July 30, 2025 | By Sam Fisher and Luke Lennon
On a hot and muggy Arkansas morning in July, I returned from a routine bike ride, surprised at how fatigued and dehydrated I felt. How hot was too hot for exercise? I decided to check out the “feels like” temperature from the National Weather Service.
But like any good data scientist, I started asking questions to ensure the model was helping me make optimal decisions.
In data science – especially in finance – we often assume that every widely-used model has been fit to and validated on a sizable sample of data from a real-world population. Yet to my surprise, the Heat Index had come to exist in a different way entirely – a discovery that underscored how important validation is when working with expert-driven models.
What is the Heat Index (HI)?
The Heat Index (HI) is the “feels like” or “apparent” temperature that takes into account both air temperature and relative humidity to predict how much stress a sweltering summer day will place on our bodies. Ever consider how dry, desert heat is sometimes more bearable than a muggy Arkansas day? That’s because high humidity hinders sweat evaporation, impairing the body’s ability to regulate temperature. The HI is a key input to crucial decisions such as heat warnings, school closures, and work safety regulations.
For decades, however, the index was accepted without traditional data-driven validation. How can a model with such wide applications, almost taken as fact, have gotten by on expert knowledge alone? And what does that mean for using expert knowledge in other high-stakes and regulated applications?
A Brief History of the Heat Index
The Heat Index was originally based on a physical model of human thermoregulation presented in physicist Robert Steadman’s July 1979 paper “The Assessment of Sultriness. Part I: A Temperature-Humidity Index Based on Human Physiology and Clothing Science.” The model assumed a 147 lb 5’7” adult wearing light summer clothes and walking in full shade on a day with a light breeze.
Then, for every combination of air temperature and humidity, Steadman solved for the apparent temperature. The objective was to achieve a more precise understanding of perceived heat or cold by incorporating humidity and wind’s effects on the body. Steadman’s table below illustrates the apparent (or feels-like) temperature for each combination.
Around the same time period, Florida meteorologist George Winterling brought the index into the TV news mainstream. This early use prompted the first adoption of the index by the National Weather Service (NWS) in 1979.
When we think of the Heat Index, however, we don’t just think of the apparent temperature – we think of heat warnings, or indications of when we face an elevated risk of heat-related illness.
These familiar risk categories are attributed to climatologist Robert Quayle and meteorologist Fred Doehring, who layered easily digestible descriptions on top of Steadman’s calculations. This was critical in allowing weather forecasters and the public to take action on HI values – so much so that Quayle and Doehring’s guide won official adoption by the National Weather Service and OSHA. The issue, however, was that Quayle and Doehring’s risk categories did not have any clear scientific basis. Instead, Quayle and Doehring categorized these based on their hypotheses regarding how Steadman heat index values correlated to heat-related illness risk.
It wasn’t until recently that the validity of this medical interpretation was highlighted in a high-stakes arena: a courtroom. In 2016 and 2017, a judge dismissed the majority of OSHA’s claims against the US Postal Service for exposing workers to dangerous heat-related conditions, in part because she found that there was no scientific basis for the Heat Index risk categories.
Physicists Yi-Chuan Lu and David M. Romps took the OSHA case as a wake-up call: a model so widely trusted had essentially skipped the usual step of empirical model validation. Lu noted that a previous regression model developed by the NWS to approximate Steadman’s model was often misapplied to predict the heat index at extreme temperatures not (yet) supported by Steadman’s model. This resulted in a dangerous under-estimation of the heat index.
In 2022, for the first time, Romps and Lu developed and validated an extended form of Steadman’s heat index model against controlled laboratory data collected by physiologists at Penn State. This model rendered sound predictions at extreme temperatures – a major breakthrough for public health and safety in light of climate change.
From Heat Index to Credit Risk: The Role of Expert Knowledge
This is all interesting meteorology, but what’s the relevance to finance?
The heat index teaches us that expert-driven modeling is not heresy – it’s often a necessity when data is limited and we have strong theoretical frameworks. Rather than making guesses or waiting years for data, modelers can use domain expertise and economics to build a reasoned model upfront, much like Steadman used physics to model heat stress.
It’s easy to assume that modern credit scoring and underwriting models are objective number crunchers. But in practice, especially in credit risk and lending decisions, there’s often a practical need to integrate expert judgment and data-driven model fitting. Banking regulators have formally acknowledged that models can include expert-based components (under SR 11-7).
In credit risk, for example, you might face a scenario where an economic shock has no precedent in your dataset, but experienced risk managers have a solid idea of what may happen. Or you might launch a new lending product with no historical data, and expert assumptions must guide the initial model until there’s enough data. Evaluations of portfolio risk incorporate the forward-looking perspective of economists on the interest rate environment in concert with computer simulations that explore many possible future rate trajectories.
Of course, the flip side is that expert-based models must be approached with humility. Not all expert knowledge is equally likely to hold up to empirical testing. Models grounded in sound economic theory or domain knowledge are likely to outshine opinions that lack such grounding. Strong assumptions can introduce bias, so models should be continually challenged and updated as real world data comes in – just as the heat index continues to be empirically refined decades later. The goal is to get the best of both worlds: use expert knowledge to inform models and use data to improve them.
How to use Expert-Driven Models: Key Takeaways
- Models aren’t always born from big data – The Heat Index was built on physical first principles long before it was validated by empirical data. This did not stop it from becoming widely trusted, useful, and accurate.
- Expert opinion can fill data gaps – In both weather and finance, there are situations where historical data is sparse or non-existent. Well-founded expert judgment (e.g. physical laws and economic theory) can guide model construction when data alone cannot.
- Widespread use doesn’t guarantee scientific proof – It’s easy to assume a ubiquitous model must have been empirically proven, but the blind acceptance of Quayle and Doehring’s Risk Categories demonstrate otherwise. Always be aware of a model’s origin and limitations. Never close your eyes and hope that a third party vendor’s model will work well for your borrower population and product, no matter how ubiquitous it is.
- Validate when you can – Relying on expert-driven models is often a practical necessity, but it should be coupled with a plan to test those models. As data becomes available (be it lab experiments for heat stress or performance data for a new credit product), use it to validate and recalibrate the model. This ongoing “trust, but verify” approach keeps models robust.
Don’t Sweat It, as Long as You Validate and Refine
For business leaders and model developers, the heat index story underscores our belief at Stratyfy: that the best models often blend the art of expert judgment with the science of data. Leveraging expert opinion in credit decision models can provide a head-start in automating and optimizing decisions, as long as we use theoretically sound logic, assess the risks of model failure, and validate or update our models as data becomes available.
Interested in more original insights from our team? Subscribe to our quarterly newsletter.