Synthetic hedges, cointegration, mean reversion and similar stuff | Page 11 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Simple Mean Reversion 621 replies

Pairs Trading: Reversion to the Mean 22 replies

Forward test of my new mean reversion strategy 11 replies

Question for synthetic and non-synthetic currency pairs 0 replies

Cointegration, Synthetic hedges, mean reversion in R, Tech Thread 0 replies

Attachments: Synthetic hedges, cointegration, mean reversion and similar stuff
Exit Attachments

Synthetic hedges, cointegration, mean reversion and similar stuff

  • #201
  • Dec 15, 2010 2:25pm Dec 15, 2010 2:25pm
tested trendomat on USDJPY m5 288 bars, taking the pairs and lot sizes from arbomat.
Both baskets closed in profit.

First, buy basket as it was going fast for the outermost limit.
Second, sell basket, as it was above the outtermost line in both trendomat and arbomat.


Best regards,
Attached Image(s) (click to enlarge)
Click to Enlarge Name: UJ.jpg Size: 92 KB
Click to Enlarge Name: arbomat_UJ.jpg Size: 121 KB
Click to Enlarge Name: trendomat_UJ.jpg Size: 109 KB
  • #202
  • Dec 15, 2010 2:41pm Dec 15, 2010 2:41pm
If the cointegration allows us to trade 1 or 2 trades max in profit ,ill be happy.

For trendomat i suggest we use some kind of maximum loss rather than waiting it to go back to the mean.

I am thatif a trade runs away from the ranges ie,arbomat..then that could mean something is wrong with hedge isnt it? Maybe the calculation is inaccurate as it was supposed to be in the same range forever...? If we can get a range that goes like 1- 2 months until we could recalculate..that would be a blessing indeed.
Never thought that playing ping pong would be so much fun.We all know that that the ping pong table had boundary in the squared circle...its not supposed comeout from the range....
Im currently testing a 4 pair with arbomat...which happen to go above the range.

All i wish is a stable range that is trade, i dunno if we can calculate the max drawdown out of any range.
Trendomat looks promising since we see a straight uptrend...time series.
buy..buy ..buy...couldnt be more easier.
We couldnt do triarb which is negated by the spread.
I wanna see other ways we can exploit the cointegration to trade single pairs..maybe we can see the direction its heading.Its like a MA which the price reverts to eventually .One other thing, i think that jagged blue line can show a major trend by its slope angle,not very sure unfortunately.

All of this will take a looot of time to test.......................
  • #203
  • Dec 15, 2010 2:58pm Dec 15, 2010 2:58pm
Quoting superbandam3
All of this will take a looot of time to test.......................

test until it breaks and gain knowledge in the process.
  • #204
  • Dec 15, 2010 3:05pm Dec 15, 2010 3:05pm
7Bit,

Have you considered a means of backtesting? I'm using JForex (Java) to backtest a basket EA and its historical data is accurate.

THis is the only way to really see the value of a strategy... test it against the times you get into DD. Then have a exit strategy to deal with this situation (doing your best to not fit a curve)

Jag
  • #205
  • Edited 9:43am Dec 16, 2010 9:16am | Edited 9:43am
thank you for sharing this brilliant method..

please excuse my ignorance, i read somewhere that the EAs write the amount of units to adjust in some text files (broker OANDA). could you tell me where you store those files again? i was looking back but couldn't find..

thanks again
  • #206
  • Dec 16, 2010 11:19am Dec 16, 2010 11:19am
finally found it, the files are stored in filesoanda-tickets

however i can't get the buttons to work, tried clicking on them but nothing happened (the files weren't generated) what am i missing?

another question, is the arb chart supposed to have horizontal lines? mine doesn't have. it looks exactly like the trend chart

thanks again
  • #207
  • Dec 16, 2010 2:24pm Dec 16, 2010 2:24pm
Quoting richy
finally found it, the files are stored in filesoanda-tickets

however i can't get the buttons to work, tried clicking on them but nothing happened (the files weren't generated) what am i missing?
doubleclick to select and then move a few pips. Unfortunately there are no user definable buttons in mt4 so moving text labels is the closest I can get.

Quote
another question, is the arb chart supposed to have horizontal lines? mine doesn't have. it looks exactly like the trend chart
thanks again

If use_diff=true then it will just plot a regression channel through the fitted data (should be more or less horizontal if the time window is long enough). the model does not enforce it to be horizontal in this mode, there are no such constraints, it will fit only the returns and allow the absolute price to drift where it wants, only some of them will be horizontal! Therefore I don't plot a horizontal channel (because this is not part of the model) but instead I plot a channel of what the basket actually does.

If use_diff=false then it is *enforced* to be a horizontal range, no matter what, this is defined by the model itself, price has no other choice, it is stationary by definition, therefore in this mode i can simply plot the horizontal channel into which the price was fitted and price will be inside it.
  • #208
  • Dec 16, 2010 2:32pm Dec 16, 2010 2:32pm
Quoting getch
1. You analyzed, what best Koef?
2. You calculated what symbol through linear regress?
AUDUSD = c1 * EURUSD + c2 * USDJPY +....?
You quoted me wrong. I said: I sold at 2*sd (two standard deviations, the second green line)

The coefficients are plotted into the chart at the left side, they are already converted to lot sizes so that they are directly usable for oanda units or MT4 lots. The regression is done as explained earlier in this thread. You can also look into the source code of arbomat.mq4 to see how it is done (you might also need to look up the documentation for lm() in the R manual)
  • #209
  • Dec 16, 2010 2:37pm Dec 16, 2010 2:37pm
Quoting sidhujag
Have you considered a means of backtesting?
Not yet, I would have to write a lot of additional code to do backtesting in MT4 with more than one currency at the same time and I am lazy.

At the moment I am still busy doing one profitable live forward trade after the other :-)
  • #210
  • Dec 16, 2010 3:14pm Dec 16, 2010 3:14pm
Quoting 7bit
Not yet, I would have to write a lot of additional code to do backtesting in MT4 with more than one currency at the same time and I am lazy.

At the moment I am still busy doing one profitable live forward trade after the other :-)
That's why I posted lol.. I'm doing JForex right now and you can test multiple currencies at the same time and the tick quality is 100% since you get bid/ask feeds seperately. Maybe I can help?


The api calls are similar and if you must use mt4 you can use a bridge to emulate the calls to their platform but that would defeat the purpose since you couldn't backtest.

Jag
  • #211
  • Dec 17, 2010 6:22am Dec 17, 2010 6:22am
hi 7bit,

something's really weird. i installed and set up the same thing in 2 computers. everything's fine on 1. but on the other computer, the charts showing only 1000 bars although i set the EA parameter to 2000...

any idea what went wrong? thanks
  • #212
  • Dec 18, 2010 4:55am Dec 18, 2010 4:55am
Quoting 7bit
You quoted me wrong. I said: I sold at 2*sd (two standard deviations, the second green line)
Why two standart deviations? Why not another (Koef * sd)?

Quoting 7bit
You can also look into the source code of arbomat.mq4 to see how it is done (you might also need to look up the documentation for lm() in the R manual)
Your regressand is current Symbol. It is bad. Example:

1. Use linear regression (current Symbol is AUDUSD):
coef1[AUDUSD] * AUDUSD = -coef1[EURUSD] * EURUSD -coef1[GBPUSD] * GBPUSD - ...., where coef[AUDUSD] == 1.

2. Use linear regression (current Symbol is EURUSD):
coef2[EURUSD] * EURUSD = -coef2[AUDUSD] * AUDUSD -coef2[GBPUSD] * GBPUSD - ...., where coef2[EURUSD] == 1.

Then vectors coef1 and coef2 aren't proportional. It is bad.

Regress use is an incorrect way since regress isn't equal in rights to the members.
  • #213
  • Dec 18, 2010 1:53pm Dec 18, 2010 1:53pm
Quoting 7bit
I am still hoping that someone else (who is more experienced in R and in statistics than me) will take the existing code and replace the simple regression with something else or otherwise enhance the script and publish the modified version here. I am doing some experiments myself but have nothing ready for publishing yet.
Quoting getch
Regress use is an incorrect way since regress isn't equal in rights to the members.
getch, could you replace the simple regression of your way of calculating the coefficients and to publish the modified version Arbomat here?
  • #214
  • Dec 18, 2010 3:06pm Dec 18, 2010 3:06pm
Quoting richy
hi 7bit,

something's really weird. i installed and set up the same thing in 2 computers. everything's fine on 1. but on the other computer, the charts showing only 1000 bars although i set the EA parameter to 2000...
You don't have enough history. open a chart of each symbol and scroll to the oldest bar (Pos1 key on the keyboard) (this triggers download of older data) and do this until you reach the end. Do this with all symbols and all timeframes you need. The EA will print into the log which pair it was that had the least amount of bars.
  • #215
  • Dec 18, 2010 3:28pm Dec 18, 2010 3:28pm
Quoting getch
Why two standart deviations? Why not another (Koef * sd)?

There is no special reason. Its an arbitrary decision. sd and 2*sd are just very commonly used. I sold near the level where I believed was the reversal and i was lucky because it did indeed what I wanted it to do.

Your regressand is current Symbol. It is bad. [...] Regress use is an incorrect way since regress isn't equal in rights to the members.
I'm not sure whether I understand you corectly and what exactly you mean with "not equal in rights to the members"

There is no "correct" way, there are many ways and I chose one of them. I cannot simply do

0 = b1*pair1 + b2*pair2 + ...

since this would result in all coefficients being estimated to exactly zero. I need some kind of constraint that enforces it to be non zero. The easiest way I have found is to simply force one of them to always be 1 (regress one against all others). I try each one and select the "nicest" result for trading.

If you have a better Idea then please tell me.

If I have 7 pairs I can do 7 different regressions of this type: each pair against the 6 others. I usually try all of them but most of the time I end up using GBPUSD against the others. Also interesting on longer timeframes seems USDCAD vs. the others but I haven't yet traded this.
  • #216
  • Dec 18, 2010 5:16pm Dec 18, 2010 5:16pm
Quoting 7bit
I cannot simply do

0 = b1*pair1 + b2*pair2 + ...

since this would result in all coefficients being estimated to exactly zero.
The second condition is |b1| + |b2| + ... = 1.
  • #217
  • Dec 18, 2010 7:05pm Dec 18, 2010 7:05pm
Quoting getch
Why two standart deviations? Why not another (Koef * sd)?



Your regressand is current Symbol. It is bad. Example:

1. Use linear regression (current Symbol is AUDUSD):
coef1[AUDUSD] * AUDUSD = -coef1[EURUSD] * EURUSD -coef1[GBPUSD] * GBPUSD - ...., where coef[AUDUSD] == 1.

2. Use linear regression (current Symbol is EURUSD):
coef2[EURUSD] * EURUSD = -coef2[AUDUSD] * AUDUSD -coef2[GBPUSD] * GBPUSD - ...., where coef2[EURUSD] == 1.

Then vectors coef1 and coef2 aren't proportional. It is bad.

Regress...
Yes, it seems that there is some asymmetry, but I've seen this asymmetry in quite a bit of published econometric literature.

Suppose you took the second equation and multiplied through by -coef1[AUDUSD]/coef2[AUDUSD].

If everything is done properly, the result should be the same as the first, right? To me, that seems to be the way to test it is working.

The asymmetry also bothered me as well. The following paper showed an approach based on PCA which eliminates the asymmetry of the 'reference'.

http://www.scribd.com/doc/23407376/R...Multi-Variance
  • #218
  • Dec 19, 2010 1:34am Dec 19, 2010 1:34am
Quoting mbkennel
Suppose you took the second equation and multiplied through by -coef1[AUDUSD]/coef2[AUDUSD].

If everything is done properly, the result should be the same as the first, right?
No. Look here.
Quoting mbkennel
The following paper showed an approach based on PCA which eliminates the asymmetry of the 'reference'.

http://www.scribd.com/doc/23407376/R...Multi-Variance
I can't see this. Please upload it to another place.
  • #219
  • Dec 19, 2010 5:31pm Dec 19, 2010 5:31pm
Can you read this?

http://papers.ssrn.com/sol3/papers.c...act_id=1404905

Sorry, I didn't understand eontents of the link you gave regarding the two different regressions, and why they are not equivalent. Can you explain more? Thanks.
  • #220
  • Dec 19, 2010 5:40pm Dec 19, 2010 5:40pm
Quoting mbkennel
Can you read this?

http://papers.ssrn.com/sol3/papers.c...act_id=1404905

Sorry, I didn't understand eontents of the link you gave regarding the two different regressions, and why they are not equivalent. Can you explain more? Thanks.
From what I understand from this paper (90% of it is over my head) he is using PCA, isn't this a completely different concept than just fitting regressions with OLS?
0 traders viewing now