When in doubt, just model it. Modelling uncertainty

There’s a pattern I’ve noticed across projects: when something is hard to measure, people tend to either ignore it or collapse it into a single number. Both choices feel safe. Both choices are wrong.

Probabilistic modeling exists precisely for this reason — it forces you to be honest about what you don’t know by representing it explicitly, as a distribution rather than a point estimate.

Uncertainty

Uncertainty shows up whenever you’re making predictions with incomplete information, which is always. In practice, this means the parameters of your model aren’t fixed truths — they’re estimates with some range of plausible values around them.

Take linear regression. A frequentist will give you a coefficient and a confidence interval. A Bayesian will give you a full posterior distribution over that coefficient. Both are capturing the same underlying reality: we fitted the model on a finite sample, and had the data been slightly different, our estimate would have shifted. The distribution is a way of making that explicit.

This matters beyond statistics. When you hand someone a single number — a predicted price, a risk score, a forecast — you’re implicitly claiming a precision you probably don’t have. A probability distribution is more honest, and in many applications (insurance, medical decision-making, financial risk) it’s also more useful. Knowing that the outcome is likely between X and Y, and that there’s a 5% chance it’s worse than Z, is a fundamentally different kind of information than knowing the expected value alone.

Deep Uncertainty

Regular uncertainty is manageable — you have a model, it has parameters, you can estimate them and quantify your confidence. Deep uncertainty is harder. It arises when you’re not even sure you have the right model structure, when the data-generating process might be fundamentally different from what you assumed.

This is common in complex systems: economic forecasting, climate modeling, anything involving human behavior at scale. The variables interact in non-linear ways, there are regime changes, feedback loops, and structural breaks that no parametric model cleanly captures. In these settings, running a Bayesian regression and looking at the credible interval doesn’t tell the whole story, because the uncertainty in the model specification itself is much larger than the parameter uncertainty within any given model.

A few approaches have been developed to tackle this:

Robust optimization finds solutions that perform acceptably across a wide range of possible model specifications, rather than optimizing under a single assumed model.
Scenario analysis sidesteps the need for a probability distribution entirely — instead of asking “what’s the probability of outcome X?”, it asks “what happens if X occurs?”.
Bayesian non-parametric models (Gaussian processes, Dirichlet processes, etc.) place priors over entire function spaces, letting the data determine the model complexity rather than fixing it in advance.
Bayesian deep learning — things like MC dropout, deep ensembles, or Bayesian neural networks — attempts to bring uncertainty quantification into neural networks, where the model class is expressive enough to handle complex structure but traditionally gives only point predictions.

The honest answer is that deep uncertainty is never fully “solved.” You can manage it better or worse, but any model is a simplification, and the further your data is from the training distribution, the less you should trust your confidence intervals.

Why it matters in practice

There’s a tendency to treat uncertainty quantification as an academic nicety — useful in papers, ignored in dashboards. I’d push back on this. The most consequential decisions are usually the ones made under the most uncertainty. A model that only returns a point estimate is actively hiding information from the people who need to make those decisions.

When in doubt about how to handle something messy or unknown, model it. Give it a distribution, make your assumptions explicit, and let the uncertainty propagate through to the output. The result will be less tidy, but it will be more truthful — and that’s what you actually need when things don’t go according to plan.