Note: the views and opinions expressed in this essay are those of the author and do not reflect any official policies, positions, or commitments of the relevant institutions or employers.
A Never-Ending Stream of Haikus
Most people haven’t heard of a TAF. Yet most of us have benefitted from them. If you’ve ever flown into or out of a major hub, your departure was informed by one. Here’s an example:
TAF EGBB 222302Z 2300/2324 07008KT CAVOK
TEMPO 2300/2308 BKN008
PROB30 2302/2307 4000 BR
TEMPO 2322/2324 BKN010=
This is a Terminal Aerodrome Forecast: a terse, textual prediction of near-future weather conditions, at a specific place, for a specific length of time, for a specific audience. TAFs are produced essentially 24/7/365 by Operational Meteorologists (OpMets) around the world, and they tell the story of the weather.
They’re like a never-ending stream of haikus. Except these poems directly inform aviation operations. The phenomena they encode and the thresholds they signal determine individual pilot actions, collective airport procedures, the unfolding of regional and national traffic patterns, and global aviation state.
Unlike a market, where volatility is somewhat generative, aviation occupies a safety-maximising paradigm. Predictability is the north star and perfect information is its nirvana. So imagine the world’s busiest airport, and a pilot coming in to land who can’t see more than ten metres ahead. If the system waits until that lack of visibility is manifest to respond, it’s too late: too dangerous, too expensive. Instead, air traffic control widens the gaps between landing aircraft ahead of time, based on a previously issued forecast, itself verified against observed conditions. The higher-order effects cascade forwards and sidewards: to ground crews, to airlines, to people dozing on airport seats, to other transport modalities, to supra-national bodies.
You may think: “Why can’t natural language be used to state what the weather will be?” We may get there in the future. But for now the terse, textual TAF reigns supreme. To see why – and to see what is actually at stake in changing how forecasts are issued – let’s make like Taylor Swift and go on a rapid tour through the eras.
The Eras Tour
Artisanal (1940s–50s)
TAFs didn’t start as formal, regulated products. They were akin to fireside tales shared amongst friends. The forecaster would walk the airfield, note the wind in the sock, feel the humidity on their skin, spy the shift of cloud over a nearby hill line, then talk to the aviator about to rattle down the runway.
The aim was not exhaustive description but a selective account that packaged meaning and minimised risk. When these oral communications went digital, the same functional constraints persisted: narrow channels, scarce compute, sparse observations, expensive storage. TAFs – compact, meaning-full, domain-specific textual encodings – occupied that evolutionary niche.
Scaling (1950s–70s)
Post-WWII aviation scaled and the artisanal arrangement fractured. Observations became systematic and instrumented: forecasting became a distinct act attributable to a node in a wider network, and the operational meteorologist emerged as the professional whose defining function was to issue a forecast the system could stand behind.
In tandem, aviation went transnational. Open observation networks, shared telecommunications standards and cross-national data exchange made a forecast issued at one aerodrome legible and actionable thousands of miles away. This demanded stability. Thus, interoperability became a higher-order good, and the pressure hardened the TAF into a standardised textual artefact. The form ossified. The cost of change was no longer technical but institutional, so change rarely happened.
Profit (1970s–90s)
When mass-market air travel boomed, forecasts were no longer just about the weather. They informed airline scheduling, crew rostering, slot allocation, fuel planning, network-level traffic flow. The system’s ultimate end was efficiency and margins, and the oft-erratic effects of atmospheric physics interacting with topography weren’t acceptable. A grand apparatus, with OpMets at the core, emerged to smother this volatility and replace it with a coherent, continuous story about the weather.
When volatility broke through, the systemic response was to ratchet: tighter governance, more checks, the art of post-hoc defensibility. A forecast that didn’t come to fruition always had a human backstopping it, able to narrate deviations and justify intent. Informal and shadow practices flourished around the forecaster and were tolerated precisely because accountability remained anchored in the human at the point of issuance. This distinction – between what may inform a forecast and what may be the forecast – became one of global aviation’s most enduring boundaries.
Climate, Security, Information (1990s–2020s)
Later eras layered on purpose without touching the kernel. Aviation’s environmental impact made forecasts inputs to fuel-burn, emissions and resilience decisions. 9/11 added security to the co-directors of scale, profit and climate, shrinking the appetite for risk further still.
The information age further transformed the forecaster’s decision environment, yet the TAF didn’t move an inch. It couldn’t. So everything around it did. Numerical weather prediction, practically infinite data sources, specialised tooling, instantaneous communications; it all bled in.
Shot Making
Today and across all prior eras, TAFs encode operationally relevant insight as a blend of deterministic, temporal and probabilistic statements. A deterministic baseline states prevailing conditions. Change groups then structure time: BECMG expresses confident, gradual evolution; TEMPO signals short-lived, tactically relevant deviations; PROB30/40 flags uncertainty. Here’s a more complex TAF than the one this essay began with:
TAF EGLL 252256Z 2600/2706 22007KT CAVOK
PROB40 TEMPO 2601/2609 BKN007
BECMG 2609/2612 22015G25KT BKN035
BECMG 2618/2621 22012KT
TEMPO 2621/2706 7000 RA BKN014
PROB30 TEMPO 2621/2624 22015G25KT
PROB40 TEMPO 2622/2706 4000 RADZ BKN007=
In effect, this is the weather story at a particular site compressed to around 200 characters. Let’s decode it.
The TAF covers 30 hours for London Heathrow. It runs from midnight to 0600 the next day. The baseline is CAVOK. Visibility is at least 10km. There is no significant cloud or weather. The wind is light, from 220°, at 7kt.
A PROB40 TEMPO (01–09Z) adds a hedge to this baseline: a 40% chance of broken cloud at 700ft. The first BECMG (09–12Z) keeps the wind at 220°. It raises the wind to 15 kt, gusting 25 kt. It ends CAVOK. Cloud becomes broken at 3,500ft. The second BECMG (18–21Z) drops the wind to a steady 12kt. The gust stops.
From 21Z, a TEMPO applies until the TAF ends at 06Z. It forecasts moderate rain, visibility down to 7,000m, and broken cloud at 1,400ft. Two PROB groups sit inside this TEMPO. The PROB30 TEMPO (21–24Z) gives a 30% chance the wind gusts back to 25kt. The PROB40 TEMPO (22Z–06Z) carries the main risk: a 40% chance of visibility down to 4,000m, rain mixed with drizzle, and cloud broken at 700ft.
The TAF does not forecast this outcome; it communicates that the risk is real on this night and that a hedged plan is the right response. The TAF does not end with a return to good weather. The rain-and-drizzle risk is still live when the TAF ends. The next TAF must describe what happens after that.
Because readability and salience matter more than completeness, TAFs are intentionally succinct. Accuracy that changes no decision is functionally irrelevant, and excessive signalling of uncertainty clouds important thresholds. Because of this, forecasters exercise a tacit, cultivated taste when deciding how much nuance to surface and how to express risk without overloading users.
Whilst TAFs as a collective are designed to provide a continuous, coherent weather story, the operational heart of an individual TAF is discontinuity. Magnitude on its own is not consequential; threshold crossings are. Two forecasts may differ by the same amount, yet only the one that crosses a meaningful threshold triggers operational change.
Variability within a threshold band is absorbed; oscillation across bands is disruptive, expensive, sometimes unsafe. Amendment rules encode this asymmetry: when a forecaster no longer expects the current TAF to hold, it must be re-issued to re-anchor the commitment. When the forecast fails outright and reality departs from the commitment, that failure is known as a bust. And busts happen.
Forecasting, as an art and a science, is closer to field goal or free throw percentages in basketball. Players take shots, yet misses are a certainty. But that does not (and should not) undermine the confidence that the next attempt is going in.
Promise Production
A TAF is a promise about the future, and every promise invites a reckoning. In aviation, the reckoning arrives every half hour, and it’s called a METAR: a standardised, instrumented observation of what the weather at an aerodrome actually is. When forecast and reality diverge, nobody acts on the forecast. Pilots, controllers and dispatchers act on the METAR, and the TAF is summoned for judgement.
That judgement – verification – inherits the threshold logic. The system doesn’t much care about pointwise numerical error. It cares whether the forecast preserved threshold status, hit the right category, landed the timing. Timing is the cruellest dimension, because timing governs when the entire system’s risk posture should shift. Fog that arrives an hour early doesn’t just embarrass a forecaster; it activates low-visibility procedures prematurely and drags resource allocation out of phase with the sky.
Unfortunately, busts cluster where the atmosphere moves faster than the update cycle. And because a false alarm costs inefficiency while a missed deterioration is potentially catastrophic, forecasters learn a quiet, defensive lean. They opt for slightly wider BECMG windows, deterioration timed a touch early, the more conservative category when uncertainty straddles a boundary. The verification regime doesn’t merely measure forecasters; it trains them.
The stakes of that training are regional, not local. A single threshold flip at one aerodrome reshapes flow within minutes: runway changes ripple through taxi-time assumptions, alternates siphon traffic to neighbouring airports, upstream rotations shuffle to protect crew-duty margins. The never-ending stream of haikus is also a never-ending stream of synchronisation pulses, and a mistimed pulse dislocates the choreography across entire sectors.
So who writes these promises, and who signs them? It helps to separate two questions usually collapsed into one: where forecast content is generated (human or machine) and where issuance authority resides (human or machine). Cross them and you get four structurally distinct modes.
The global baseline is human-generated, human-issued. Even heavily tool-assisted workflows live here, because decision aids don’t change authorship.
The “centaur” mode has machines drafting and humans editing, contextualising and committing.
Its mirror – human-crafted content, machine-carried authority – is conceptually possible but operationally rare.
And full automation relocates both the epistemic and the accountability functions into the technical system.
These are not rungs on a maturity ladder. They are discrete authority structures, each with its own tolerances. Nor does the TAF stand alone in any of these structures.
Substitutes compete with its compact syntax: ensembles express uncertainty the TAF cannot, nowcasts displace early-period reliance.
Complements support it: local tooling, decision matrices, and – above all – OpMet tacit knowledge, the embodied sense of how fog forms along one particular valley, which no model currently reproduces.
Shadows bypass it: weather models, internal drafts, chat-channel heads-ups, adjacent products and services, the half-approved system running quietly in the corner.
The TAF’s real operating envelope is defined as much by these adjacencies as by its own syntax. Any future approach to producing TAFs starts inside this envelope, not outside it.
Evolutionary Constraints
Complexifying further: the TAF ecosystem does not evolve on its own terms. Mega-trends bound what TAF production may look like over the next decade. Below are the pertinent ones.
Human scarcity. The operational meteorologist is a rare and slowly made creature: years of training, untransferable local knowledge, a temperament suited to holding asymmetric risk at 3am. The pipeline is thinning, and the tacit knowledge that anchors TAF quality walks out the door with each retirement, faster than it is replaced.
Technical debt. The infrastructure beneath TAF production is old and load-bearing: message-switching networks with telex ancestry, fixed-width formats designed for channels that no longer exist, national systems integrated through decades of patchwork. Every future inherits this substrate, and the substrate has sharp teeth.
Technical progress. Pulling the other way: numerical weather prediction keeps improving, machine-learning models have begun outperforming physics-based systems on headline metrics, and the cost of generating a plausible draft TAF is collapsing toward zero. Such progress makes forecast generation abundant while doing nothing, by itself, to resolve issuance authority. The bottleneck migrates from generation to assurance.
Regulation and compliance. The assurance ratchet turns one way: layers added, never removed. This is not a pathology; it is the point. But the gap between what is technically demonstrable and what is institutionally certifiable widens with each crank. Regulation is simultaneously the strongest brake on automation and – should a regulator ever bless a pathway – its most potent accelerant.
Operator economics. Airlines and airports run on razor margins, and meteorological provision – historically bundled, subsidised, invisible – is increasingly itemised. Operators want cheaper forecasting and the volatility-smothering, defensibility-providing service that only well-resourced human forecasting has historically supplied.
Consumer cost of living. Passengers discipline airlines, airlines discipline suppliers, suppliers discipline cost bases, and meteorology feels the compression from three layers up. Delay and cancellation have become politically visible consumer rights issues, raising the price of forecast failure. The consumer demands both cheaper weather and better weather, without ever knowing a TAF exists.
Cultural fragmentation. The mid-century settlement that hardened the TAF – open data exchange, shared standards, technical communities operating beneath geopolitics – is fraying. Data nationalism, commercial enclosure of model outputs, and diverging professional cultures all strain the interoperability on which a single global forecast grammar depends. The “globally standardised” TAF starts to look less like a settled fact and more like an achievement requiring proactive maintenance.
These pressures do not point in a single, obvious direction. Scarcity and cost push toward automation; debt, regulation and fragmentation resist it; technological progress says so much more is possible whilst raising hard questions about labour and capital.
The forecaster’s craft, at its heart, has always been deciding how much confidence the evidence will bear, committing to it publicly, and amending without ego when the weather inevitably disagrees. That is also, precisely, the craft this moment demands of those with a stake in TAF production and usage.
The question was never whether machines can write the haikus; of course they can. It is whether we can build a system worthy of signing them, and what the shape of that system should be.
Five Futures
Whatever the shape of the future system is, it will not be something entirely novel. It’s possible to construct five end states that answer the critical questions: who writes the promise, and who signs it? Like the generation-issuance quadrant above, these aren’t maturity stages; they’re distinct alternative futures.
Frozen Centre, Accelerated Periphery
Nothing changes at the core – and that is precisely the story. Human-issued TAFs remain the institutional centre while high-frequency forecasting, autonomous actors and data-rich platforms accelerate around them.
The pace-layer separation widens yearly: a slow, assured core; a fast, experimental periphery; a growing translation burden between them. Eventually they stop describing the same world.
Analogue: global payments, with its stable institutional core and fintech innovating furiously at the edges of a behemoth it cannot touch.
Centaur Under Extreme Load
The upstream explodes: automated systems generate more candidate forecast content than any human could ever produce, and continuously refresh it. Yet the commitment boundary does not move – machines may inform without limit, but a human must still be the forecast.
The forecaster’s craft migrates from writing the haiku to selecting it: triage, curation, sign-off under ever-heavier load. The bottleneck is no longer what can be known but what can be committed to.
Analogue: radiology ML/AI, where machines produce the analyses, clinical sign-off stays human, and the gate becomes the pressure point.
Fallback During Exceptions
This future splits the weather itself. Routine regimes are automated end to end – machines issuing TAFs for perhaps 90% of operational hours – while rising uncertainty snaps authority back to humans via explicit triggers.
The uncomfortable implication: humans end up handling only the hard cases, precisely where busts have always clustered, while the routine hours that once built a forecaster’s feel for a site belong to the machine. The architecture is clean; whether expertise can survive on a diet of exceptions is anything but.
Analogue: automatic train operation and autonomous-driving stacks, with their well-documented safety upside contrasted with worries about the human as the fallback.
Machine-to-Machine
Both the writing and the signing go technical. Nobody stands behind each individual TAF; the architecture stands behind all of them. Trust anchors migrate from human interpretability and post-hoc defensibility toward reproducible pipelines, audit-friendly data structures, and algorithmic verification.
The commitment boundary – the most enduring line in the institution – becomes a machine boundary. This future relocates the most and accordingly carries the heaviest assurance burden: it must reconstruct, in technical form, every property that a century of human-anchored governance was built to provide.
Analogue: cryptographic consensus systems, with commitments made by protocol and humans governing the rules (not the outputs).
Non-Centrality
The quietest future, and the strangest: the TAF survives untouched and simply stops mattering. Still issued, still verified, still formally load-bearing – while operational influence and decision-making drain away to ensemble probabilities, nowcasts and machine-readable feeds consumed directly by planning systems.
No decision ever kills it. It is not defeated but routed around: each migration to a richer product is locally sensible, and only the aggregate constitutes an ending.
Analogue: fixed-line telecoms after mobile being maintained, even mandated, but no longer driving majority behaviour.
Reality will likely braid several of these. But the braiding is not yet fixed. Capability is developing quickly all around, pressure is binding disparate sociotechnical systems together, and governance conversations that have been closed for decades are finally opening up. Landscapes like this are not merely receptive to opinionated movement; they’re made by them.
Thank you for sharing this!
I would love to add another boundary: the conditions under which another institution may rely on a forecast for a particular decision.
Aviation benefits from rapid feedback: forecasts can be checked against frequent observations as the weather unfolds.
In my work on assessing simulations of living landscapes for climate adaptation I am facing a verification problem. A model might estimate how much a wetland reduces flooding, but an insurer, investor or public authority still needs to establish which commitments that evidence can support, who maintains the conditions it depends on, and what happens when those conditions change.
Using the same model to justify a commitment and check its success risks circular reasoning unless independent evidence can challenge its assumptions.
The thing to formalise is a commitment to make the evidence, assumptions and limits explicit, keep the assessment open to challenge, and respond when its basis changes. Since in the case of nature based solutions, a financial commitment may run for decades, the obligation to check whether its supporting assumptions still hold must as well. There is a lot to study and learn from the processes and systems developed in the aviation domain, open banking and how we approve new pharmaceuticals.
Thank you again for sharing your thoughts!
Best wishes, Jan