Alternative Data: The Good, The Bad, and the Useless (?)

Last week I started a new post thread entitled "NO Price Data at All!" to extol the virtues of "alternative" (i.e. non-price) data as a source of alpha. We saw that the Morningstar Fundamentals data set does in fact provide alpha, as can easily be demonstrated, even in the absence of any price data at all. Now I'm going to play devil's advocate and see if I can provoke a few (friendly) arguments that might cast at least some items of alternative data in a rather different light. Here goes......

When you build trading systems/algos, do you have an underlying philosophy? I do, and mine is that a good system/algo should be a blend of art & science that combines various parts together in such a way that a) each part works on its own, and b) all the parts work together in a synergistic and harmonious way to produce something that is greater than just the sum of the individual parts. However, at least for me, each component must be good enough to produce something of value in its own right. If not then take Occam's razor and remove the non-productive complications. All else equal, simpler is better, less to go wrong, and easier to fix if it does. I can't think of any really good trading-related examples where something is genuinely useful but incapable of demonstrating any merit in its own right on a stand-alone basis.

Many of the Morningstar Fundamentals items can clearly demonstrate their value individually. However, the same cannot be said for some other pieces of alternative data. So why should I trust something like that? If a piece of data can't demonstrate at least some small measure of value (i.e. bring at least a little bit of alpha) in its own right, then why bother with such a thing that is evidently useless (or worse...) ?

I have been playing around with stock tweets, specifically the "PsychSignal Stock Twits Trader Mood" data set , i.e. stocktwits.total_scanned_messages, bull_scored_messages, bear_scored_messages. Sure, if I combine these with momentum & some fundamentals I can get a reasonable-looking result. But to me that doesn't mean anything. I already know that the momentum and the fundamentals are good signals, but if I take them away and just play with the twits, I don't see anything that is clearly distinguishable from noise. Therefore I conclude that, although Mr. Trump may enjoy going around twittering & tweeting, actually they are not useful from an algo perspective and only serve to obscure where the alpha is really coming from.

Maybe I'm just doing something wrong and haven't got the little twitty bits connnected up right. Or maybe they really aren't as useful as I had hoped after all. Please, I invite you (.... anyone ....??) to prove me completely wrong about this by posting a purely "twitty" algo that actually has some alpha significantly different from just background noise on any time-frame longer than a day.

Also, if you have a nice working algo that uses twits as well as other input, try just turning off the twitts part and see if it makes any difference.

10 responses

I always find this hard to reason about: can a predictor be useless on its own, but useful when combined?

As a non-trading example: the height and weight of a person can predict, to some degree, how likely it is that person will suffer a heart related condition, but just the height or just the weight won't tell you much (except maybe extreme values of the weight).

For trading: maybe the recent average volatility has predictive value when combined with some other directional indicator, but on its own it might not be very predictive (I know low volatility is associated with a rising stock market, but maybe this doesn't hold for all instruments).

Most of the twitter data sources seem to mostly be about bullish or bearish sentiment though, which is directional per definition. So we'd have to think of a reason that a directional prediction only has value when combined with another indicator, but not on its own.

Hi @ Ivory Ant
Thanks for your comment, and your good analogy of Body Mass Index (BMI) = units conversion factor * Mass [kg] / Height [m]^2 = a measure of obesity, and therefore of health risk, whereas mass & height alone are not. Let's pursue this analogy a little further and see where it leads.

I stated my notion that (at least I believe) each component of a good system should add value in its own right, and what I called "component" there is somewhat similar to what you have called "predictor". I think the key point to note it that each single variable itself is not necessarily a predictor, because an important concept is that of NORMALIZATION. In the Body Mass Index example, height^2 is required to normalize the weight of each person to put everyone on a common basis for comparison and determine whether they are really over-, under-, or normal-weight .... for their height. In a similar way (see another of my posts about Fundamentals) regarding the Piotroski F-score, most of the items require normalization by comparing with their values one year previously. In the Altman Z-score to determine corporate risk of bankruptcy, most of the items are normalized with respect to (i.e. by dividing by) Total Assets, which is a measure of the size of the company from an operational perspective, and analogous to height in your example. It is not a predictor in its own right, but it makes a good normalizing factor for other financial statement items which can then become useful predictors AFTER normalization.

With regard to the "sentiment indicators", or at least the ones I looked at, honestly I don't know if they have any predictive value or not. Maybe they do and I just haven't applied them properly. Or maybe they don't and the stuff written about them is basically just advertising by the vendors wanting to make money selling data irrespective if whether useful or not. I agree with your comment that [if they don't work and some other input is also required] ... we'd have to think of a reason that a directional prediction only has value when combined with another indicator, but not on its own.

With the fundamentals data there is no shortage of evidence that it is useful. With the "sentiment" data i just don't know. I'm open minded, but yet to be convinced, so I'm hoping someone will step forward and convince us with an algo showing a favorable result. :-)

@ Ivory Ant,

I always find this hard to reason about: can a predictor be useless
on its own, but useful when combined?

Simply yes.

In the field of machine learning's feature selection methods there is this recommended read:
In the section "Small but Revealing Examples":
1. Noise reduction and consequently better class separation may be obtained by adding variables that are presumably redundant
2. A variable that is completely useless by itself can provide a significant performance improvement when taken with others
3. Very high variable correlation (or anti-correlation) does not mean absence of variable complementarity

btw, i also have had some trouble extracting alpha from psychsignal. Like you, I could not rule out the possibility that i was doing it wrong. Some algorithms have performed very well, but are hard to extract as a factor (and would not compatible with Q fund allocation).

@Lionel,
Many thanks for this "recommended read". I spent lots of time on this topic when I was playing around with Neural Networks for my own home trading account years ago, but I know that state-of-the-art has moved on a long way since then and I look forward to catching up agin. Best regards, Tony.

Thanks, especially 3.3 is interesting to me:

Can a Variable that is Useless by Itself be Useful with Others?

@Lionel, @Luca,
Thank you. I realize that you are certainly correct about "useless" variables becoming useful ...... AND to make it even more complicated, and also more realistic, at times some "even worse than useless" (i.e. counter-productive) variables can become useful under certain conditions, i.e. there can be sign reversals with respect to their utility.

Here's an example: Let's try to predict the probability that I will get wet in the rain today. We can use some "indicators" such as seasonals (is it rainy season now?), autocorrelation & short-term memory effects (was it raining yesterday and/or the day before?), "fundamentals" (is it cloudy? because if no clouds now then no rain now, and if no rain then certainly I will not get wet in the rain), "alternative data" (what is the wind direction? maybe the wind will bring clouds later). But there can also be un-expected sign reversals in some "indicators". For example, is it already raining RIGHT NOW, before I leave home? If it is already raining before I go, then paradoxically I probably will NOT get wet in the rain. Why? Because in that case I will definitely take my umbrella with me and thereby stay dry, whereas otherwise I might not.

So the question for us becomes how to develop good general methods to find hidden/unexpected useful relationships efficiently, and bypass the sometimes useful / sometimes not useful / ambiguous ones. Now this gets a whole lot more interesting than "curve fitting" ........................

Re: sentiment, maybe sentiment only becomes predictive when no big move (change in volatility relative to the average) already occurred in the last X days.

@Ivory Ant & others,
There is an interesting thread by @jacob shrum entitled: "Robinhood PennyStock Rotation (Using PsychSignal Data) - 100%+/Year" which i want to go through in detail but haven't had time to yet.

What impressed me about Jacob's work is that evidently he is getting VERY GOOD results using exactly the SAME data source that i found to be dismally useless. Now of course it's possible that there is just some screw-up in my code, but i also started to think about whether there might also be any other possible reasons for the differences. Yes, i think there is.....

The first thing i notice in Jacob's code is that he is specifically targeting stocks in the range between $3.20 and <$ 5.00 (not exactly what i would call "penny stocks", but let's just let that go by) whereas in my case there is no upper limit on the stock price, so most of the stocks that i'm working with are larger ones, and with no preference at all for any specific sectors (like tech, etc).

So if we think about who are the sort of people (other than Mr. Trump) who are most likely to be doing lots of twittering & tweeting, and what sort of stocks they are likely to be communicating about, my guess is mostly young high-tech & bio-tech companies, startup or recently graduated out of the startup category at high speed, with lots of growth potential but perhaps (in many cases) still not yet with well established stable earnings records or even with any earnings at all, but lots of "blue sky" potential. Basically the relatively higher-risk higher-reward stockmarket plays, especially appealing to young traders / investors who are most likely to be actively tweeting about their successes.

On the other hand, if we look at a broader cross section of the more mature market, and in particular think about large, well-established companies like GE, or BAC, or KO, for example, these are no longer the hot-shot high-flyer super-growth companies, are not likely to be closely followed by enthusiastically hyped-up twitterers and, even if those people were tweeting about these bigger companies, it is questionable if any serious mature investors would actually care or have their investment decisions regarding those companies influenced much by the tweeters? Probably not.

So my hypothesis (as yet still purely speculative and completely un-tested) is that there is probably likely to be an INVERSE relationship between the success of tweet-based sentiment as a source of alpha and corporate maturity-related factors such as:

• Age of the company (e.g. years since first listing),
• Market Cap,
• Number of shareholders,
• Size of institutional holdings,
• Stability of Revenue,
• Stability of Earnings (if any)

Who knows, perhaps even the postcode of the corporate HQ, e.g.Silicon Valley, or Boston, might be a useful indicator of the effectiveness (or not) of tweeted sentiment as a possible source of alpha .... although if anyone come us with profitable new indicators such as BookValue/Postcode ratio, i would be suspicious of dubious data mining practices and skeptical about robustness!! ;-))