Improve sale rate computations

[opatut]

I didn't get this done before the event, but here is a few things I'd like to do:

• The linear regression is bad because it allows negative stock predictions and isn't capped at 0, so the sum of the stock prediction functions is useless because "out of stock" is not handled at all.
• We should use segment-wise linear interpolation for all known count entries to get a more variable sale rate over time.
• We have to filter outliers especially in the time axis. No need to look at data points before or after the event.
• Maybe we can differentiate between items handed out via preorder and actual sale.