Thread by @choffstein

1/ 📚 Random Research Note

While doing research in 2006-2007, one of the interesting phenomenon I found was that sampling a data on a non-time-related domain created beneficial statistical properties.

2/ In other words, I would transform a time-series so that the x-axis was no longer a constant time measure, but rather something like constant volume, ticks, or even variance.

3/ Suddenly, "returns" became a lot more normal and stable looking.

(Image source: github.com/Jackal08/finan…; credit @JacquesQuant)

4/ My hypothesis for this phenomenon was that information flow into the market was not constant over time. So by sampling in the time domain, we were over-sampling during calm periods and under-sampling during chaotic ones.

5/ You can "see" this by looking at something like accumulated variance in the time domain.

6/ I had never heard anyone else really talk about this until I recently read @lopezdeprado's book Advances in Financial Machine Learning.

Paraphrasing from github.com/Jackal08/finan… (credit @JacquesQuant):

7/ Beyond the potential statistical benefits, there are other interesting aspects of domain transformation as well.

For example, an indicator in a fixed volume domain might appear *dynamic* in a fixed time domain.

8/ e.g. Here I take $SPY and transform it to a fixed accumulated variance domain, run a 200 bar SMA, and then transform it back to the time domain.

Note how much "faster" it tracks the market in 2008 than the fixed 200 day SMA.

9/ Obviously this has interesting applications in the trend-following domain.

But it might in others as well (e.g. rebalancing on a fixed accumulated variance vs fixed time horizon).

Back to the lab... 👨‍🔬

FIN.