Four forecasts on screen shouldn't create more doubt, they should create more confidence. Learn how i4cast turns agreement and lead time into a practical recommendation.
"if the four forecasts diverge, won't that make me feel less sure when it's time to decide?"
It's a fair question. It helped shape the tool. Divergence between forecasts isn't noise. It's information. Each forecast is generated independently, by different institutions and methods. When independent sources reach the same conclusion, the chance of the event happening increases. When they disagree, in many cases the signal is different: it's still too early to call it.
i4cast's role is to do that reading for you. For each event of interest, the platform answers two questions and combines the answers into a recommendation:
Agreement: how many forecasts point to the event?
Lead time: how long until it happens?
Next to each forecast's name there's a number: i4cast 3 km, ECMWF 9 km, ICON 13 km, GFS 25 km. That number indicates the resolution — the size of each grid cell used to compute the forecast. The smaller the cell, the greater the local detail.
Think of a photo. With few pixels, you recognize the landscape. With many pixels, you can read the ship's name plate. On a 25 km grid, an entire port terminal — with its channel, breakwater, and neighboring hill — can fit inside a single cell. On a 3 km grid, the same area is covered by dozens of cells. As a result, local phenomena show up more sharply: the breeze that shifts in the late afternoon, the fog that closes the harbor entrance, and the wind channeled by the terrain.
That's why i4cast, with 3 km and AI-based local refinement, is the reference in the comparison. It's also why the tool warns, about GFS: "Lower resolution: may fail in mountainous areas and in cloud and rain forecasting."
Global forecasts aren't necessarily wrong. They see from farther away. That difference in perspective makes the comparison more valuable, because it shows when independent sources are pointing to the same scenario.
With an event active on the combined forecasts screen, the platform evaluates its conditions — such as gusts above 25 knots and visibility below 1 mile — in each forecast and at each hour. A forecast "captures" the event when it indicates all the required conditions at the same time in at least one instant of the period.
The result appears as a count, for example "3 out of 4", accompanied by one of these readings:
"All forecasts point to the event in this period." It's the strongest signal the tool can give. Independent sources, with different methods and resolutions, reach the same conclusion.
"i4cast plus N other forecasts point to the event." There is partial agreement. The event is likely, but divergence still exists. This is the most common scenario at medium lead times.
"Only i4cast points to the event in this period." This calls for attention, but it does not mean a false alarm. i4cast can capture local conditions that global forecasts don't see. Harbor-entrance fog or channeled wind may simply not show up on a 25 km grid.
Two details prevent confusion:
The count is per occurrence. If the event appears twice within the horizon — say, tonight and three days from now — each window has its own count.
When one of the event's variables exists only in i4cast, such as waves or currents, the tool indicates "only i4cast forecasts this variable" instead of displaying a misleading fraction. In those cases, agreement is calculated only over the variables in common.
Here an important rule of meteorology comes in: forecast error grows with lead time. A forecast for tomorrow is structurally more reliable than a forecast for a week from now. As the event gets closer, the forecasts also tend to agree more with each other.
The time pill in the tooltip — such as "in 36h", "in 4 days", or "ongoing" — places the event in one of four windows:
Less than 48h: mature forecasts. Agreement carries a lot of weight, and divergence also deserves more attention.
2 to 5 days: planning window. Treat the scenarios as likely and follow the daily updates.
5 days or more: window labeled "Distant forecast", with the recommendation box always in amber. As the platform itself puts it: "It's normal for forecasts to disagree several days out, and they can still change." No reading in this window should become a final decision.
Ongoing: the event is happening now. The recommendation shifts focus: monitor real-time conditions.
Each tooltip ends with a badge summarizing the two readings, such as "Very high confidence · All forecasts agree". The logic is this:
| Less than 48h | 2 to 5 days | 5+ days ("Distant forecast") | |
|---|---|---|---|
| All forecasts agree | Very high confidence | Very high confidence | High confidence |
| Partial agreement | High confidence | High confidence | Moderate confidence |
| Only i4cast | Moderate confidence | Moderate confidence | Under watch |
The pattern is simple: agreement sets the starting point, and distance in time can lower the confidence level. There is never "very high confidence" for an event a week away. Even when all forecasts agree, the horizon still imposes uncertainty.
The color of the recommendation box follows the type of event. Restrictive events, which close an operating window, appear in red. Permissive events, which open a favorable window — like an expected calm — appear in green. In the distant window, all of them turn amber: wait for confirmation.
Scenario 1: all agree, event in 30h. Your gusts-above-25-knots event appears with "4 out of 4" and a "Very high confidence" badge. The recommendation is: "Count on the event happening." It's time to act: reschedule the maneuver, bring the operation forward, or alert the teams. Waiting for more confirmation here tends to add little.
Scenario 2: 3 out of 4, event in 3 days. "i4cast is backed by 2 other forecasts. Treat the event as likely and follow the daily updates." It's not yet time to halt the operation, but it is time to sketch out plan B. If agreement rises to 4 out of 4 as the event approaches, you decide with more lead time and less improvisation.
Scenario 3: only i4cast, event in 18h. This is the case that raises the most doubt, and it's also the scenario where local resolution matters most. "Only i4cast points to the event, but it captures local conditions the other forecasts don't see. Treat the event as likely, have a plan B, and follow the next updates." Ignoring the warning because the other three forecasts show nothing may be exactly the mistake the 3 km resolution helps you avoid.
Scenario 4: all agree, event in 6 days. Even with full agreement, the box stays amber: "The forecasts agree, but the event is still 6 days away and they can change until then. Follow the updates." Use this signal for the week's planning, not for a final decision.
In conversations with users, it became clear that the comparison plays different roles depending on the operation:
Short-term operational routine: if i4cast has already proven its accuracy in your microregion, your day-to-day can stay in it, using the meteogram, table, and thresholds, without needing to open the comparison. The feature is optional, not a mandatory step.
Extreme events and long horizons: for decisions 3 to 10 days out — such as shutdown scheduling, special operation windows, and severe events — agreement between forecasts is the best available estimate of probability and of variation in the timing of occurrence.
The comparison is one click away when you need it, and out of the way when you don't.
What counts as "capturing" the event? The forecast indicating all the event's conditions satisfied at the same time in at least one instant of the period, respecting the configured "AND" and "OR" operators.
Why does the count change between two periods of the same event? Because it's calculated per occurrence. Each forecast window has its own agreement count.
What about missing data? Missing data never counts as a forecast event. If a forecast doesn't cover a variable, the screen warns: "This forecast model does not provide this data". If the point doesn't have enough data to evaluate the event, the warning is also explicit.
Why is the recommendation box sometimes green? Permissive events, which indicate a favorable window rather than a risk, get a green recommendation when they're near and confirmed.
Agreement answers "how likely": the more independent forecasts agree, the higher the confidence.
Lead time answers "how mature": the closer it is, the more weight the reading carries.
Only i4cast pointing to the event is not a false alarm. Often, it's the 3 km resolution seeing what the global grid doesn't.
"Distant forecast" is radar, not a decision. Follow the updates and wait for confidence to rise — or for the event to disappear — before acting.
Divergence between forecasts has always existed. Before, it was scattered across several browser tabs. Now, it appears on a single screen: measured, contextualized, and translated into a recommendation.