MAE = mean |prediction − reference|
Mean absolute error scores the size of a typical absolute miss on the chosen data.
Verification in P4Choose how to read
The full view retains the technical content, figures and sources. The plain-language explanation introduces the ideas without changing the findings.
Why the best rainfall forecast depends on the score
How to use this companion
Use the controls to inspect the relationship described by the paper. The interactive display is explanatory; it does not rerun the archived training experiment or supply independent evidence of forecast skill. Read the linked full manuscript for the executed comparison.
What the study asks
Can pretrained Timestral models forecast Jamaican daily rainfall better than local models when they receive the same rainfall history?
How the comparison works
Two pretrained model sizes are evaluated against local controls at an island average and one grid cell. The study checks different forecast horizons, heavy-rain days and delays in the latest observations.
What the result means
Both Timestral models have lower next-day mean absolute error than the local controls. Persistence has lower next-day RMSE and identifies more heavy days. Ridge has lower RMSE over longer horizons. A model can do well on typical errors while missing the larger errors that another score emphasises.
Where the evidence stops
Only a small number of weekly test cases are extreme. Finalised rainfall records and imposed delays do not reproduce an operational service. The archived checkpoint revision is not pinned, and pretraining overlap is unknown.
Use the full view for the complete technical discussion, figures, interactive elements and sources.
P4 · INTERACTIVE 3D COMPANION
Compare rainfall errors and heavy events
Watch a constructed rainfall sequence through two different forecast styles. Inspect ordinary days and heavy peaks separately.
The question
Real geographic contextCalculated synthetic fields
ReferenceEstimate
Loading the 3D scene…
The WebGL 3D view is unavailable
A 2D fallback may appear in its place. You can also follow the study with these steps and the full paper.
- The question Can a forecast stay close on ordinary days but miss the largest rainfall events? Watch the smaller rain areas and occasional heavy peaks in the constructed reference.
- The method Compare a smoother, typical-day estimate with a simple rainfall-history baseline on the same sequence. Use the typical-day weight and wet-spell onset controls to inspect the trade-off.
- Check the result Switch to the difference layer and examine the error where rain is heaviest. Mean absolute error and root-mean-square error weight the misses differently. Input delay can change the comparison again.
Weight in a transparent synthetic stand-in; not a Timestral configuration option.
Assumed chance of a new wet spell before the generator’s seasonal adjustment; not a measured frequency of heavy days.
Added information delay in the constructed comparison.
Drag to rotate. Focus the scene to zoom with the wheel; arrow keys rotate and + / − zoom. Click a cell to inspect its transfers.
The question
Can a forecast stay close on ordinary days but miss the largest rainfall events? Watch the smaller rain areas and occasional heavy peaks in the constructed reference.
Synthetic forecast stand-ins illustrate the comparison; no Timestral checkpoint or paper prediction array is run here.
Data, views and methods
Calculated values
What this browser model represents
The browser uses transparent synthetic forecast rules to show the MAE/RMSE and delay questions. These are not Timestral-5M or Timestral-28M outputs.
The scene fits conditional means and medians on earlier synthetic days. Its 20 mm heavy-day threshold is a teaching choice, not the manuscript’s fitting-period percentile. It does not reproduce checkpoint pretraining, inference, probabilistic intervals or the reported rankings.
The complete paper reports the two executed checkpoints against four matched local controls. Read its tables for measured outcomes.
Method in the paper
RMSE = √mean (prediction − reference)²
Squaring puts more weight on large misses. A different model can lead on this score.
Verification in P4Geography and figure design
Terrain is a sampled Mapzen/Tilezen elevation mesh, with a 62.5 m display spacing over an 8 km square. Source tiles have approximately 36 m pixel spacing at this latitude; display spacing does not establish survey accuracy. Colours show the calculated field, while terrain heights provide location context. WebGL displays a 129 × 129 terrain grid; the CPU compatibility view resamples the same elevations to 33 × 33 for responsiveness. Neither display resolution is an accuracy claim. The location selector does not change the numerical model. P6 soil columns and P7 touching heat cells use diagram dimensions. Changing the cube count changes the synthetic lattice and replays the elapsed scenario; it does not refine an atmospheric forecast. P8 markers use requested pilot coordinates, with illustrative model values. P1 column heights repeat rainfall, or absolute error in Difference view, at diagram scale; they are not millimetres on the terrain. P8 column heights repeat the probability or absolute error; their signs come from the colour scale and inspector. P6 shows zero in the reference column when displaying error. Coastline extrusion, sensor shapes, heat-volume sizes and particle speeds are explanatory symbols; they are not surveyed buildings, physical parcel sizes or measured motion.
Sources: Mapzen/Tilezen Terrain Tiles; USGS and NOAA terrain attribution; Natural Earth, public domain. Regional outlines are at 1:10 million, displayed with an equirectangular projection at 17°N. Terrain uses a local equirectangular display. Elevation exaggeration is shown above the scene.
Sequential field colours follow viridis; signed errors use a centred blue–orange scale. The colour limits stay fixed as the example runs; values outside them are counted in the legend. The plan view removes perspective. Figure design follows Crameri et al. (2020); pause and reduced motion follow WCAG motion guidance. The browser export retains a source caption. No atmospheric model skill follows from this animation.
Asset versions and checksums · Coordinate definitions · Numerical model source
Source sections
Zero-shot Timestral-5M and Timestral-28M against matched local baselines for Jamaican daily rainfall
Terms explained
Select a term to read its meaning in this study.
Zero-shot
Using a pretrained model without fitting its weights to the local task. This does not rule out overlap between its earlier training data and the evaluation record.
MAE
Mean absolute error. Add the sizes of the errors, ignoring whether they are positive or negative, and divide by the number of predictions.
RMSE
Root-mean-square error. It measures the size of prediction errors and gives larger errors extra weight. Lower is better when the data and units are the same.
Persistence
A forecast that carries the most recent observed value forward. It is simple, but it can be difficult to beat over a short period.
Extreme event
An event beyond a stated threshold. The threshold, location, averaging period and number of examples determine what the evaluation can show.