[aatchb]

A few years ago I graduated with a PhD in statistics with lots of ML inspiration. Since then I have always dreamed of applying my knowledge and skill in this domain. However, despite the belief I was 'probably' in a decent position to do so, I consistently read about how impossible it was. I have a boring 'normal' persons job, but, posts like this are somewhat reassuring that I made a reasonable decision to abandon a life of fruitless datamining and overfitting.

[alvis]

I am keen to second here. With a PhD in probability & loads of experience in data analytics, my experience has told me that we are too ignorant and sometime too ambitious to try to predict a the outcome of a stochastic process (e.g. financial time series) without knowing that the amount of information required to make a sound prediction is far beyond those we have. Unless there's a very clear dominating signal among thousands of information sources, very often we are trading on noise.

Although I don't necessarily agree with all the points in this article, it just reminds me what Poincaré said:

`You ask me to predict for you the phenomena about to happen. If, unluckily, I knew the laws of these phenomena I could make the prediction only by inextricable calculations and would have to renounce attempting to answer you; but as I have the good fortune not to know them, I will answer you at once. And what is most surprising, my answer will be right.' -- Poincaré, H. (1913) The Foundations of Science. New York, The Science Press. p. 396.

[chillingeffect]

I don't think the message here is "don't do it," but "have domain knowledge." The crux of the paper was scientists applying ML to a bunch of data without really understanding trading.

[dsacco]

You can actually have scientists find signals in data they have no domain experience in. In a typical hedge fund the quantitative researchers will be a different group from the quantitative developers and traders. There are fuzzy lines between those depending on culture, but those three groups are broadly the front office. You really need domain experience for execution and risk management, but pure insights can be derived without necessarily needing any domain experience.

That said, quant researchers typically understand how the market works. They are just able to quickly excel without a background in it.

[murbard2]

It's easy to forget that this is a highly competitive field.

You're used to see the techniques you work with capture signal because there isn't an army of PhDs in math, physics, and computer science working around the clock to trade any signal out of that data.

In the end, it doesn't even matter if you're the best statistician in the world: whatever signal you detect may simply not be worth the effort you put into detecting it.