In the previous article I argued that bird strike records are invaluable but biased, and that any credible forecasting effort has to model the reporting behaviour alongside the risk. This article is about the constructive half of that argument: what you can build once you take the data seriously.
The short version: fuse strike history with meteorology, and express the output not as alerts but as risk windows. These are bounded, contextual, explainable statements about when and where risk is elevated.
Why weather is the right partner for strike data
Bird activity is substantially weather-driven. Frontal passages shape migration timing. Wind direction and strength influence soaring and commuting behaviour. Temperature affects insect availability, and with it the feeding activity of the species that matter most around aerodromes. Visibility and precipitation change flocking and roosting patterns. Sunrise and sunset, modulated by season and cloud, anchor the daily activity cycles that put birds and aircraft in the same airspace.
From a modelling perspective, meteorology has three properties that make it the ideal complement to strike records. It is measured continuously and consistently, free of the reporting biases discussed in article one. It is available everywhere aircraft operate. And, decisively for a forecasting application: it is itself forecast, with usable skill, days ahead. A model that learns the relationship between weather conditions and strike occurrence inherits the forward reach of the weather forecast.
Strike history contributes what weather cannot: the local, site-specific record of where aircraft and wildlife actually meet. Habitat, attractants, species communities, and operational patterns differ enormously between aerodromes; two airports with identical weather can carry very different risk. The historical record, treated with the care it demands, encodes that local structure.
What a risk window is
The output format matters as much as the model. Operational teams do not need another binary alarm. Alert fatigue is a well-understood failure mode in every safety domain, and a system that cries wolf gets switched off.
A risk window is a different kind of statement: for this location, over this time band, under this assembling set of conditions, strike risk is elevated relative to baseline, with a stated confidence level and the drivers visible. Tuesday 05:30–09:00, approach corridor, elevated: post-frontal tailwind conditions, peak seasonal activity for the locally dominant species group, historical clustering in comparable conditions.
Expressed this way, the forecast becomes something an operation can actually use. A wildlife team can schedule patrols and dispersal against it. An operations team can weigh it alongside the dozen other factors they balance every morning. A safety team can log it, audit it, and evaluate it after the fact, because a bounded, explained prediction can be scored against what happened, and a system that can be scored can be held accountable.
Explainability is the price of admission
Aviation safety runs on evidence, audit, and challenge. A model that outputs a bare number, however accurate, cannot participate in that culture. When a safety review board asks "why did the system flag Thursday morning?", the answer cannot be "the neural network said so."
This is why every Avigilance forecast carries its drivers and its confidence as first-class outputs, not as an afterthought. It is also why we are open about limitations: forecast skill varies with data density; low-reporting aerodromes carry wider uncertainty; rare species-specific events are harder to anticipate than aggregate activity. A tool that states its limits is a tool a safety manager can defend to a regulator, and one that quietly hides them is a liability regardless of its accuracy.
We are currently validating this approach with real operational data from a major Latin American carrier, testing forecasts retrospectively against held-out strike records and refining the feature set with practitioners who manage this risk daily. The early lessons, including the humbling ones, are shaping the platform more than any design document did.
The conversation we want to have
On Wednesday 5 August at 09:00 I’ll present this methodology in full at the aviation wildlife management conference in Cleveland: session CS1A-P7, "From Reports to Risk Windows." It is a first industry look, in front of the community whose scrutiny matters most, and the question-and-answer time matters to us as much as the talk.
If you manage wildlife risk at an aerodrome, run safety for an operator, regulate this space, or research it: we want to hear what a useful forecast would look like in your world. Meet us in Cleveland, or book a Teams call. Both start here: https://www.avigilance.com/wildlife#register