To solve this problem I used a second NN to quantify the news to get a sentiment which I also included for the price prediction. At the moment I am using headlines and started also including tweets from twitter.
I would like to know what you think about this approach.
I find this interesting but I'm pretty skeptical of the results you could produce at this point. What kind of logic do you employ to quantify/qualify news sentiment or Twitter sentiment? I agree that sentiment is extremely important to trading this market, but reading sentiment is highly subjective. In my case, it's taken years to become acutely aware of trader/speculator sentiment and learn to trade against it.
The most useful indicator that I can think of (in this context) would be detecting decidedly bearish vs. bullish opinions/charts to give a broad overview of position commitment. When everyone is bearish, you should expect bullish resolution. When everyone is bullish, you should expect bearish resolution.