4 ms·

Correlating financial time series with micro-blogging activity (2012) [pdf]

26 points by ColinWright 5y ago

Bostonian 5y ago
The paper is from 2012 (which should be mentioned in the post title), which explains why it looks at tweets but not at Reddit posts. Has there been published research on the impact of wallstreetbets Reddit posts?
- WolfOliver 5y ago
  Did the popularity of Reddit change that much?
  - Kkoala 5y ago
    From 2012 compared to now? Yes, even 2018 to 2019 was 30% increase in active users https://redditblog.com/2019/12/04/reddits-2019-year-in-review/ https://redditblog.com/2019/12/04/reddits-2019-year-in-revie.... According to this https://backlinko.com/reddit-users https://backlinko.com/reddit-users they have 10x their active users from 2012 to 2019. And then there are wallstreetbets and bunch of other subreddits that it spawned.
plaidfuji 5y ago
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.
- alpineidyll3 5y ago
  When evaluating trading systems looking at returns over time is meaningless because of the large fat tailed variance of financial price series. Instead if you want to quantify anything in a rigorous way you need to think about units of information and tests of null hypotheses...
  - kruxigt 5y ago
    Can you elaborate a bit?
- regularfry 5y ago
  At the time this was written, I think I'm right in saying that the Twitter Firehose was something you could just ask for access to. I'm sure someone here can correct me if it was more complicated than that.
chollida1 5y ago
Yes that is around the time we implemented sentiment analysis into our trading systems. I remember the firehose and $<Ticker> symbols, good times. Though it didn't last. If I remember correctly it lost its predictive value around 2014. Bloomberg came out with a sentiment product right around then and nlp libraries got good enough that anyone could do it in under a month of work, heck now you see weekend projects almost daily on reddit of people who are still trying to do sentiment analysis on twitter, mostly for crypto to trade off of. There was a time a few years ago when you could use the Robinhood hot list to trade from, well technically you'd trade against it as what was hot on Robinhood inevitably crashed as those types of trades have no conviction or fortitude to hold and they're always pulling money to put it into the next shiny thing, ex GME of course, that one seems to have sticking power!!
- noam_compsci 5y ago
  Please post a link to where these weekend projects are
  - chollida1 5y ago
    reddit.com/r/algotrading

26 points by ColinWright 5y ago

The paper is from 2012 (which should be mentioned in the post title), which explains why it looks at tweets but not at Reddit posts. Has there been published research on the impact of wallstreetbets Reddit posts?

WolfOliver 5y ago
Did the popularity of Reddit change that much?
- Kkoala 5y ago
  From 2012 compared to now? Yes, even 2018 to 2019 was 30% increase in active users https://redditblog.com/2019/12/04/reddits-2019-year-in-review/ https://redditblog.com/2019/12/04/reddits-2019-year-in-revie.... According to this https://backlinko.com/reddit-users https://backlinko.com/reddit-users they have 10x their active users from 2012 to 2019. And then there are wallstreetbets and bunch of other subreddits that it spawned.

plaidfuji 5y ago

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.

alpineidyll3 5y ago
When evaluating trading systems looking at returns over time is meaningless because of the large fat tailed variance of financial price series. Instead if you want to quantify anything in a rigorous way you need to think about units of information and tests of null hypotheses...
- kruxigt 5y ago
  Can you elaborate a bit?
regularfry 5y ago
At the time this was written, I think I'm right in saying that the Twitter Firehose was something you could just ask for access to. I'm sure someone here can correct me if it was more complicated than that.

chollida1 5y ago

Yes that is around the time we implemented sentiment analysis into our trading systems. I remember the firehose and $<Ticker> symbols, good times. Though it didn't last. If I remember correctly it lost its predictive value around 2014. Bloomberg came out with a sentiment product right around then and nlp libraries got good enough that anyone could do it in under a month of work, heck now you see weekend projects almost daily on reddit of people who are still trying to do sentiment analysis on twitter, mostly for crypto to trade off of. There was a time a few years ago when you could use the Robinhood hot list to trade from, well technically you'd trade against it as what was hot on Robinhood inevitably crashed as those types of trades have no conviction or fortitude to hold and they're always pulling money to put it into the next shiny thing, ex GME of course, that one seems to have sticking power!!

noam_compsci 5y ago
Please post a link to where these weekend projects are
- chollida1 5y ago
  reddit.com/r/algotrading