Thanks everybody, was really a nice discussion and an good turnout. I have also received a few private messages about it, so for those and for the benefit of the attendees, below are some of the discussion points.

• try to use transformations such as C/O - 1 (note these ratios and percentages naturally lend themselves to beta regression and with cauchy errors it seems to make marginal improvements)
• use random forrest or boost regression and just the "financial" data (e.g. open, high, low ...) not social media stuff seems to work
• set the missing values to zero
• segment the sample e.g. on market cap or when H - L is "high" and fit different models to the segments
• fit a combination of main and error correction models where the main model uses financial and the error correcting uses social media stuff
• the asymmetry and outliers are problematic and need to be addressed