Can we get negative mean square error?

I am getting negative mean square error, while training a SVM model. What is confusing me is that, squared errors should not be negative, is this any bug or is it something that I am missing?

[Matt J]

I would view it to be a bug, but it's possible that MSE is calculated with constant terms dropped. That is, if MSE has the general form,

MSE(θ) =

the term is independent of the design parameters θ and can therefore be dropped without affecting optimization. This would allow negative values.

[Soumyabrata Bhattacharjee]

Yes Sir, I too believe it to ve a bug, as in this case, I am working on a time series problem, so y is not independent of the predictors, rather it itself is the predictor, albeit, past values.

[the cyclist]

Are you able to post the data, and the model that gave negative MSE? I know that sometimes R^2 can be negative (when constructed in certain ways for an ML model), but I am also surprised to get MSE negative.

[Soumyabrata Bhattacharjee]

Hey Cyclist, thanks a lot for your response. I am using Optimizable SVM, from Statisctcs and Deep Learning Toolbox. I have attached the data file, for your kind perusal. Please let me know if I can provide any more details.