Which is the best weather forecast?

There's a catch in this question that we rarely notice.

Weather forecasting, whether for wind, rain or waves, is the result of processing numerical models. These models analyze environmental data, run complex equations and produce the numbers that show up on your favorite website or app.

And here's the trick: most forecasting sites and apps show the same information, just in different formats. Some are more visually appealing, others more complete. That's why the best forecasting site or app for you is the one that delivers what you need in the easiest and most complete way.

But that only solves half the problem.

Most of the time, what we really want to know is which forecast is more accurate. Which one gets it right more often. Which one is best suited for a specific decision, such as sending a vessel out to sea, authorizing a lift or releasing a maintenance window. To answer that, you need to look beyond the layout and understand the models behind the screen.

In 2026, this question became even more interesting. Artificial intelligence has fully entered weather forecasting, and that changed part of the answer. Let's break it down.

What is the best weather forecasting model?

Anyone who follows forecasting sites and apps knows there are different forecasts for the same place and the same day. That happens because each one comes from a different numerical model, developed by large regional or national government agencies, such as the American NOAA.

You've probably seen the acronyms of the main global models: ECMWF, ICON, GFS, WaveWatch III.

These are the results that most weather apps use, dressed up with a simpler, more pleasant layout.

Worth repeating: choosing between free sites, and some paid ones, comes down to how you prefer to visualize the information, since they all draw from roughly the same source. To go further, you need to assess which model performs best for your region and your specific need.

There are a few ways to evaluate this:

1. Ideally, you'd have an accuracy analysis of the model for your location and your use case. That's what we did, for example, in the wave forecast accuracy analysis at the Dan Tysk wind farm, for Vattenfall.

This type of analysis provides the confidence needed for decisions that affect the safety of people, the environment and assets. No generic forecasting service delivers this out of the box.

2. Another option is to check the forecast's overall accuracy, information available on some sites, though not all. It helps, but it's limited: a forecast can be excellent for light winds, up to 15 knots, and fail precisely above 25 knots. If your concern is strong wind, overall accuracy doesn't tell the whole story.

3. In the absence of a customized analysis, looking at the technical characteristics of numerical models already helps a lot. They show what the model is capable of capturing and give you more confidence when deciding between conflicting forecasts.

The AI revolution in weather forecasting: what changes for you

i4seaWhich is the best weather forecast?The question has a hidden catch. Learn how spatial resolution, temporal resolution, and update frequency determine which weather forecast model is best for your needs.i4sea | 25 mar | Adicionado por Aurélio de Exú[14h55]

Since this post was first published, global weather forecasting has entered a new phase. In February 2025, ECMWF put AIFS into operation, its first artificial intelligence-based model (ECMWF, Feb/2025). Months earlier, in December 2024, Google DeepMind published GenCast in the journal Nature, a global probabilistic model with a 15-day horizon and a resolution of 0.25 degree, roughly 25 km at the equator (Nature, doi: 10.1038/s41586-024-08252-9).

The result caught the attention of the entire industry: GenCast outperformed ECMWF's physical ensemble system in 97.2% of the 1,320 metrics evaluated, and proved particularly better at signaling extreme events (Nature, Dec/2024; Valor Econômico, Dec 5, 2024).

This is good news. It's also less magical than it looks.

What AI actually gains: brutal processing speed, ensembles (multiple probabilistic scenarios) that are much cheaper to generate, and a sharper signal for extreme events. Where an agency once needed hours of supercomputer time to run a set of scenarios, today it takes minutes.

What AI doesn't solve: local scale. The grid used by GenCast and other global AI models still sits around 25 km, coarser than ECMWF's physical model at 9 km, and far coarser than what a port, a substation or a mining slope needs to make safe decisions.

In other words: AI made global models statistically smarter. It didn't zoom in on them. It still sees Brazil, not berth 3 of your terminal.

How do you know which is the best forecast?

To assess which sea and weather forecast is best for your location and your activity, three characteristics of numerical models matter more than any polished app badge:

• Spatial resolution
• Temporal resolution
• Update frequency

These are what indicate the circumstances under which you can trust a forecast more, or less.

What is spatial resolution in a forecasting model?

The first step to understanding a forecast is knowing the spatial resolution of the model used. Spatial resolution works like the number of pixels in a photo: the more pixels, the smaller the distance between them, the greater the level of detail in the image.

It ranges from a few meters, in hyperlocal models, to tens of kilometers, in global models. It indicates the model's ability to capture spatial variations in weather and sea conditions.

The best resolution for you depends on your objective.

To know whether a storm is approaching the Brazilian coast, a low-resolution model already works well. But if you need to know what will happen in a specific stretch of coastline, such as a port, the higher the resolution, the better.

What is temporal resolution in a forecasting model?

Spatial resolution isn't the only thing that matters. Temporal resolution also increases or reduces a model's ability to predict a specific weather event.

The most common temporal resolutions are hourly (6am, 7am, 8am, 9am) or every three hours (6am, 9am, noon, 6pm). The impact of this is simple to understand.

If you only have data every three hours, any phenomenon that occurs within that interval goes unnoticed. A strong wind lasting a few minutes, or even an hour, may never show up in a three-hourly forecast.

The same applies to hourly forecasts: the data for 12:00pm may not hold true for 12:45pm. And since free forecasts don't provide results every 15 or 30 minutes, the next data point only arrives at 1:00pm, missing the event entirely.

This detail defines what's reasonable to expect from a forecast. You can't expect an hourly model to predict an atmospheric phenomenon that lasts 10 minutes.

How often are weather forecasts updated?

[14h55]

The third factor is how often the forecast is updated throughout the day.

Every forecast tries to diagnose the future based on recent past data. The fewer times a day the model updates, the more outdated the reading becomes relative to what's actually happening, because the atmosphere changes fast.

The most common frequency ranges from once to four times a day, meaning every six hours. This means global conditions have been reprocessed and a new forecast is available for whoever uses the model.

This frequency determines how often it's worth checking the forecast and how well it reflects recent reality. Generally speaking, more daily updates mean better-quality forecasts.

There are other factors that also affect forecast quality, but with these three you can already reliably assess which is the best option for your location.

What is the best forecasting model for my region?

Of the three characteristics, spatial resolution is by far the one that carries the most weight. It's no use updating a model ten times a day if it still has low resolution: that doesn't change its ability to see the specific stretch of your coastline.

GFS, the most widely used free atmospheric model, and its ocean counterpart, WaveWatch III, operate with a spatial resolution of approximately 22 km, equivalent to 48.4 soccer fields.

Because of this coarser resolution, GFS tends to fail in mountainous regions and in rain and cloud patterns. WaveWatch III performs well in open ocean but fails in shallow zones, such as ports, bays and coves.

ICON is more modern, with a resolution of 13 km, and delivers good results globally, though it performs better in Europe, the continent for which it was optimized.

Among traditional physical models, ECMWF remains the one with the highest spatial resolution, 9 km, equivalent to about 8 soccer fields per grid point. It delivers both atmospheric variables (wind, rain, temperature, humidity) and ocean variables (wave height and period), with hourly results updated every six hours.

And what about AI models, like GenCast? They sit around 25 km of spatial resolution (Nature, Dec/2024). In other words: they gain in statistical intelligence and speed, but lose in zoom, even compared to the physical ECMWF, already considered the most effective among traditional global models.

In practice, among free global models, ECMWF remains the safest bet, especially outside Europe. But all of them, including the most advanced AI models, share the same limitation: none of them see what happens inside a port, a slope or a substation.

How to make better decisions based on weather forecasts?

After this rundown of models and their characteristics, an uncomfortable truth needs to be said: every forecast will eventually be wrong.

You can't expect total accuracy. It's still technologically unfeasible. The atmosphere is too complex, and to deliver a 10 to 15-day forecast within a reasonable timeframe, every model needs to make approximations. That applies to physical models and AI models alike.

Even so, adequate sea and weather forecasts remain the best tool available for planning ahead, managing risk, reducing costs and seizing opportunities.

What changes the game is the cost of deciding based on a forecast that's wrong, or too generic for your asset. And that cost, in Brazil, has been mapped out.

Between 2022 and 2024, the country recorded R$ 184 billion in climate-related losses, an average of R$ 60 billion per year. Of that total, 91% had no insurance coverage whatsoever (CNseg/EY, COP30 2025).

[14h55]

In ports, the problem shows up in the form of demurrage: US$ 2.3 billion, about R$ 13 billion, paid in 2024, a 15% increase over 2023 (Bain & Company, via Valor Econômico, Apr/2025). In Santos, 84% of vessels were delayed in 2024, with an average wait of 12 days (Datamar/CNT).

And the trend isn't easing. Extreme weather events became three times more frequent and severe over the past decade (Atlas Digital MDR). Where most people see the unexpected, this figure points to a management variable that should already be on every operation's risk spreadsheet.

Given this, a few points help you get more out of any forecast, physical or AI-based:

1. Global models are good at predicting extreme conditions. Even when they get the magnitude wrong, the signal that something significant is coming usually shows up.

2. They're also good at capturing the most frequent local patterns, the typical climate of that region.

3. But they don't identify local differences in wind and waves, such as between the open-sea navigation channel and the inside of the port, or between one berth and another within the same terminal.

A resolution of 9 km means the data point you're looking at could be up to 9 km away from your actual point of interest. In a coastal zone, that changes everything. A resolution of 25 km, like today's most advanced AI models, widens that gap even further.

Breakwaters, jetties and other port structures reduce wave intensity inside the port. No global model, physical or AI-based, captures this difference between sheltered and exposed areas. The same goes for bays, coves and other coastline particularities.

Spatial resolution, even at ECMWF's 9 km or an AI model's 25 km, doesn't "see" these nuances of coastline, depth or infrastructure. For that, only a high-resolution model built for your specific location works: the hyperlocal model.

Tips for making better decisions with sea and weather forecasts

With expectations aligned on what each type of forecast actually delivers, here are some practices to get more value out of the information you already have:

1. Favor sites and apps that show the time of the model's last update, the spatial resolution used, and hourly data, not just three-hourly.

2. Check the forecast right after the most recent update, and keep tracking changes with each new run.

3. Leave the final decision as close as possible to the event. The closer to the actual time, the more accurate the data tends to be.

4. When forecasts diverge from one another, accept that this is normal, and build more than one planning scenario.

5. Whenever possible, compare with local sensors to find out whether the forecast tends to underestimate or overestimate the reality of your asset.

Weather forecasting is not climate intelligence

All the progress AI has brought to global models is real and welcome. It just solves a different problem than the one faced by whoever operates a port, a power plant, a railway or a mine.

Weather forecasting, even AI-powered, still speaks the language of a country or a continent. Operational decisions speak the language of the berth, the substation, the slope, the 200-meter stretch of railway that floods first.

[14h55]

That's the gap climate intelligence tries to close. At i4sea, we work with a resolution of 1 to 3 km, compared to the roughly 25 km delivered by today's most advanced public models, including AI-based ones. This is backed by more than 10 years of proprietary climate reanalysis and 18 hydrometeorological hazards continuously monitored, now present in more than 100 critical assets across Latin America and Europe.

The question "what's the best weather forecast" got a new answer in 2025 and 2026: global models became faster and statistically smarter with AI. But for whoever makes decisions about a yard, a berth, a transmission line or a mining face, the right question has changed. It's no longer just "which model gets Brazil right more often." It's "which data actually sees my asset."

That's what drove us to develop i4sea's Climate Agent, a conversational climate risk copilot built on this same hyperlocal data foundation. It doesn't replace weather forecasting. It translates forecasts, historical reanalysis and risk into a direct answer for whoever needs to decide today, without waiting for a report.

Weather forecasting is not climate intelligence. And confusing the two comes at a steep price, as shown by the demurrage figures and uninsured losses we've seen here. Decide with the climate, not by the climate.

Sources

• ECMWF. Operational launch of AIFS (Artificial Intelligence Forecasting System), Feb/2025.
• Nature. "GenCast: probabilistic weather forecasting with machine learning", Dec/2024. DOI: 10.1038/s41586-024-08252-9.
• Valor Econômico. Coverage on GenCast and its performance versus ECMWF, Dec 5, 2024.
• CNseg/EY. Study on climate-related losses in Brazil, presented at COP30, 2025.
• Bain & Company, via Valor Econômico. Survey on demurrage at Brazilian ports, Apr/2025.
• Datamar/CNT. Vessel delay data for the Port of Santos, 2024.
• Atlas Digital de Desastres no Brasil, Ministry of Regional Development (MDR).
• i4sea. Wave forecast accuracy analysis for Vattenfall, Dan Tysk wind farm.
• i4sea. Proprietary climate reanalysis database, 1 to 3 km resolution, 18 hydrometeorological hazards monitored.