Relative MSE

[gduthe]

Hi @NickGeneva! Thanks for the nice work!

I'm working on re-implementing the relative mean-squared error (MSE) metric as described in your paper. However, I'm having difficulty finding a clear, universally accepted definition of relative MSE in this context. Could you please provide the exact formula you used to calculate these values?

Specifically, is the relative MSE calculated as:

• MSE(y_pred, y_true) / MSE of y_true against a baseline (like mean pred, last pred or zeros)
• MSE(y_pred, y_true) / Variance(y_true)
• MSE(y_pred, y_true) / ||y_true||₂
• ||y_true - y_pred||₂ / ||y_true||₂
• Something else?

Thanks!