3 ms·
I'm by no means among the "Greatest A.I. Minds of my generation" but count my among those who tried to use twitter as an input to an "A.I." system and had to fi
by chollida1 11y ago
I'm by no means among the "Greatest A.I. Minds of my generation" but count my among those who tried to use twitter as an input to an "A.I." system and had to finally admit I couldn't tame it.
In my case it was an automated trading system where twitter was one of about 50 different inputs that drove a hidden markov model that spit out buy/sell/hold signals.
I couldn't figure out how to "clean" the twitter stream in real time, either fast enough, or thoroughly enough to make the inputs usable.
Even when I scaled back to using only StockTwits input the data was so noisy that it wasn't usable by me.
It's a very hard problem. Bloomberg spent a lot of money trying to develop new sentiment indicators and after following it for 6 months I found they are no better than a 50-50 guess, and this is a product they want $10,000/month plus for.
- lightbendover 11y agoIt's really not so surprising that you couldn't reproduce Dataminr, a ~$200M company, in your spare time.
- crosbyventure 11y agoDataminr specializes in 'event detection' which is different than deep (semantic) processing.
- crosbyventure 11y agoAI is hard. But we're always here to help! Watching folks build trading neural networks powered by our data has been awesome.
- faitswulff 11y agoWho is "we" in this case?
- traek 11y agoLooks like "we" is StockTwits: https://twitter.com/crosbyventure https://twitter.com/crosbyventure
- jodrellblank 11y agoCan you clean the Twitter stream in your head? Can you, if reading tweets directly, get any kind of sentiment analysis better than a 50-50 guess? (i.e. for one specific stock or industry, or slowly and not in realtime) ? If not, do you know for sure that there is any relevant information content in the Twitter stream?
- adrianN 11y ago50-50 guess is pretty good if you have like 4 different sentiments. Language is hard, short tweet doubly so.
- TheOtherHobbes 11y ago50-50 is extremely good if you're charging $10k/mth. You could skip the NLP altogether and hook up a PRNG and apparently people will still pay you. If there's a lack of intelligence, it's not in the code.
- logicallee 11y agomeanwhile you likely could have sent a DM to anyone who mentioned the given ticker symbol and say "Re your latest tweet - so should I sell or buy?" and then you've simplified it to analyzing the response of anyone who responds for whether their response includes "sell" or "buy". if it's a random sampling you won't even need to tap the same people repeatedly. maybe follow-up DM's to the same person could be "how about now? MSFT, buy or sell?" and so forth. then it's not so blatantly obvious that it's a bot...