IBM SkillsBuild Hydropower Climate Optimisation Challenge
Helping the World
$3 000 USD
Completed (~1 year ago)
Prediction
Forecast
1233 joined
462 active
Start
Mar 03, 25
Close
Apr 13, 25
Reveal
Apr 14, 25
On 'kwh' train/test distribution
Help · 2 Apr 2025, 19:23 · 6
Has anyone tried tweedie as regression objective (irrespective of model type or CV strategy)?
It appears tweedie regression demonstrates strong perf on local cv but not great on public LB.
Anyway, it's still a struggle attempting to harmonize my local cv and public LB, but I'm curious to know if anyone also noticed this trend.
I made this observation whilst experimenting with xgb and lgbm. Happy to know your thoughts.
Yes very poor actually. But good cv. Tweedie gave me like 2 RMSE score on CV but 11 on lb
I see... I got like 1.2 rmse cv but 13.5 on lb
I am not sure that's a great option though.
I guess so
Tweedie is meant for 0 inflated distributions. The training data has this distribution however the test has nowhere close to the distribution you observe in train.
why do i hypothesise that?
- lookat the target shift between last 31 days in training and the data prior to that. my best guess is that in test we would have a distribution similar to the last 31 days in training if not more :)
I see your perspective. I need to look at the data from that angle. Thanks for your insights.