Seems like without any sentiment analysis, there’s a strong correlation with daily trading volume, but not as strong with actual stock price. Still, they manage to build a model+trading strat that outperforms DJIA by about 5% over 90 days.

I’m always curious what the cost and feasibility would be to implement these simulated trading strategies. Seems to me you would need a couple more things to actually do this: 1. Consistent access to near-real-time tweets or other scraped blog content at very low cost via an API, and 2. A service that allows you to make very frequent automated trades at very low cost, also via an API. Or a better modeling strat that knows how to find the features that correlate with a steady increase in price over time to reduce required trading volume.

I assume this is basically what high freq trading is, but the cost of access to those two services is probably so large that you need your base capital to be in the tens of millions, or to have a ridiculously performant model (like +20% vs indices) to actually make a profit.