Fossil Demand Forecasting Challenge
$5 000 USD
Completed (over 3 years ago)
Forecast
1009 joined
200 active
Start
May 24, 22
Close
Aug 28, 22
Reveal
Aug 28, 22
Our Solution
Notebooks Ā· 1 Sep 2022, 09:01 Ā· 6
I was waiting for the top 3 to share their solutions, but sadly no one did.
Btw, for anyone interested here is part of our solution:
This is not the final one, but it's very similar to it. This would give a score about 165000.
Please put a star if you found it useful.
thanks for sharing!!
I was wondering how you checked your model performances. Your notebook doesn't seem to include cross validation part. Does this mean, you didn't check your perfomance before submission? maybe check how the model improves on public leaderboard?
I want to ask one more thing... is Catboost the only model in your final solution too? If yes, why did you not try any other popular models, such as lightgbm, xgb...
thanks in advance,
The cross validation part was a bit messy in my notebook, so I deleted it from this notebook. But, I just used the last 4 months as a validation set, and the rest of data as a training set.
Tbh, I didn't have a lot of time as that I entered the competition a little bit late. So I just stucked with catboost and foucs on feature engineering.
thanks for your reply...!
i see your point.
I asked question because i wanted to know how other participants conducted validation. tbh i stuggled a lot to appropriately validate my model. also my CV score didn't indicate leaderbord score well...
by the way, did your CV score has good correlation against submission score (public/private)?
Yes, it had a very good correlation with the lb. When forecasting just make sure you make your validation period the same as the test set (4 months) and avoid adding features lagged less than this period.
Also the train set must represent the past values, and the validation set must represent the future.
Hey Mohamed, thank you for sharing, I also went for a catboost :). I tried other models (RF, Arima, lgbm, xgb, MLP) but when I checked the performance using cross validation, the performance fluctuated a lot so I went for a catboost.. at least the score was stable through the testing sets and it will perform well in the future I hope :)
It's my first competition on zindi, can we share the code before/after the final leaderbord ? If a zindi staff member can tell me I would be happy to share too !
Well Done! Getting the 2nd place in your first competition is amazing!
I think you can share a brief summary about your approach whatever is your place in the lb. Not sure about the code though especially if you were in the prize area.