Just to ram home my assertion about how bent Metatrader is, look here: http://www.forexfactory.com/showthread.php?t=70582
Its quite appalling and one reason why there needs to be much tighter control of fx trading.
Pair Trading Strategy Journal
- Thread starter jonnysharp
- Start date
Just to ram home my assertion about how bent Metatrader is, look here: http://www.forexfactory.com/showthread.php?t=70582
Its quite appalling and one reason why there needs to be much tighter control of fx trading.
I started my paper trading on Nov 17 and by Dec 17 I was about 4.6% when margined 100% with RegT.
Out of more then 250 completed trades I lost only 12% of them with average win of around 4.5% and average loss of around 9%.
Another thing I noticed is that entering correlation has no effect on profitability at all - you can enter a trade with current correlation of 10% as well as 90% - statistically, it doesn't have any dependency on the profitability of the trade, only Z what really matters as well as an average 100 day correlation.
Also, I closed all trades which take more then 15 trading days to converge.
Also, I haven't traded any stocks with value less then $10.
Yes, virtually (on paper) though.
I have a universe of over 800 backtested pairs with the only criterias of a) pair being correlated with avg R>0.7 over last 100 days; b) earning more then 2.5% per trade (on one side); c) having more then 8 trades a year.
My current goal is to execute as many trades as possible for 2-3 months in order to get valuable statistics and then I will reduce the universe to 250-300 pairs with better profitability targets and then will just use my daily "signals" to enter and exit positions on those trades.
Let's see how it will work. I expect the final win/loss ratio to be around "classic" 75% with approx. 1.0/2.0 win/loss distribution (i.e. my usual avg. loss is approx. twice as more then avg. win mainly due to the fact that I do not wait for the trades to converge and close "bad" trades after 3 weeks (15 days). I've noticed that if the trade is profitable - it will converge within first 6-7 days on average, around 95% of profitable trades converge within 3 weeks time).
I have a universe of over 800 backtested pairs with the only criterias of a) pair being correlated with avg R>0.7 over last 100 days; b) earning more then 2.5% per trade (on one side); c) having more then 8 trades a year.
My current goal is to execute as many trades as possible for 2-3 months in order to get valuable statistics and then I will reduce the universe to 250-300 pairs with better profitability targets and then will just use my daily "signals" to enter and exit positions on those trades.
Let's see how it will work. I expect the final win/loss ratio to be around "classic" 75% with approx. 1.0/2.0 win/loss distribution (i.e. my usual avg. loss is approx. twice as more then avg. win mainly due to the fact that I do not wait for the trades to converge and close "bad" trades after 3 weeks (15 days). I've noticed that if the trade is profitable - it will converge within first 6-7 days on average, around 95% of profitable trades converge within 3 weeks time).
Ive been looking through the thread for a few weeks and have been testing PTF. It seems like an interesting strategy, does anyone mind sharing the stats for last year?
Id be interested in
max risk per trade
biggest winner
biggest loser
If possible do you mind sharing the pair/date/entry direction of biggest winner/loser? I am trying to back test using this strategy with options. Using PTF the winners look good, I'm just trying to find bad trades.
Id be interested in
max risk per trade
biggest winner
biggest loser
If possible do you mind sharing the pair/date/entry direction of biggest winner/loser? I am trying to back test using this strategy with options. Using PTF the winners look good, I'm just trying to find bad trades.
I analyzed a trading record of 93 pair trades from October/November of 2010.
I calculated the Profit/Loss for these trades as if they had been sized each of the following four ways:
Method 1. Equal Number of Shares
Method 2. Equal Dollars (Face Value)
Method 3. Inverse to Beta
Method 4. Equal Long-Term Risk (Volatility, 10 year/max std dev)
Code:
ALL TRADES
Method 1 Method 2 Method 3 Method 4
Avg P/L 4.93 6.15 7.16 17.08
Std P/L 286.12 295.97 277.35 293.46
Sample Size 91 91 91 91
T-statistic 0 0.03 0.05 0.28
0 0.02 0.25
0 0.23
0
P-statistic 100.00% 98.00% 96.00% 78.00%
100% = 100.00% 98.00% 80.00%
insignificant 100.00% 82.00%
100.00%
Code:
WINNING TRADES
Method 1 Method 2 Method 3 Method 4
Avg P/L 159.69 158.1 158.04 185.23
Std P/L 110.78 104.17 112.26 124.9
Sample Size 62 65 61 59
T-statistic 0 -0.08 -0.08 1.19
0 0 1.31
0 1.25
0
P-statistic 100.00% 93.00% 94.00% 24.00%
100% = 100.00% 100.00% 19.00%
insignificant 100.00% 21.00%
100.00%
Code:
LOSING TRADES
Method 1 Method 2 Method 3 Method 4
Avg P/L -325.91 -373.71 -299.63 -292.95
Std P/L 216.2 228.83 207.75 213.31
Sample Size 29 26 30 32
T-statistic 0 -0.79 0.48 0.6
0 1.26 1.38
0 0.12
0
P-statistic 100.00% 43.00% 64.00% 55.00%
100% = 100.00% 21.00% 17.00%
insignificant 100.00% 90.00%
100.00%Method 4 (Risk) would have produced the highest average profits, highest average winning trade, and least average loss. The 10-year volatility was computed from after the fact using in-sample data so it is not the exact number available at the beginning of the trade. For stocks with less than 10 years, the maximal available timeframe was used. Since this underestimates volatility and since the early history is probably not representative of the equilibrium, greater care should be used in real trading.
Method 3 (Beta) and Method 2 (Dollar Neutral) did equally well on winning trades. Method 3 (Beta) did quite a bit better on losing trades, but not statistically significantly.
Method 1 (Equal #) was probably treated unfairly. The entry prices differed by as much as 11 to 1. Someone really trading by equal number of shares would presumably have forgone of these trades, but the sample size is too small to fairly compare it with these trades removed either.
I would love to have more trades to add to this study, and I would love to have a working C++/C# Johansen cointegration module to compare against these results. (Having trouble getting even the core components of GRETL to compile under Visual Studio.)
I've put together a simple Java program which reads all the stocks from Yahoo Finance, gets the 500 day historical prices for all the stocks for which it's available and then backtests each pair possible per industry (updating historical prices with latest available).
As long as it literally takes 10 mins to backtest around 60k pairs for correlation & profitability (and to select lucky 800 out of them), I update my pair trading universe with fresh pairs every week.
I tried to do some backtests on the same conditions with Pair Trade Finder and to some reason it's times lower then my Java program.
Currently I am putting together another program on top which will work with IB API and measure Z in real time and then send enter/exit signals to the market when needed.
It took me a week to put everything together, but I believe that I am a way ahead of Pair Trade Finder now, although not user-friendly as PTF
Here are the services that I have seen listed in this thread:
Free:
http://www.catalystcorner.com/index.php?m=pair_tool
http://market-topology.com/
http://www.impactopia.com/
http://www.sectorspdr.com/correlation/
And my own http://toolkit.entonesoftware.com/Correlation/CorrOvODisplay.aspx
Not free:
http://www.pairtradefinder.com/
http://www.spreadprofessor.com/ (?)
http://www.csidata.com/ (In Windows App, no longer work there)
I don't know that any of these do cointegration, proper. I have been working on adding it to my tool.