment-level exposure methodology from Kruttli et al. (2025).

Exposed firms experience significantly negative CARs following hurricanes—confirming that physical climate events impose real financial costs. The central question is whether prior adaptation dampens these losses.

The answer is yes—but specifically through physical protection:

A one-standard-deviation increase in physical protection before a hurricane increases CARs by approximately 0.5 percentage points over a 20-day window, offsetting roughly 26% of the negative stock price reaction.

The effect grows over longer horizons and is especially strong for:

• Smaller firms, where physical protection can offset up to 37% of losses
• Industries with large physical footprints (real estate, manufacturing, utilities)
• Firms with higher prior physical climate exposure, where investors appear to value a broader set of adaptation actions

Other adaptation categories—adaptive operations, financial reserves—matter for larger, more diversified firms where operational flexibility is valued by investors. Risk assessment, the broadest and most shallow category, has no significant mitigating effect on its own.

The interpretation is intuitive: physical protection is the most direct, tangible, and verifiable signal of genuine preparedness. Investors appear to treat it as credible evidence that a firm can withstand a hurricane—while other adaptation forms are priced selectively, depending on firm characteristics.

The Credibility Problem for Constrained Firms

For financially constrained firms, the mitigating effect of adaptation largely disappears. Investors appear to discount adaptation disclosures from firms that may lack the resources to actually implement or sustain protective measures. This points to a credibility channel: markets reward adaptation only when the firm appears able to follow through.

Application 2: Financial Constraints Block Adaptive Action

The second application directly examines how financial constraints affect firms’ adaptation behavior.

Using the Whited-Wu index as the measure of financial constraints, the paper finds a strong negative association between constraints and adaptation actions:

• A one-standard-deviation increase in financial constraints is associated with a 4%–14% reduction in overall adaptation actions relative to the sample mean
• The largest effects are for physical protection (9%–19% reduction), adaptive operations (12%–21%), and financial reserves (8%–23%)
• Risk transfer and risk assessment show smaller but still significant reductions (3%–13% and 5%–11%)

These effects are robust across specifications, controls, and alternative constraint measures (SA index).

Post-Hurricane Underreaction

Using a stacked difference-in-differences design around hurricane landfalls, the paper shows that:

• Exposed firms increase adaptation actions by 10%–11% relative to the sample mean following hurricanes—suggesting firms respond to realized shocks by adapting (or at least disclosing existing adaptation)
• But financially constrained firms respond significantly less: a one-standard-deviation increase in financial constraints offsets roughly one-third of the post-event increase in adaptation actions

This is a materially weaker response precisely when adaptation becomes most relevant. The interpretation is not merely that constrained firms have lower baseline adaptation levels—they also fail to adjust dynamically when new risk information arrives. Financial frictions actively limit adaptive capacity.

This has a systemic implication: financially constrained firms may amplify the economic impact of physical climate shocks. As extreme weather events become more frequent and severe, this group of underadapters could become an increasingly important source of systemic risk in financial markets and supply chains.

Related initiatives at Climate+Tech

This line of work aligns with several active Climate+Tech initiatives:

Climate Risk Intel — Our open climate risk tool hub and stakeholder collaboration work connects practitioners with datasets and tools for physical risk and resilience. Firm-level adaptation measurement and disclosure analysis complement that mission: both help move from abstract exposure to actionable preparedness.

The Adaptation Exchange — Climate+Tech is a founding partner of this cross-sector platform that translates climate risk awareness into coordinated resilience investment. The firm-level evidence in this paper—showing that physical protection measurably mitigates hurricane losses, and that financially constrained firms systematically underadapt—directly informs the platform’s business case and resilience standard work.

AI Benchmark for Sustainability Report Analysis — Both initiatives emphasize transparent, reproducible NLP on regulated corporate text, and open artifacts (models on Hugging Face, annotated benchmarks) so performance claims can be checked rather than taken on trust.

Contributions and Context

The paper makes three main contributions:

To the climate adaptation literature: It provides the first scalable, comprehensive firm-level measure covering the full depth and breadth of adaptation—not just reactive behavior to realized shocks, but pre-emptive preparation across five distinct categories. This allows systematic comparison across firms, industries, and time going back to 2003.

To the physical climate risk literature: By combining exposure and adaptation measures, the paper enables a fuller assessment of firms’ preparedness. Adaptation is shown to meaningfully mitigate financial losses—but only when credible.

To NLP in finance: The paper translates a complex, multi-dimensional taxonomy into an open-source LLM classification system that outperforms both keyword-based and keyword-discovery approaches. The teacher-student pipeline ensures scalability and reproducibility without dependence on proprietary APIs.

Access the Research

The paper “Firm-level Climate Change Adaptation” by Tobias Schimanski (University of Zürich) is available on SSRN (April 2026, 70 pages).

All models, training data, and the human-annotated benchmark dataset are openly available:

• HuggingFace: climate-adaptation — fine-tuned Qwen2.5-7B-Instruct classifiers, training and test data

The underlying firm-level adaptation dataset covers over 13,500 US public firms from 2003 to 2025 and is constructed from SEC EDGAR Form 10-K filings.

---

This summary is based on the working paper version posted on SSRN in April 2026. The paper was funded by the Swiss National Science Foundation (Grant Agreement No. 207800).