Can you give us some more details about the problem you're solving? Like what type of SVM is it? how is the acc measured? Did you compare the inference results with the ground truth? (or else how can you measure acc), what does the 1day 2day ahead mean?

For the SVM, we used libsvm, https://www.csie.ntu.edu.tw/~cjlin/libsvm/

as it uses RBF.

For the data set, its a 24-Hourly dataset of precipitation satellite images.

so for 1day ahead prediction, we take 2 images as an input and we predict (rain/no-rain) to third image in the sequence (sliding window). If it were 2day ahead prediction, then we would skip the third image and predict the fourth.

we used accuracy_score to get the above results against ground truth data.

Okay so basically you're training a kernel-SVM. Okay, LIBSVM is a package to train a SVM; it does not necessarily use RBF. You can use RBF or polynomial, etc. to train your classifier.

Anyhow, to be able to answer your question, we need to know how are you calculating the loss function? Is it the hinge loss function? Have you performed grid search to select your hyperparameter values (C and gamma)?

yes we did, and we got different C and gamma for each model

we wrote that in the terminal: python grid.py 24train21.txt

to test the performance after doing grid search we did

1- svm-train -c 2 -t 2 -g 0.00012 24train21.txt

2- svm-predict 24test21.txt 24train21.txt .model output_file

and the numbers above is from using classification accuracy

Classification Accuracy = (correctly predicted data / total testing data) × 100%

I honestly dont know more than what I wrote now, so I suppose that the classification accuracy is the same as a loss function

Thank you for letting me know the fact that it can use different kernels

and also excuse me if I am little confused, first time to use SVM in general