B-TIER skill
Quant alpha discovery primer — Kalman filters, stat arb, XGBoost
100k ft — What is this?
Quant alpha discovery primer — Kalman filters, stat arb, XGBoost
50k ft — Why care?
@sorhan.hq profile of Guangting Yu (applied math PhD, published NeurIPS). Lists the alpha-hunting toolkit: statistical arbitrage, Fama-French factors, Kalman filters, mean reversion, momentum, XGBoost/NNs. Plus covariance formula. Good intro-level quant reference.
10k ft — Key takeaways
(no takeaways yet)
5k ft — Tools & techniques
(no tools tagged)
Fama-French factors, Kalman filter, Kalman filters, arbitrage, mean reversion, statistical arbitrage
1k ft — Deep-dive / try it
Meet Guangting Yu ( @yugtmacs ):
- Applied Math PhD. Quant
- Finance MS (UMich).
- Worked on alpha discovery using Kalman filters. Published - AI/ML research (NeurIPS).
Quants don’t guess markets. They hunt alpha. Hidden edges in data.
Statistical arbitrage exploits tiny pricing gaps
Factor models like Fama French explain what drives returns.
- Kalman filters separate signal from noise in real time
- Mean reversion bets prices snap back
- Momentum rides trends early
- Machine learning models like XGBoost and neural nets predict moves before they’re obvious.
Covariance Formula:
Cov(X, Y) = E[XY] − E[X]E[Y]
Measures how two variables move together
Positive means they rise together, negative means they move opposite
Core signal used to detect relationships and extract alpha
While mos…
Transcript excerpt
Alright, so you were a quant at hedge funds helping them predict daily stock returns using machine learning. Yes, so it's called finding alphas. Finding alphas, and how accurate did it get? We can reach around 5% to 10%. How do you calculate that? We use historical stock data to get a prediction on stock returns on the next day, and then compare it with actual next-day returns. This algorithm to use historical to predict the next days is called finding alphas. Our relation is calculated by multiplying the predicted returns and actual returns, and summing up for all stocks. To get that 5% to 10%, and if you constantly can predict that, they pay you $100 to $200,000 for the algorithm. Yes.