StatArb

Knowledge Snippet
What is a factor?
I feel I never quite understood this idea until I started trading them very actively.
Factors are not anything special, they are just important alphas - NOTHING MORE.
You use factors to say:
Hey these alphas explain a large chunk of the variance and I don’t want to find them again. In crypto that might be a momentum factor so to avoid finding 20 versions of the same effect we use a xs regression to remove our momentum feature from returns and can then test against specific factor returns (returns minus the returns explained by factors basically).
The first

Effects:
1h reversal (crypto)
5-21 day momentum (crypto)
decay effect (crypto)
Small cap anti-premium (crypto)
1-7 day reversal (equities)
long term behaviour characteristics (ts argmax) (ie who had the largest volume spike in the last 3 years) (equities)
Momentum (9-18 months) ( equities)
A lot of the same effects, different timeframes

Crypto alphas tend to work in equities but not equities alpha in crypto I find.
** specifically price/volume alphas, alt datasets don’t compare **

Digging through WQ training material online for new transforms and found this gem from a scam network copy pasting alphas between them as WQ consultants

Ever since Claude code came out there has emerged a group of people into algo trading that didn’t have the IQ previously to produce a working trading algorithm but now post as if they’re RenTech. I heard one say he “spoke to a Quantum at Jump” and saw the secret sauce

MM knowledge - microprice and fair value:
I sometimes get asked on the best fair value metric to use. The options typically include:
- mark price
- mid price
- weighted mid price
- microprice
In general, for liquid instruments mid price is usually correct. Maybe there are questions about how to do it cross exchange (volume weighted, OI weighted, maybe some custom method), but it is the generally accepted standard.
In options you often find less liquid options are best to use the mark price or some weighted mid price because they’ll be heavily skewed in their book. You can play around with th

Rust is very good for use with LLMs. Checks a lot of errors naturally in the compiler and very token efficient as a language. An effective choice if you need to vibe code something which is performant.

When I am on a plane or train I open a notes file and just sit there thinking of features or ways to make more money.\n\nIt keeps my idea pipeline busy.

Believe it or not, the strongest features tend to be the most linearly behaved I find.\n\nWhich obviously also implies that the feature vs forward return curve being wiggly only in weak features is a nature of the lack of signal.\n\nYou can find certain non linearities in some specific types of features but generally they map linearly to forward returns

Grok find 1 million alphas above 3 sharpe.
Make no mistakes.