Predictions should _never_ be below zero.

[huevosabio]

It might be worthwhile to look into log-based scales again:

• They are never going to give predictions below zero
• I suspect (we can discuss this next week) that our slight biased might be helped by assuming that the logs are normally distributed (https://en.wikipedia.org/wiki/Log-normal_distribution).