Shan, you are right to pursue this strategy. Keep up the investigative work. Baskets will likely model smoother than simple pairs, but the tradeoff is increased execution slippage and complexity. There is money in it, because it is going to be much more consistent than taking directional flat price risk. And you can lever the piss out of consistency in a hurry.
Agreed, simple is always better - just be careful that simple doesn't mean taking on risks you're not aware of, which is often the flip side of 'simple'. IMO, your interests are worth spending a bit of research energy on, gl.
Tour, We have profit targets and stop-loss levels we set at the time of trade entry. If the model flips back over before the stop-loss is reached, that mechanism will get us flat as well - it is an and/or proposition. My apologies, but propriety would keep me from further discussion on specifics.
I would think a moving average crossover method may be a simple solution here. You could probably test this pretty easily. I don't know if you have been to the ACD thread, but i've always though that bracketing the intra day opening range and getting long or short on an A-up or down would be a creative solution! Also I don't know if you are familiar with the Hurst exponent, but you can test for the predictability of a time series and estimate if it has a random, anti-persistent, or persistent behavior. In the divergence approach, theoretically is persistent spread would lead to some opportunity. I have not used the Hurst exponent in this manner, maybe someone can help here. Lots of things one can do here!
Yep, the potential number of combinations are exponential - you are literally building your own markets. You can tune the volatility and market behavior characteristics of the spread differential by adding legs or by filtering and selecting each component leg according to its' statistical correlation. Many stat arb traders want a fundamental link between the leg components in terms of the business market sector and the company's cap, while other stat arb traders just want to meet a certain positive correlation threshold. If you have a creative bone in your body, you find this niche specialty trading technique to be very stimulating.
Yes, this is a good summary. My desk does intraday equity pair stat arb; the actual code was written in python by a quant. My role is a combination of research/trading/risk management/operations. I am constantly making adjustments to algorithms, what pairs we trade, stop-loss, etc., based on market conditions and various news. So I guess you can say that it's a "gray-box" system. We trade stocks and etf's from the same sector or country (eg., FXI-SNP, EWZ-PBR) that exhibit fairly high correlation. I'm thinking of running some co-integration tests but not entirely sure if that will be helpful for intraday models.
The baskets and the spread combinations featuring multiple legs will exhibit better mean reversion tendencies than the simple pairs from my experience to date - unless you are dealing with a super high correlator match ( closely aligned or akin to an arbitrage ).
This seems right. One disadvantage of intraday pairs is that the upside is limited while the potential downside is unlimited. If news comes out, and there is a breakout in a given stock while its corresponding leg remains stationary, we get crushed on that pair, since the odds of a mean reversion occurring that day is rather low.