Thread by @TerribleQuant

Early poll results indicate feature examples are wanted. Let's do sentiment.

>> A word of caution: alternative data is not a holy grail, there is loads of it and it is hard to gain inference from. It also is not a massive part of the variance.

1/n

Plot sentiment, and it tends to look like noise. There is so much data for sentiment that it doesn't really matter what model you are using versus how you use it. More data usually means you need to think out of the box, because the marginal increase of model performance...

2/n

gets smaller and smaller, and there is a LOT of text data. If you are a beginner just looking for alpha, just get out something like textblob or spaCy, don't waste time with LSTMs and CNN and other NN models. First step, understand your flows and how you should...

3/n

weight the data. Your model will be able to grasp some of the relations between sentiment and its impact, but how will it know the source. Get the data for viewership when weighting your sentiment. Is it good/bad, how sure do they sound, what is their viewership? ...

4/n

The next part is just using headlines, full articles give very little extra inference, and you should manage time efficiently to get other features. Analyst ratings help too. Model how different classes/ caps are affected by different sentiment sources. There...

5/n

are even models with plenty of alpha for using CEOs' facial expressions on CNN, but don't use a CNN model XD, there are far better models for computer vision nowadays. Google searches as well are great. The key here is to remember that sentiment doesn't directly drive the..

6/n

market and there are heavy reflexive effects from other algorithms using them. So use statistical tests like Portmanteau Statistic, Hurst, KPSS, T-Values, etc to build meta-features. If you are doing HFT, you can use microstructural flow metrics like Kyle's Lambda...

7/n

and all the different variations of VPIN that exist can be ensembled like GARCH models would be. Just don't spend too much time playing around with sentiment, although it can be good if you prepare it well. If you are doing an exclusively alt data approach or being...

8/n

mid freq (but still intraday) or daily, just not HFT, then you will need to isolate the idiosyncratic effects of sentiment and alternative data variance since otherwise the noise will be overpowering. This is why you should make sure your portfolio is constantly hedged...

9/n

This will hurt your alpha in the short term, but on the mid-daily timeframes (sentiment means fuck all long term) you will need to stay delta neutral by shorting your top xx (maybe exponentially weighted) sells and longing your top xx buys...

10/n

When you do this your effects become more idiosyncratic so focus on sharpe instead of returns as returns will likely be smaller because it is a smaller part of the variance. You can ensemble this to improve capital efficiency or just lever it up and use...

11/n

a modified kelly that assumes way fatter tails than you observe in the implied PD. Careful with leverage of course as it eats into profits and is risky (I personally don't touch leverage one bit).

DONE

12/12