Discussion in 'Strategy Building' started by MeanerRevert, Sep 18, 2003.
has anybody had any material success implementing cointegration
in pairs selection?
I am working hard on a tool.
we tried it for several months a year ago and failed to bring test sharpe ratios above 1.25. we still consider going back to start all over, but haven't done so yet.
i must add that our initial approach was not so much pairstrading but trading the forecasts derived from the cointegration process itself. thus we could end up trading baskets of stocks against others. most of the time we would have traded pairs, but it could have been triples or quadruples as well.
i think that the concept of cointegration as such is as overvalued as was AI in the beginning of the nineties. as such cointegration is not tradeable. it might be one of many possible starting points, but it is not an answer per se. no big surprise, given the amount of publicity the concept gained.
Could somebody post a definition or a short description of what is meant by "cointegration".
Thank you kindly,
cointegration is a loose kind of correlation and goes back to a johanson. assume you have two time series and want to find if they are cointegrated = have something in common. very simplified: you try to find a third time series that is derived from the two basic series by means of multiplying each with a factor. if this new time series fluctuates around the zeroline, you say the two basic time series are cointegrated. now the task is to find such a factor set for a given set of time series. the task is always to find a fluctuation around zero, which is called a stationary process.
Man, I have a technical question â how did you calculate Sharpe Ratio if you traded the strategy just a few months? Avg(MonthProfit)/StandardDeviation(MonthProfit) ?
If so, i think in most cases it will be less then funds that calculate it by year.
we never traded cointegration - i referred to backtesting. we calculate sr by dividing (annualised returns - risk free rate) by annualised vola. over ten years time it will not make any difference if you calculate annualised daily data or annualised monthly data. then you have enough data points that the two converge.
the key is the annualisation. you can calculate a sr after two trading days - it won't mean anything, but in principle the figure is totally comparable to any other sr.
we have been trading co-integration on currency-pairs. the perfect co-integration example is the 3-pair-cross of eur/usd, usd/jpy, jpy/eur. they always sum to one (multiplication), otherwise arbitrage would be possible. therefore all historic data and future predictions show co-integration properties.
we tried 4-hour-prediction of currency movements and achieved a significant 55% prediction rate. getting started into trading the hit-ratio (and positive pay-off) we got hit in 2001 by a succession of currency interventions by the japanese central bank. losing heavily through the interventions and discovering that only 4-ticks were the expected profit per trade (before slippage and costs) we quit the co-integration trading of currencies...
Clive W J Granger got a Nobel prize for this few days ago. I think it's a great material success.
Separate names with a comma.