First Rule as a Quant: Think Nonlinear...

...because financial data is not linear!

Fat tails do not exist

…

in textbooks ;-)

Well, let’s look at some the most important Features (out of > 25 AI Factor Systems):

[…]

FINANCIAL DATA IS NOT NORMALLY DISTRIBUTED

It has fat tails. We can love it. Hate it. Better to accept it.

Especially the most important Features / Factors have fat tails.

ChatGPT:

• “Total Return Strategies Are Naturally Fat-Tailed”

• “While the trimmed [cut the left and the right tail of the distribution] dataset is less volatile, it still underperforms on a risk-adjusted basis, meaning that keeping the tails produces better returns even when accounting for volatility.”

I had to read that three times, yikes!

So, which tools can capture the right tail of the distribution and avoid the left tail?

• NOT (!) Statistical Methods that assume a normal distribution of the data (remember LTCM?). They cut the left and the right tail of the distribution. And based on that you can produce beautiful math and statistics. I get it, the math is beautiful, but unfortunately the assumption is false.
  Karl Popper wrote critical about that since the 50s (“Assumptions of economic models are off”), Taleb writes about that since 2007.

• Ranking Systems do a good job (the expression is linear, but the difference between a 99% ranked stock and a 95% ranked stock is not, Ranking Systems implemented right make no assumption about the distribution of the data).

• Non-linear Machine Learning Models do the best Job
  

Here is my cup of tea:

Accept reality, use methods that handle nonlinearity well and build total return strategies (that capture the right fat tail) in the first place and (if you must) take care about the volatility later.

https://www.portfolio123.com/port_summary.jsp?portid=1826066
Unhedged:

Hedged:
https://www.portfolio123.com/port_summary.jsp?portid=1825610

Best Regards

Andreas