How the jellyfish forecast works: the model in plain words

Every day, for every beach, we estimate how likely you are to meet jellyfish today and in the next few days. This is how that estimate works, without formulas.

What a risk level means

There are four levels: low, moderate, high and very high. Each is an estimate of the chance of meeting jellyfish, not a promise. "No reports" does not mean there are no jellyfish: follow beach flags and lifeguards.

  • Low: a usual day for the season, no recent reports.
  • Moderate: the season or conditions raise the chance, but nothing unusual.
  • High: a clearly more jellyfish-prone day than usual: steady onshore wind, recent reports or the peak of the season.
  • Very high: rare days, usually with recent confirmed reports at the beach itself.

Three jellyfish groups, not one

Jellyfish off Thailand behave differently, so the model scores three groups separately:

  • Bluebottles (Portuguese man o’ war) float at the surface and their sail catches the wind. They reach beaches after several days of onshore wind, most often on exposed west-facing Andaman beaches in the monsoon season.
  • Box jellyfish stay in calm, warm water near the shore, more often after rain. Dangerous but rare.
  • Other jellyfish and stinging plankton: large jellyfish, mauve stingers, stinging larvae. Each area has its own season for them, for example the dry season around Phuket.

The beach level is the highest of the three. The "Why this level" block on a beach page shows which group sets it and what raises it.

The season of the area

For every area and month the model has a background: how often each group has been met there at that time over many years. It is based on official warnings1 and beachgoer reports. That is how it knows, for example, that Phuket has two seasons: bluebottles in summer and other jellyfish in the dry season.

The background is only a starting point. What matters more is how a given day differs from a typical day of that month at that beach.

Weather: what raises the risk

Weather comes from a wind, wave and rain forecast refreshed several times a day2. The orientation of each beach is taken into account: what counts is wind from the sea.

  • Onshore wind that holds for several days in a row: the main signal for bluebottles. The steadier the wind toward the shore over the last 7–10 days, the higher the risk.
  • Onshore wind and waves over the last two days.
  • Rain over the last 3 and 14 days, for box jellyfish.
  • Calm sea: box jellyfish are more common in calm water, bluebottles less.

The model compares these conditions not with fixed numbers but with what is usual for that beach in that month. The typical day of each beach is computed from the same weather data as the forecast, for 2022–2026. The model was trained on many years of weather history3 and thousands of dated jellyfish observations.

Recent reports and drift

Official warnings and beachgoer reports raise the risk for the jellyfish group reported: strongly at the beach itself, less at nearby beaches. Their weight halves every few days. "No jellyfish" reports lower the estimate slightly.

Fresh sightings are also carried forward with forecast currents and wind for 3 days. If a track reaches another beach, that beach is warned in advance. For bluebottles the model allows for their sail deflecting them from the wind direction4.

From probability to level

Each group has its own thresholds, so that "high" means a genuinely unusual day rather than every day of the season. Across all beaches, high or very high comes up on about 8–10% of days; more often on exposed beaches in peak months.

A level is a comparison with a typical day, not an absolute guarantee. A low level in season does not mean there are no jellyfish.

How accurate the forecast is

The model was tested on years it did not see in training. For bluebottles it is clearly better than the previous formula; for box jellyfish and other jellyfish only slightly, because the exact day they appear is hard to tell from the weather. For them the season and recent reports matter most.

Every day we store the 5-day forecast for each beach and later compare it with what beachgoers and DMCR reported: "jellyfish seen" or "no jellyfish".

Forecasts stored since 2026-10-24 (1 days).

Same-day forecastForecast 1 day(s) aheadForecast 2 day(s) ahead
Forecast–report pairs0 (0 / 0)0 (0 / 0)0 (0 / 0)
Mean risk when jellyfish were seen———
Mean risk when there were none———
Brier score (lower is better)———
ROC-AUC (0.5 random, 1 perfect)shown once there are at least 5 reports of each kindshown once there are at least 5 reports of each kindshown once there are at least 5 reports of each kind

Accuracy of the trained model (v1)

Tested on years the model did not see in training. "Day vs typical day" is how well the model separates days with jellyfish from ordinary days of the same month; "present vs absent" is how well it separates "jellyfish seen" from "no jellyfish" reports.

GroupDay vs typical dayPresent vs absentObservations
Bluebottles0.79 (previous formula: 0.71)0.87 (previous formula: 0.77)318
Box jellyfish0.59 (previous formula: 0.54)0.50 (previous formula: 0.52)132
Other jellyfish and plankton0.53 (previous formula: 0.51)0.64 (previous formula: 0.50)1134
All together0.55 (previous formula: 0.53)0.60 (previous formula: 0.54)2644

"No jellyfish" reports matter as much as sightings: without them accuracy cannot be measured. Report jellyfish →

Fig. 1. Model accuracy on held-out years and the ongoing check of forecasts against reports.

What the forecast cannot do

  • Predict a particular jellyfish at a particular spot on the beach.
  • Resolve small bays and headlands that weather models cannot tell apart.
  • Know about jellyfish nobody has reported.
  • Replace beach flags, lifeguards and common sense.

If you see jellyfish, report them on the site: every report sharpens the forecast for nearby beaches.

Sources

  1. DMCR (Department of Marine and Coastal Resources, Thailand) — marine hazard warnings
  2. Open-Meteo — weather and marine forecast API
  3. Hersbach H. et al. (2020). The ERA5 global reanalysis. Q. J. R. Meteorol. Soc. 146, 1999–2049
  4. Lee E. A., Schaeffer A., Groeskamp S. (2021). Drifting dynamics of the bluebottle (Physalia physalis). Ocean Science 17, 1341–1351

Questions and answers

Why is the risk low when I saw jellyfish?

The model does not know about jellyfish nobody has reported. Send a report: it raises the risk for that beach and nearby beaches right away.

How far ahead can I trust the forecast?

Best for today and tomorrow. Further ahead the wind forecast error grows, and jellyfish drift is only computed for 3 days.