I'd always go with NaN. In dataprocessing, it's common to encounter NaNs due to bad data - these records can be filtered out to allow the process to continue. T

I'd always go with NaN. In dataprocessing, it's common to encounter NaNs due to bad data - these records can be filtered out to allow the process to continue. The records which resulted in NaN can be flagged for follow-up. As a program author, I'd expect that it's my choice how to handle this situation. I don't want libraries making arbitrary decisions about how my code should work.