When building a system that surfaces interesting trends, you have to be careful about curating which historic examples you show. People will craft their unders

When building a system that surfaces interesting trends, you have to be careful about curating which historic examples you show. People will craft their understanding of predictive power based on those examples... but if the examples aren't representative (i.e. randomly sampled) of the algorithm, then you're selling people rainbows and unicorns.