Thread by @HangukQuant

Sub-Thread 4

Numerical Expression of the Thought Process

“Horizontal Alpha Research”; Alpha Fracking

Alpha Probing

📌 Numerical Expression of the Thought Process

1)

-1 * cor ( returns(1) , delay( returns(1) , 1 ) , 90)

“The alpha is the negative correlation between time-series of daily returns and its 1-lagged series, taken across, say, 90 days.”

2) The strategy of numerically ranking the outputs from above and going risk-neutral L/S quantiles on 30 stochastically sampled S&P assets:

3) I call this the “vertical” alpha research, where a hybrid approach was adopted to find pricing factors that persist “in infinite-dimensional spaces of market factors”.

📌 “Horizontal Alpha Research”; Alpha Fracking

4) Next we like to perform “horizontal” alpha research.

“Alpha fracking” in the subspace of the region identified by the vertical. Once we strike oil, we extract as much alpha as we can like a greedy shale company.

5) If our set of parameters, {1, 1, 1, 90} had “profitable edge”, who is to say that those parameters were optimal? The alphas should persist in a reasonable range of parameter values in the 4-dimensional Euclidean space of settings. A related thread

https://twitter.com/hangukquant/status/1437052690077392902?s=21

6) The return surface should be smooth, and profitable for a large combination of values. If it is has sharp peak/jumps along parameter axis, there is a chance that our alpha was a “false positive”. It shows us the “path dependency” of our return distribution in the alpha space.

7) Simulating on different correlation window lengths, we see that most are stable/profitable:

8) We can also vary other parameters:

9) Varying these parameters also allows us to obtain multiple variants of our strategy, that are not perfectly correlated. This has multiple advantages:

https://twitter.com/hangukquant/status/1443119514707628039?s=21

10) A separate case study displaying enhancement in risk-adjusted returns was written in my newsletter:
hangukquant.substack.com/p/1500-trading…

📌 Alpha Probing

11) Another technique we like to perform as part of the “horizontal” alpha research is “alpha probing”. If the alpha worked for a particular dataset, why should the trading edge not exist in other markets?

12) We do stochastic sampling over multiple datasets, ensuring different combinations of assets, different markets (US, emerging, commodities, FX), and hope to see that majority of the samples display positive returns.

13) Simulating on Different Datasets

14) These are just some of the techniques we apply for our “premium alphas”, and we perform other numerical tests such as alpha decay (how long alpha gets exhausted), signal decay (how long alpha’s signal gets exhausted) et cetera.

15) MAIN THREAD

https://twitter.com/HangukQuant/status/1445464869587746818