Why are prediction intervals on fitlm NaN

Using

mdl = fitlm(x_mlr,y,'quadratic'); [y,pi] = predict(mdl,x,'Prediction','observation');

results in perfectly fitting model but in a pi matrix containing NaNs only. Any suggestion why this is the case. It works fine for:

mdl = fitlm(x_mlr,y,'linear');

Thank you in advance