Predicting Sept Minima to Within 2% using Jan-Feb regional extent data

[SimonF92]

I made this over the last few weeks and to be honest i find the prediction accuracy remarkable.

It uses the MASIE regional extent data to predict the Sept minima for that year, only letting the model train on the Jan and Feb data for that year, on a sea-basis. I had a feeling the regional data was far more powerful for predictions than total data as it describes what will melt quickly where for each year. It is accurate back to 2006 to within 2%.  Its so accurate that I actually dont believe it.

Is there anyone out there who can check this for errors/mistakes? Is there anyone who can convince me that this is actually reliable? Can anyone recommend a cross validation test for this kind of continuous data?

FIRST image is the prediction if the model is given all Arctic seas
SECOND is if the model is given 5 Arctic seas
THIRD is if the model is given 10 Arctic seas

Accuracy does appear to scale with how many seas it is given to train on as features

https://github.com/SimonF92/Arctic/blob/master/Masie_Neural_Network.ipynb

[crandles]

Is this a case of giving it 15 data series to predict 15 outputs and it gets it perfect? ie overfitting?

Can you train it on 8 years and see how it does on the 7 years it hasn't been trained on?

I could be talking rubbish as I am not very sure/ not really investigated what machine learning/neural net you are using.

[SimonF92]

Is this a case of giving it 15 data series to predict 15 outputs and it gets it perfect? ie overfitting?

Can you train it on 8 years and see how it does on the 7 years it hasn't been trained on?

I could be talking rubbish as I am not very sure/ not really investigated what machine learning/neural net you are using.

Thanks crandles, I was thinking about overfitting. I will try this and see how it looks; theres a train:test split function in sklearn, its for classification data but should be workable.

It gets a shape of (60 days x 15 seas) per year.

Its a convolutional neural net so a little more tricky than what i have worked with previously, hence why i am suspicious

[uniquorn]

Tried running this but got a 'no keras module' error. Probably due to running spyder without anaconda. Sometime I'll update everything properly, good luck.

[SimonF92]

Spyder without Conda? Dark magic indeed 

[uniquorn]

thanks, will try that. Clearly flying spyder without a licence here.