Hi Goran, thank you for reach out to us. I am sorry, but in this case is an expected result. The first 'problem' I see is the lack in data, if you see you have only 11 datapoints, where around 30% of them is more 'negative' than the others, and two of them are at the end of the datapoints which can introduce a negative slope.
If you see my example (attached) I added at the same 'negative-values' more of them, thus now the forecast function is more stable.
And, regarding the percentage under zero, I can see in the y-axis is in a numerical/decimal scale, not percentage. You could try to use percentage values/scale instead.
In here
https://www.gooddata.com/docs/cloud/create-visualizations/smart-functions/ you can find more details about the Forecasting feature.