I've been working with ICON, ICON-EU and GFS forecasts, and it seems like the temperature results are "close" for the same location, 6 hour forecasts and 120 hour forecasts - MAE: 5.7 and 6.87 respectively. Is it usually the case and if it is, I would like to find out the reasons briefly.

I'm fetching forecasts from global models every 6 hours. Forecasts are in the range of 0-144 hours, hourly incremented. I calculated the error metric by using only 6 hour forecasts vs 120 hour forecasts.

Dataset example below: You can see that models forecast the next 144 hours in every run and I fetch every 6 hours so base_run_timestamp increments every 6 hours. (Base_run_timestamp is when the models made their forecasts and forecasted_timestamp is pretty self-explanatory)

Note: I realized this is a consistent error for my observation point, I applied simple linear regression and the error went straight down to ~1°C so high error is irrelevant.

Thank you!

DS_1 DS_2

  • 1
    $\begingroup$ I think I can answer your question if you provide a little more detail. What is the time step for the different model runs (is it even different?) ? What is the origin of your numbers? A mean average error of 5.7°C for a 6 hour forecast seems far from reasonable to me. Which time stepping method was deployed (this is the most important detail) ? $\endgroup$ Mar 22, 2022 at 13:37
  • $\begingroup$ @J.Fregin Sorry for the late response and thanks for your interest. I've updated the question with examples. $\endgroup$
    – UgurZCifci
    Apr 21, 2022 at 12:40
  • 1
    $\begingroup$ Sounds like you are probably in an area with significant terrain variation over short distance. And so you're at a different elevation than your gridpoint? Hopefully regression can work well often (though I'd expect there may be times it doesn't, times when the weather varies more by elevation) $\endgroup$ Apr 21, 2022 at 13:26
  • 1
    $\begingroup$ I think otherwise the basic key questions are to try to make sure everything is being done properly... how big a dataset are you using and what data source are using to verify against the forecast. Also, why are you doing just 6 hr and 120 hr, if you have a full database and are well familiar with this stuff, I would think you could pretty easily code to score all the forecasts and that might point you towards any issues you're having. $\endgroup$ Apr 24, 2022 at 12:30
  • $\begingroup$ Because even for a place with a systemic error like poor terrain resolution, the degree of error should change in most situations over that range just because models are poorer at longer range. And you're also comparing two different times of day (6 hours separate), so even if it were something straightforward like elevation/weather, that shouldn't be as constant for different times of day necessarily. $\endgroup$ Apr 24, 2022 at 12:33


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.