Not yet. We haven't collected historical discussion data (prior to June 2018 for popular coins). But it's something I planned to do in the next few months.
I doubt the algorithm would overfit for negativity. The supervised portion is trained on unrelated datasets. When we collect historical data (e.g. late 2017), or when the market recovers, I can verify.
That's true, unfortunately :) But there are still a lot of people consider bad market as an opportunity to build something useful / to experiment with their investment strategies.
Is the previous year available? Have the algos overfit for negativity?