Thread by @TerribleQuant

I think you can probably classify modeling and feature engineering into a few areas: ML, Statistics, Time Series, Microstructural, some fun extras like entropy which are super weird to work with TA, data-mined alphas, and signal processing.
1/who knows lol

I'll probably speak on each of these eventually, but today I think it'd be good to get some publicity on signal processing. It's underhyped compared to ML and just as deserving.
2/

A lot of the literature is exclusive to electrical engineering and CS, but I can tell you there is lots of alpha in the area. As the story usually goes NN models like LSTMs get a bad rap performance-wise bc of their terrible application
3/

Usually by DS guys who know nothing better than to slap an NN or LGBM on the data after some basic price transforms, and intuitively the idea of LSTM makes sense as data is related through time and are not i.i.d.
4/

Anyways, where I am going with this is that signal processing provides a beautiful solution to this issue of NN models getting destroyed usually because of the noise, lack of data (in effect, there is a lot of data but the lack of i.i.d. means you don't REALLY have much)
5/

dimensionality, and non-ergodicity in the data generally destroy these models which have highly non-linear decision surfaces which means they are much more susceptible to this noise. As an alpha architect we have to remove this noise as best we can so we can still
6/

pick up on these non-linear relationships without having a very confused model. Enter signal processing. EMD is a decomposition to break down a time series into multiple components that are non-stationary oscillating waves. This is great because we don't have to make the data
7/

stationary ourselves and break the structure (i.e. fract/log differencing & detrending), and makes our solutions more robust in the process. The key here is that we can now run LSTMs as per usual on these components and then sum up the predictions.
8/

No, I am not kidding, the RMSE drops by half if you do this. Even better the Hilbert-Huang transform can be used with this and there is a vast literature of similar decompositions that can be ensembled in the signal processing literature.
9/

Empirical Mode Decomposition (as empirical suggests it is based on the time series unlike dynamical mode decomposition), has many versions in the literature such as Ensemble EMD (EEMD), Complete EEMD with Adaptive Noise (CEEMDAN), the list goes on.
10/

The area diversifies and wavelets (much like how DL engineers create custom loss functions there is a lot of specialization to be done with certain wavelets) also are an effective decomposition. Wavenet provides an additional model other than LSTM to ensemble.
11/

A lot of quants forget that ensemble is one of the most powerful tools. On the surface level if you have two entirely uncorrelated strats that have a 20% EV p.a. and one says buy $500, and the other says sell $300 you can just buy $200. You now earn 40% EV p.a.
12.

or at least on this per trade basis, but have only created $200 of exposure (great for risk mgmt), but even better you have reduced notional by 1/4 and saved loads on fees. The best part is that this is only the most basic method of combining them and there exists
13/

a whole wealth of literature surrounding these. Using LSTM + Wavenet reduces overfitting much how judge panels attempt to remove human bias as well overall improving returns for no extra cost (free lunch yay!).
14/

Only caveats are that they need to be uncorrelated (probably won't be but hey anything helps) and trade at the same time (easy in this example, but an issue with different timeframe alphas). There is also the opportunity to stack the models on a timeframe
15/

based hierarchy as we can all tell that certain features like momentum (+autocorr) & quality signals are mostly long term, but more non-linear (-autocorr) signals are shorter term and thus it makes sense not to confuse your models and do these seperately.
16/

Hopefully, these tips on modeling better are helpful and as always I've added links below for research papers and articles to get started on this great topic:
17/

Sick ACE-EMD paper for crypto:
researchgate.net/publication/35…
Another great Hilbert Huang paper (crypto):
papers.ssrn.com/sol3/papers.cf…
Medium article I partially plagarized:
towardsdatascience.com/multiscale-fin…
Part 1 of an awesome playlist for signal processing
youtube.com/playlist?list=…

Part 2 of the playlist
youtube.com/playlist?list=…
A wavenet github example
github.com/sinusgamma/pro…
Really good implementation example of the LSTM model approach with CEEMDAN on github
github.com/Cy743652/CEEMD…
Wikipedia (such a good source for this stuff because a lot of it is non

finance related)
en.wikipedia.org/wiki/Signal_pr…

Oh and I forgot the best part (not just that it drops RMSE by half off the bat and likely 2/3 with ensemble + some extra effort), it can be used to forecast 3-month rolling volatility REALLY well!

Not just asset prices!