Cold does not simply slow an operation down — it reshapes the entire distribution of outcomes. When we began modeling airlift capacity in sub-zero environments, the first thing the data made clear was that averages hide almost everything worth knowing.
Strategic airlift is one of those systems where the constraints are physical, well-documented, and yet routinely underpriced by the models that depend on them. Ground crews, hydraulic tolerances, fuel viscosity, and runway condition all respond non-linearly to temperature. A ten-degree drop that is trivial at one end of the scale becomes decisive at the other, and any forecast that treats the relationship as smooth will be confidently wrong exactly when it matters most.
Our approach starts from the observation that throughput — sorties completed per operating day — is better understood as a censored variable than a continuous one. Below certain thresholds, activity does not degrade gracefully; it stops. Modeling that discontinuity directly, rather than smoothing over it, is what separates a useful estimate from a comfortable one.
Why the tails carry the signal
Most of the value in this kind of forecast lives in the worst five percent of days. Those are the days that determine whether a supply chain holds or fails, and they are precisely the days that a mean-based model treats as noise. By fitting the conditional distribution of throughput against observed weather, we recover a picture where the extremes are first-class citizens rather than residuals to be explained away.
The practical payoff is not a single number but a range with honest edges. A planner who knows the tenth-percentile outcome can stage differently, commit differently, and hedge differently than one working from an expected value alone. In our backtests, that difference compounded into materially better decisions across a full operating season.
From weather to decisions
The final step is the least glamorous and the most important: translating a probabilistic forecast into something a human being can act on before the weather arrives. We publish these findings not as finished answers but as instruments — tested against real conditions, revised in the open, and built to be argued with.
That is the whole premise of this publication. Predictive mathematics is only worth anything once it survives contact with the world it claims to describe.