Statistical Arbitrage trading/market neutral strategies

[hirtop]

I was wondering if anyone is currently doing statistical arbitrage trading at the retail level, since this type of trading is widely used by the hedge funds firms.
I am interested to find out, share and discuss more about this subject as follows:
…data processing and manipulation (i.e. gaps), instruments, timeframes, backtesting methods (i.e. in-sample/out-of-sample), number of observations, cointegration testing methods (i.e. Dickey-Fuller, Johansen), results interpretations & trade signals, strategies (i.e. mean reversion), automated & semi-automated solutions (API’s and third party software) etc.

 

[Izzy Fuente]

"statistical arbitrage trading at the retail level"

This is a bit like bitcoin mining. The cost of providing the electricity to power the server farms to mine the coins has reached parity i.e. it's no longer worth he investment in hardware and power. The hedge funds and more specifically the large institutional arms have saturated proximity priority and effectively edged each other out of front running capability. If they've eroded their own edge there is precious little for us to pick up on the stat arb side.

 

[IFeelFree]

Cointegration

Over the past couple of years, I've developed a mean-reversion algorithm, similar to Bollinger Band, but applied to a cointegrated portfolio of leveraged ETFs. I started with the Kalman filter approach described in Ernie Chan's , "Algorithmic Trading", extending it to multiple (more than two) price series (using the Johansen procedure).

Cointegration is the important feature. Most financial price series are not mean-reverting, so that mean reversion approaches may not work that well. However, by finding a linear combination of two or more price series that is mean-reverting, one can apply a mean-reversion algorithm with much greater success. In-sample backtesting of my leveraged ETF triplet is around 50% Average Yearly Retrun (AYR) over the past 5 years, and somewhat lower (around 45%) out-of-sample. The maximum drawdown is 8.9%.

More recently, I've written a cointegrated Vector Auto-Regressive (VAR) algorithm for forecasting the cointegrated ETF portfolio price, and I've integrated that with the mean-reversion algorithm. This has boosted the in-sample AYR by another 20%, although I have yet to test the out-of-sample performance.

In any case, I've been using this algorithm for live trading over the past couple of months. My average holding time is about 8 trading days -- the half-life of mean-reversion of this portfolio. I've made a 7% profit during this 2-month period, which implies around 50% AYR, if it continues to work.

Anyway, you sound like you're learning about trading algorithms, so I wanted to share my (limited) experience.