A flood forecast can be produced long before it deserves to be shown. Producing it takes a model and data. Showing it takes evidence that it would have been right before, often enough to be worth acting on. This is how that evidence is built, and why a forecast that fails the test stays off the page.
Testing a forecast against the past
The test is called a hindcast. You run today's forecasting system on past dates, as if you were standing on each of those days, and compare what it would have said with what actually happened. Nothing about the past is hidden from the record, so every forecast can be scored.
A hindcast has one strict rule: the system may only use information that existed on the day it is pretending to forecast. Let it see a single day of the future and it will look far better than it really is.
The record also has to be long enough to include the events that matter. A season with no serious floods says almost nothing about whether a system can warn of one, however good the score looks.
Two ways to be wrong
A warning system can fail in two different ways. It can stay silent when a flood comes, which is a miss. Or it can raise the alarm when no flood comes, which is a false alarm.
Forecasters measure the first with the probability of detection, or POD: of all the floods that happened, the share the system warned about. A POD of 1 means it caught every one.
They measure the second with the false alarm ratio, or FAR: of all the warnings it issued, the share that turned out to be wrong. A FAR of 0 means every warning was followed by a flood.
Why both numbers matter
Either number on its own is easy to game. A system that warns every day for every province would catch every flood and score a perfect POD, while being useless. A system that never warns would never raise a false alarm.
So a forecast passes only when it catches enough real events and keeps its false alarms low enough that people still listen. Where to set those two bars is a decision about people, not only about statistics.
Who pays for each kind of error
A miss and a false alarm cost different people different things. A farmer who is not warned may lose a crop that is weeks from harvest. A farmer who is warned for nothing may harvest early, move equipment or pay for sandbags, and next time may not move at all.
That second cost is easy to underrate. A system that cries wolf trains its own audience to ignore it, so the false alarms of this season weaken the warnings of the next.
Why we hold a forecast back
We test our own flood forecasts this way. The latest hindcast did not reach the bar we set for catching flood events, so we do not show forecast levels yet. When the page opens, it will show measured values only and say so plainly.
Holding a forecast back can feel like hiding work that is nearly ready. But a forecast on screen is read as a statement about the coming days, and one that has not passed its test makes a promise nobody has checked.
What to use in the meantime
For warnings and evacuation orders in Thailand, follow the Department of Disaster Prevention and Mitigation (DDPM), the Thai Meteorological Department and your local authorities. The DDPM hotline is 1784; for medical emergencies, call 1669.
When our forecast passes its test, we will say what it was tested against and how it scored, so anyone can judge it for themselves.