Why is the quant community skeptical about sentiment models?

Econometricus · Post by **Econometricus** » Sat Oct 01, 2016 10:36 pm

I am a newbie to quant finance, so please forgive my naiveté, but I noticed that the quant community is generally quite skeptical about strategies based on sentiment analysis. Would anyone care to elaborate whence such a negative view of this approach?

Maggette · Post by **Maggette** » Sat Oct 01, 2016 11:27 pm

Hi,

I certainly do not represent the "quant comunity" but I think the reason is obvious: the assumption when using sentiment analysis in financial applications is that the bullshit you can find, parse and analyse on the internet is somehow related to people (read or interpreted by them) who actualy do move markets...which most probably is not the case.

Market Makers and the algorithms they use don't give a shit on stuff you can read on social networks and I guess influential traders, hedge fund managers and portfolio managers also do not care.

Kitno · Post by **Kitno** » Sun Oct 02, 2016 12:29 am

Hitherto underlying data signal quality has been low. Also, how do you define "sentiment"? The narrower the definition is the more useful.

I know some guys working on something I regard as of potential value. What's your e-mail (it's not in your profile)?

Maggette · Post by **Maggette** » Sun Oct 02, 2016 1:52 pm

I think what might work is somewhere where retail investors are active and their actions do have impact. Like in penny stocks. Their this might be an interesting approach.

Econometricus · Post by **Econometricus** » Sun Oct 02, 2016 6:23 pm

Thanks for the replies.

@Maggette, I do agree that it is weird to expect the biggest and most liquid stocks to be responsive to retail sentiment, at least in "normal" times. But as you've noted yourself, there seem to be potential applications of sentiment analysis in less crowded areas. Apart from penny stocks, I believe story stocks of the Tesla kind (young company, negative cash flows, uncertain prospects but a LOT of hype around) could be a potential target. What do you think?

@Kitno, absolutely, definition is crucial. I've been looking at news- and Twitter-based measures. My guess would be that market-based measures (i.e. implied vola etc) are already more than well-researched and are unlikely to provide much valuable insight. Or am I wrong?

My e-mail is sereku37(at)gmail(dot)com

Feel free to get in touch, I'd love to exchange thoughts on this.

curvefitter · Post by **curvefitter** » Sun Oct 02, 2016 7:02 pm

Can't speak for the community either but if your sentiment analysis is based on twitter data don't complain when someone buys a botnet to flood twitter with made-up sentiment in order to influence your actions..

EspressoLover · Post by **EspressoLover** » Sun Oct 02, 2016 10:39 pm

Because talk is cheap. We already have a very good signal to gauge sentiment, it's called market activity. And unlike twitter, it involves a pretty credible commitment to your beliefs. If Alice is betting millions of her own money against thesis X, and Bob writes a nice little story defending X, who do you trust? When sentiment works, it's usually because it just re-confirms existing market activity. Sentiment without market activity usually means nothing. E.g. Facebook made a cool robot that everyone's talking about, but isn't going to move the bottom line. In contrast market activity, without visible sentiment, is usually still very important. There's plenty of very well-informed participants who trade without uttering a word.

At the end of the day, sentiment-funds are competing in basically the same space as traditional stat-arb. I do think sentiment does have some minor amount of mutual information to a good stat-arb system. Maybe 5% optimistically. Most of that coming from interactors rather than direct impact. In contrast the traditional complex of stat-arb alphas probably have 100%+ mutual information to the best sentiment signals. All that being said, sentiment probably does have a place as a small sub-division at major, mature stat-arb desks. But a fund trades purely on sentiment alone has a snowball's chance in hell at competing in anything other than sexy buzz-word marketing.

goldorak · Post by **goldorak** » Mon Oct 03, 2016 5:15 am

EspressoLover, your comment reminded me of a paper. price alone vs price+google trend vs google trend

ronin · Post by **ronin** » Mon Oct 03, 2016 12:21 pm

It's not even true that people don't pay attention.

A few years ago there was a mini-trend in event-driven funds automatically parsing news stories for sentiment and using it to trade. Reuters was trying to charge a fortune for some of those news feeds.

But I haven't heard of anybody making money with that. Like everybody else said, too much noise, and too easy and cheap to manipulate - on either side.

"X&Co paid their CEO a hundred million."

"Y&Co is squeezing small suppliers."

"Z&Co Head of HR reveals her interview secrets - how to get hired."

"W&Co donates money to farmers in Africa."

That influences the values of these companies how exactly?

Kitno · Post by **Kitno** » Mon Oct 03, 2016 1:01 pm

I think the only way this works is if you cut out intermediaries (like news agencies) and have a high quality source at the origin.

Twitter has to much noise and lacks consistency from the few high quality content generators, which is why I suspect it's failed.