Custom loss gradient calculation

[Arminius4]

Hey,
In the custom loss function notebook for lightgbm ln[16], there is the line grad = np.where(residual<0, -2*10.0*residual, -2*residual).
Why is the gradient negative here? The corresponding question on the medium article has not been answered yet and I stumbled upon the same issue.
Could you please explain why it doesn't work if the "mathematically correct" gradient is returned?

Best regards,
Lukas