How Investors Can Evaluate AI Heat-Risk Models for Company Valuations

• Heat-risk models must connect physical exposure to company valuations. Investors need to understand how extreme temperatures affect production, operating costs, margins, and cash flow—not simply where a company faces heat exposure.
• AI can improve heat-risk estimates, but model reliability remains a challenge. Differences in data, assumptions, and operating sensitivities can produce conflicting results, making transparency and validation essential.
• Company-level heat disclosures are critical to accurate valuations. Investors should demand measurable operating thresholds, financial sensitivities, and adaptation costs to assess whether heat-related losses are reflected in company value.

In an earlier article, I argued that heat can reduce a company’s economic capacity without destroying its assets, and that markets do not consistently price that loss. In “Can AI Help Investors Price Heat Risk into Company Valuations?”, I argued that investors need a company’s heat-response function: an estimate of how production, costs, margins, and cash flow change once relevant temperature thresholds are crossed.

Machine learning may help estimate that relationship. The question is, what would make the estimate credible and usable in valuation?

AI is not starting from scratch. Banks, insurers, asset managers, and specialist vendors already use climate scenarios, geospatial data, catastrophe models, and vulnerability functions to estimate physical losses at the asset level. But providers can produce very different results for the same company or asset because their data, assumptions, and methods differ.

Linda Hain, Julian Kölbel, and Markus Leippold compare six firm-level physical climate-risk scores and find substantial divergence, including among scores based on similar methodologies. A benchmarking exercise conducted by the Global Association of Risk Professionals for the UK Climate Financial Risk Forum compared thirteen vendors assessing the same portfolio of properties. It found wide differences in hazard and damage estimates and little standardization in heat-related metrics.

The issue is not only the size of the estimate. Providers may also be measuring different parts of the problem.

The GARP exercise treated heat mainly as a property problem. It considered air-conditioning costs and possible building-system failures but excluded worker productivity, because it’s harder to measure. Yet heat often reaches cash flow because workers produce less or equipment runs less efficiently, even when the building remains intact. A model can therefore measure the property correctly while missing the company’s main financial exposure.

AI could use more detailed data to find relationships that conventional models miss. It could also scale the black-box problem by producing outputs whose assumptions investors cannot test. The test is whether investors can trace the path from heat to company value.

A credible company-level estimate requires five connected stages:

Hazard → asset exposure → operating sensitivity → financial effect → valuation

A model may have accurate temperature data but miss an exposed facility. Or it may identify the facility but lack details about cooling, production, working conditions, or backup capacity.

The model must also distinguish heat from other causes of weaker performance. Lower production or margins may reflect heat, but they may also reflect weaker demand, maintenance problems, equipment failure, labor disruption, or higher input costs. A hot period followed by weaker results does not prove heat caused the decline. The model must estimate how much of the change is linked to heat. Even an exact-looking number may be unreliable if the data are incomplete or that link is weak.

Valuation introduces another risk: double counting. Expected heat-related losses should normally be reflected in cash-flow forecasts, not counted again through the discount rate. A higher required return may still be justified for systematic risk not already reflected in expected cash flows. Expected loss and residual risk must remain separate.

Hot river water and low river flows can require EDF’s nuclear plants to reduce output to comply with environmental limits on cooling-water discharges.

EDF’s latest figures compare annual nuclear production losses related to environmental conditions of 0.3% from 2000 to 2020 with an estimated 1.5% by 2050.

Those averages hide when losses occur and how concentrated they can become. During the late-June 2026 heatwave, EDF stated that approximately 5.5 gigawatts, or 8.7% of installed French nuclear capacity, was temporarily unavailable.

The financial impact can exceed what the lost electricity suggests. If output falls when demand is high, EDF may lose valuable sales or replace the missing power at a higher cost.

A company-level model should estimate how site-specific conditions, including river flows, cooling capacity, reactor design, and environmental limits, affect output. It should then translate the production effect into revenue, replacement-power costs, margins, and cash flow.

The model should also evaluate adaptation. In early 2026, EDF submitted its first detailed adaptation plan to the public authorities, covering planned spending through 2040 for its nuclear, hydropower, and island operations in France. For investors, adaptation is a capital-allocation test: does the present value of the future cash flows protected by adaptation exceed the present value of its cost?

EDF reports past production losses, a forward estimate, and an adaptation plan. But investors still cannot see how each plant’s output changes as heat and river conditions worsen. Without that response, they cannot translate heat into company cash flow with confidence.

Companies need to disclose that operating response. Investors already receive sensitivity disclosures for foreign-exchange rates, interest rates, and commodity prices. Companies with material heat exposure could provide similar information.

The disclosure could identify:

• the temperature or operating threshold that matters
• the facilities, activities, and dependencies covered
• the effects on production, revenue, operating costs, capital spending, and cash flow
• the expected effect of adaptation measures
• the main assumptions and remaining uncertainty

For example, a company could estimate production and free cash flow under mild, severe, and extreme heat scenarios, before and after planned adaptation spending. Investors could then discount each scenario’s cash flows and weight the resulting values by their assigned probabilities.

Such disclosure would not predict which scenario will occur. It would show how production, cash flow, and company value change under each one.

It would also give AI models better company-specific inputs. Where historical data exist, models could compare the stated sensitivities with actual outcomes during past heat events rather than infer the entire relationship from external data.

Company disclosure is only the starting point. Investors should ask:

1. Which assets and operating dependencies are included, and what heat thresholds matter to them?

2. How does the model separate heat from other causes of weaker performance?

3. How are expected heat effects reflected in cash flow, and what risk or uncertainty remains?

AI can estimate operating effects, generate scenarios, suggest probabilities, and perform the valuation calculations. But these outputs depend on choices about the scenarios, probabilities, and required return. Investors must test and approve those assumptions rather than accept the final number.

AI’s value lies in combining more data and capturing company-specific relationships that simpler models miss. The result is useful only when investors can trace heat to cash flow and cash flow to value.

Where that path remains hidden, a more advanced model does not solve the problem. It gives uncertainty a number.