Transparency reporting

How accurate are the forecasts?

No weather forecast is perfect. This page shows how closely Sunsethue forecasts matched what people actually reported using the in-app feedback reports, so you can set realistic expectations and see how the model keeps improving.

Data from Sep 3 – Nov 2, 20261,784 reports

All forecasts

Total accuracy

60.8%

The share rated mostly right or spot on (3 or 4 out of 4), regardless of whether the underlying weather forecast was correct.

1,784 reports

Sunsethue model

When cloud data was right

76.1%

Accuracy among reports that said the predicted cloud cover matched reality. This better isolates Sunsethue’s own quality model.

1,102 reports

Weather model

Cloud-cover accuracy

68.4%

The share of optional cloud-cover responses marked accurate. This measures the weather input that Sunsethue receives.

1,610 reports

Reasons for wrong forecasts

A final forecast can be wrong for two fundamentally different reasons. Both the weather model and Sunsethue's ray-based algorithm on top of this data introduce errors. Besides this, the quality of a sunrise or sunset is something subjective. Differences in how people rate the forecasts can also contribute to some reported errors.

How you can help: honestly report every sunrise and sunset you witness, regardless of the forecast or how impressive it was. Reporting disappointing events as well as beautiful ones, not only the good sunsets, helps keep these accuracy statistics representative. Accurate information on the accuracy of the forecasts is the first step to improving the model further.

Weather model

The cloud forecast was wrong

Clouds may arrive earlier, later or in a different place than predicted. Even if Sunsethue interprets the input perfectly, an incorrect weather forecast can make the final result wrong. The sunset quality is very sensitive to the exact position of the clouds, a small misprediction by the model can lead to a wrong forecast.

Sunsethue model

The quality interpretation was wrong

The clouds matched the weather forecast, but the ray-based model still misjudged the colours. The model has to make some assumptions and approximations to be feasible to run it 4 times per day, that's why even when clouds are predicted correctly, the model can still sometimes miss something.

Learn more about forecast uncertainty

Future plans to improve accuracy

Future work will target both failure cases, with the current focus being on improving the weather data source, because that is the main cause for mispredictions.

More weather-model options

Compare and combine more weather models, including higher-detail regional models where available, to improve the input cloud forecasts for the sunset prediction model.

Higher spatial resolution

Run the Sunsethue model on a finer grid so smaller cloud structures and local differences are represented more precisely.

More atmospheric detail

Use more vertical cloud levels, include metrics like fog and haze, and refine model settings to better capture how light travels through different layers of the atmosphere.

Forecast rating distribution

How users score the sunrise/sunset quality forecasts.

Spot on60634.0%
Mostly right47826.8%
Somewhat off37421.0%
Way off32618.3%

1,784 responses

Was the cloud cover right?

How users score the predicted cloud cover (optional question).

Accurate1,10268.4%
Not accurate50831.6%

1,610 responses

Reporting bias

Reports are not a random sample

People may be more likely to report a memorable, colourful sky, or to send feedback when a forecast was wrong. Both behaviours can affect the total accuracy in different directions.

  • •An under reporting of poor and fair forecasts as compared to their frequency of occurence.
  • •Clearly more reports for good and great sunrises and sunsets.

This comparison shows that there is some bias in the reports made by Sunsethue users, which can influence the results on this page. This comparison reveals where the feedback sample differs from the global baseline; it cannot tell us exactly why someone chose to report.

Sunrise vs Sunset

The accuracy is about the same for sunrises and sunsets, sunsets get a lot more reports.

GroupAccuracyAvg. ratingReports
Sunrise
60.3%
2.74 / 4642
Sunset
61.0%
2.78 / 41,142

By predicted quality

Better sunrises and sunsets were correct more often than poor and fair forecasts. Reporting bias can play a role here.

GroupAccuracyAvg. ratingReports
Poor0–20 predicted quality
50.0%
2.51 / 4350
Fair20–40 predicted quality
50.5%
2.51 / 4594
Good40–60 predicted quality
68.3%
2.90 / 4467
Great60–80 predicted quality
79.9%
3.28 / 4318
Excellent80–100 predicted qualityInsufficient data—49/ 100

By region

The forecasts have about the same accuracy in all three regions with sufficient user reports.

GroupAccuracyAvg. ratingReports
North America17–54°N, 132–58°W
61.3%
2.79 / 4910
Europe35–65°N, 16°W–44°E
62.2%
2.81 / 4590
Australia48–8°S, 100–180°E
60.1%
2.68 / 4148

By predicted cloud cover

Most ratings came in for sunrises and sunsets with either little or many clouds, the in-between partly-cloudy forecasts perform slightly worse.

GroupAccuracyAvg. ratingReports
0–25%Mostly clear
63.3%
2.77 / 4381
25–50%Partly cloudy
62.7%
2.69 / 4142
50–75%Mostly cloudy
57.6%
2.67 / 4125
75–100%Cloudy to overcast
60.0%
2.78 / 41,130

How to read this page

Details about the methodology

These results describe recent community reports, not a controlled scientific sample. People choose whether to report, locations are unevenly represented, and visual judgement is subjective. The main goal of this page is to inform you about the strengths and limitations of Sunsethue's forecasts.

Minimum sample

A percentage or average only appears once that exact group has at least 100 reports, to make sure the data is sufficiently representative.

What “accurate” means

Ratings 3 (“mostly right”) and 4 (“spot on”) count as positive. The average score is the average among all ratings.

Self-reported sample

Results are shown as submitted and are not reweighted to match the worldwide forecast distribution. There is no filtering on the reports, they all come straight from the users.

Update rhythm

The snapshot will be refreshed manually about once a month and always covers the preceding 60 days.

Last generated Nov 2, 2026 · 1,784 reports · 1,610 included an optional cloud-cover response