MFT Series: crafting features

Recap

I’m going to do a recap of what I’m looking at, and then talk a bit about the features I built last week.

I’m using Bybit mid data (pretty awesome that you can download the orderbook history for all assets, btw).

Forward returns are 24h forward returns (with a 30s delay after the feature bar ends, to avoid contamination) divided by an EWMA of the 7D std of hourly returns.

the features

I run the analysis on 3 sub-universes ranked by 30D ADV and reranked monthly. Some features work best on liquid names, some on illiquid names, and some work everywhere.

For now, I've focused only on time-series features, not cross-sectional ones. I want to keep the scope simple, and the implementation will be more straightforward with only TS features.

Here is an overview of the 14 features I have.

Feature bin plot

And their correlations:

feature correlation

I don't really know what it's worth, but I built these features from intuition, didn't apply any transforms, and most of them make sense to me.

Except for a few features, the correlations seem fine to me.

Managing turnover & execution costs

The next step is managing turnover. I still don't really know what my execution costs are going to be. Feature turnover is far from homogeneous: by my estimates, hourly turnover ranges from 3% to 40% of gross depending on the feature. A more predictive feature can probably handle more turnover, and that's fine, but for simplicity I'd rather set a hard cap of, let's say, 5% hourly turnover per feature. Alternatively, I could use a more complex metric that weighs a feature's expected returns against the costs it generates, but I want to take the simple road and go live ASAP.

I think I'm just going to smooth the features with an EWMA so they all end up at around 5% turnover, discard the ones that become trash after smoothing, and then move on to the fitting.

The fitting

I think I'm going to fit a rolling 90D ridge regression per ADV universe and see what it produces and how stable it is...

Next steps

• Smooth the features to a decent turnover and re-evaluate whether they're still worth using

• Find a simple way to fit the features to forward-return predictions

• Go from predictions to a portfolio

• Design the execution system (maker, taker, or both; maybe HFT prediction...)

As always, I love reading advice and feedback on what I'm doing.