Hacker News Reader

Comment by xyzzy123 | Hacker News Reader

Probably because you've never bet a lot of money on natural language processing before. Backtest and check alpha. cpercival's idea is pretty good though. I'm reminded of: https://xkcd.com/793/

View on HN

The headline of that news was "Intel in Talks to Buy Altera". You are acting like it would be difficult to write software that would be able to figure out what that means. It's not. These headlines also seem to follow a very specific pattern [1].

So if WSJ has a headline that X is in talks to buy Y, that's pretty much that. Also you could eliminate most of your risk by placing a limit order based upon the last trade prior to the news coming out. The market will either a) confirm that you are one of the first people to understand what has happened by filling your order at a price that doesn't reflect the news, locking in profits, or b) your limit order won't be filled because the price already is already above your limit, in which case you have assumed no risk.

[1] https://www.google.com/#q=site:wsj.com+%22in+talks+to+buy%22

let us know how it works out, then. :)

It would be easy to write software to parse that particular phrasing of that particular relationship ("Company X in Talks to Buy Company Y"), but to catch any arbitrary relationship between two entities in a given domain? That's no longer an easy problem. Not to mention that you'll need to represent that world knowledge coherently in ontologies / knowledge bases and write complex logic around it to make the data actionable.

All you need is a positive correlation between an action taken on a weighted evaluation of a headline containing key words, and profit. You're overestimating the difficulty of the problem. It's model training; there's lots of historical data to tune on.

> It's model training; there's lots of historical data to tune on.

Is there a timestamped archive of DowJones Newswire articles? I found it difficult to find archived news articles the last time I was looking for them.

Sentiment analysis has plenty of research around it, and once you've got a big enough training and validation set, it can give very good results. My old boss works for a company that runs sentiment analysis on comments made about companies to automatically highlight positive/negative messages - it's not a massive leap from there to repurpose that to analyse positive or negative news reports about a company.

These bots don't actually do much analysis. They just weight keywords and sources. "Buys" or "acquisition" in a WSJ headline has a pretty clear cut meaning. Financial headlines also tend to be pretty easy to parse (e.g. "X Buys Y").

Do you think "buys" would be less lucrative than "in talks to buy"? I imagine by the time company A buys company B, there has already been rumor and the markets have already moved to reflect that speculation.

Reply on news.ycombinator.com