Information For Beginners

So I’m trying to diversify and i came across this thread and i currently trade EUR/USD and my partner trades EUR/JPY. so if I’m understanding this thread correctly then I’m pretty much S.O.L. But we have managed to do alright except for the pretty loss we took two weeks ago. I have been experimenting with the USD/JPY cross and it seems to do well. But can’t you essentially do the opposite between the EUR/USD and EUR/JPY and come out a winner?

Some historical data, 20-day correlations:
EURUSD-GBPUSD
http://understandmarket.com/images/eurusd-gbpusd20.gif
EURUSD-GOLD
http://understandmarket.com/images/eurusd-gold20.gif
USDCAD-GOLD
http://understandmarket.com/images/usdcad-gold20.gif

First of all, welcome to the forum!

The studies I've done are medium to long term and they should factor out a lot of noise you get with daily correlations.

I see a high correlation with the two pairs in question so I would be cautious about trading them together. That's not to say you won't get divergence as the last poster shows but you have a constant correlation problem to manage if you do.

I'm not entirely sure this belongs here, but maybe someone wouldn't mind taking a stab at this.. I'm still a pretty big fx newb around here so please bear with me.

First, is this a basically correct understanding of correlation as it relates to trading?
--------------
Correlation is simply two pairs moving in tandem to a varying degree, either directly or indirectly. The goal of focusing on correlation is to avoid overexposure to very similar elements of risk.
--------------

So now for my real question, can I take signals from correlated pairs and still be safe by only trading one pair at a time? For example, I currently am monitoring and taking trades on GBPUSD, EURUSD and USDJPY. I (hope to) avoid the overexposure problem by only taking trades on one pair at a time.. if I get signals on more than one in the same direction, I might take the trade on GBPUSD since it is generally the most volatile. At the same time if there are conflicting signals, I might decide not to trade at all. Can someone confirm for me that my thinking is correct on this?

thanks for any help,

nate

Yes you've got a pretty good understanding there, nate.

That was one approach I did consider as well. Say you trade the USD "top six" - 4 majors + AUD & CAD, can you avoid correlation by trading them one at a time only?

Well again, yes and no! Yes because you avoid the problem directly by trading just the one pair. (If you just stick to one pair at all costs then you avoid the problem for good.)

No because if you switch from pair to pair for each trade then you will run into correlation indirectly. Here's an example: say you get a buy signal on EURUSD, and since this pair has a high correlation with other majors they signal a buy. The decision is easy, just take a long position and trade it according to your system.

Next you exit the trade and on the same pair you get a sell signal but the others are still signalling buy. What do you do now? Stick with the sell and ignore the buys? Sell the euro and buy the others (and risk cancelling out your overall position)?

If you get conflicting signals then thats where a truth table helps, it picks what pairs you can trade safely together based on historical correlation analysis. So for me I would just trade one USD major to the exclusion of all other pairs (i.e. don't bother with them at all).

Also if you have two highly correlated pairs that signal buy or sell in tandem then you can halve your position on both trades and go ahead without risking double exposure on the overall position. (So where two pairs are false in the truth table you can override by this approach).

Hope this helps.

Iso, thanks for this food for thought, glad to hear I'm atleast on the right track. I will ponder on what you said, I'm going to give those truth tables a closer look as well. Thanks again,

nate

The following article was written by Fijitrader. It is copyrighted, and is reproduced here with permission from the author. It is just one of many articles and posts of similar depth and quality to be found in Fijitrader's Money Management and Psychology section of the James16 Group of Forex Factory.


Asymetrical Leverage

I hear the above term bandied about on the forum but few seem to understand it. It is probably easier to understand than it is to explain but the most common mistake of all when it comes to position sizing is not having the ability to execute the position size that will provide optimum trading results.

For example if you decide to risk 2 percent per trade and you have a 5k account and your stops are 35 pips on every trade you are going to risk 100.00 per trade. 100/35 is about 2.85 contracts. Therefore if you risk 2.85 mini contracts, you have a 28,500 position size with 35 pip stops. If the trade is a loser your maximum loss should be 99.75. Your risk would have been slightly less than 2 percent or 100.00

The only problem here is that with a mini account you cannot trade 2.85 contracts which would give you a position size of 28,500. You must round down so as not to exceed your maximum of 2 percent per trade. Now you only have 2 contracts working for you which is a 20,000 position size. That is only 1.4 percent risk. If your trade is a winner your account equity only increases by about 70 percent of what it could have were you to have put on the 2.85 contracts your risk profile dictates. You left 30 percent of your profits from that trade on the table.

Now suppose you have a winning trade with the 2 contracts instead of the 2.85 contracts -- it would take a win of 150 pips to have an account equity of 5300.00. However if you had put on the position size of 28,500 you would only need a 106 pip winner to increase to three contracts. The problem with a mini is that the account is denominated in 10,000 position sizes which makes it impossible to place the correct position size.

There are numerous iterations I could cover to illustrate this but you must postulate your own scenarios in order to see all the sides of this. Most people are just too lazy to really work this through and see all the facets of it. I'll cover a bit more so that you can get the basics but to make compounding work for you there is no substitute for learning it untill you can describe the phenomena from memory. It is not sufficient for you to read this once and leave it be.

Moving on. Let's look at what happens now that you have 5300.00 in your account. Your risk is 2 percent and your stops are still 35 pips. 2 percent of 5300.00 is 106.00.

106/35 ($1.00 per pip mini on /usd contracts) is equal to 3.085 contracts or a postion size of 30,850.00. However we round down to three mini contracts or 30,000.00. Now suppose this trade is stopped out and we lose our 106.00 which means we are back to trading two contracts again.

Stop here and think a minute... we had to trade two contracts since the account was at 3500.00. At first we were risking almost exactly 2 percent of our equity. 3500 * .02 = 70.00. Two contracts (or a 20,000 position size) risking 35 pips is exactly two percent of our account working for us on each trade at that account level. However as we chalk up winners we cannot increase our position size until we earn another 1800.00. Each time our equity grows our ability to increase our risk (and our position size) does not -- until we have pocketed that 1800.00 in our winnings.

Our leverage decreases as our account increases up until the point where we can increase our position size by another 1800.00 in the example shown. At this point our leverage increases again to 2 percent but again as our account increases our leverage decreases. Many people have been able to grasp this point of asymetrical leverage and explain it but what is missing in most cases is the impact on our trading system and our account growth. I'll try to explain it in some more detail below for those who do not yet see the full implications.

That means that some of our trades are going to increase our account equity by a different percentage than others even if we get the same number of winning pips. The same is also true for losing trades. Some losers will cost us 2 percent of our account equity and some as low as 1.4 percent. This can dramatically skew the results of our trading.

Let's look at trading systems. Because they are designed to capture particular kinds of inefficiencies in the markets the result in most trading systems are streaks of primarily winning trades and streaks of primarily losing trades. That is not always so but it is a common attribute of most systems.

So what happens if your losing streaks tend to appear just as you get over the next position size increment. Lets look at the most common problem in trading mini accounts:

3000.00 account.
35 pip stops.
10,000 minimum trade size.

You begin trading with one contract because there is no position size smaller than that. Your risk is only 1.16 percent and you cannot trade two contracts because that would put you over 2.32 percent. One common and suicidal mistake most people make here due to greed is to round up and begin with two contracts. If the first trade is a loser it will decrease your account size by 70.00 and therefore increase your leverage to almost 2.4 percent. Having already broken a rule, the trader resorts to trading another 20,000 position size, at the higher leverage and increased risk because the trader realizes that if he/she trades a single contract (10k positions) -- that instead of taking a nice 40 or 50 pip profit on the next trade he/she will have to make at least 70 pips just to break from the last loss. This dilemma is the single most frequent opportunity where traders screw up due to asymetrical levearge. Greed has them rounding in the wrong direction because they don't want to work doubly hard or be doubly patient in order to compensate for a serious deficiency in position sizing flexibility.

Getting back to trading systems -- remember that streaks of winners and losers are not uncommon. What happens if your winning streaks constantly occur during the lowest leverage regions and your losing streaks constantly occur during your highest leverage ratios? When this happens you'll find many glass ceilings in your trading. This not only amounts to a major inefficiency in your equity curve it also destabilizes your trading psychology whether you are aware of it or not.

If you are aware of this problem then you'll be kicking yourself (either consciously or unconsciously) for your lack of professionalism due to being lazy about getting the right type of account for your account size. If you are not aware of the problem you'll be struggling more than you should and intuitively you'll sense that something is not right but you'll never quite put your finger on it. This is also destabiling because it undermines your confidence and to the degree that you are ignorant of the issue you'll also feel more at the mercy of the market than you really are. The sad part is that you'll not even know why.
FT

i think this a recent addition by oanda, but at least i can check my method with better accuracy.

here is the oanda results for euro & gbp taken over the weekend:

so i went back two years to check, mine are within a few %ge points as well, so that means:

a. my method is pretty much correct after checking against two independent data sources
b. strategy builder fx data is pretty good and
c. i don't really have to do this checking anymore.

(oanda values are in bold underneath - i used the two year values to check against)

going back 27 years. note there is no eurusd since its data only goes far back as 1989 on sbfx:

The yellow bold cells are my new selection, the other hues of orange are other interesting combinations. I've left out a yen pair for now since I can't find one to my satisfaction.

In hindsight, taking 2000 as a starting point didn't give a large enough dataset for analysis. These results should be more reliable in comparision and for picking pairs.

from the same date: sep 04 - who is best..?

here is the oanda data by way of comparison - enjoy folks!

it occurred to me that the same problem that confronts "manual-ers" will also happen to mechanical traders. namely managing correlation risk.

so howabout testing these correlation tables on systems that you've developed in the past? correlation problems will affect programs and humans alike, so what works for one should work for the other.

what you would have to do is construct some kind of lookup table based on the tables above, so for example let's pick the following 4 pairs from the last table:

remember that T is the correlation threshold in decimal, so 50% = 0.5. so values above or equal to 50 are treated as correlated and hence false, conversely values below 50 are treated as uncorrelated and true.

table rules:

1. pairs that are 'T' can be traded together if signals arise simultaneously, so for example you could trade +/- EURCHF with +/- EURGBP where + = buy and - = sell. no problem so far (i hope).

2. pairs that are 'F' must follow some kind of correlation policy you construct. one such policy could be:

a. if you get the same direction (++ or --) then ignore both signals or pick just one.of the two to trade, but not both (see XOR rule below).

e.g. +EURUSD and +USDCHF since they would normally cancel each other out.

b. if you get +/- then take both trades but halve trade size for each, or just pick either one.

e.g. +EURUSD and -USDCHF.taken together at 1/2 size each.or trade +EURUSD XOR -USDCHF (XOR = exclusive or, one or the other but not both!)

Test over a number of systems you have. Just change the code that before you take a trade, check against the lookup table and implement your policy like above.

What I would like to see is the following:

1. Test for different values of T from 0.1 to 1.0 in 0.1 intervals. The correlation buffer zone is generally 0.4 to 0.6, or 50/50% with 10% error margin. So I would expect the values in this zone to work best. Below 0.4 is too conservative and above 0.6 is too aggressive.

2. What time period (t) is best to build such a table? You need to construct a table from correlation data so that the last entry is the last day before your first test data entry, then count back the required time interval.

So for example say you use test data from 2000-01-01 and split it into various datasets for testing. Your correlation data should end on 1999-12-31 and if you had 5 years lookup data then that would wind back to 1995-01-01. Ten years to 1990-01-01. What values of 't' work best to construct a lookup table? 2 years? 5 years? 10 years? This is a problem I'm wrestling with right now. My guess is 5 years minimum and perhaps 10 is optimal.

With automation you can quickly test a range of values for 'T' and 't'. Whatever values are best for a computer should also be good for a human!

If anybody wants to do this then I can provide them with all the data I have and start a new thread for test results. Read this thread again for more information.

...Quite a long post. I hope somebody out there understands then benefits of what I;m saying here...OVER TO YOU!!!

I like this thread, and after seeing some results I have another idea, I would like to test. When you look in shorter time-frames (1 week or 1 month), some of the pairs change their behaviour. That means their correlation with other pairs are different from the normal values. And now I am wondering whether these changes predict a reversal, which means they change somehow back to the normal situation.

Suppose the EURJPY is normally around +50%. One week or month it is suddenly -10% (see your message at 4-9-06 of oanda values). Would this mean that it will reverse in the next weeks, which means it will be go back to +50% or even higher?

Hi Iso,

i think i know what you try to achieve and i do agree, that this topic is really important... traders must be aware about how correlation can impact their results. it all falls under the money management/risk management/position sizing cathegory.
on the other side i think, you are aproaching this issue from the wrong side.
your method is kind of brute force method, where you just try different combinations and search for the optimum using backtests and comparing the results. besides that, your method would only concentrate on correlation between two currency pairs, but there is a lot more to correlation than that.
this is only correlation between two instruments, but there are a lot more correlations of differnet kind, that have impact on the results. How about correlation between three or four or five or more instruments at the same time? what about correlation between different trading strategies traded at the same time, or one parametrised strategy traded with different sets or parameters or on different timeframes, etc.

the bottom line is, there are a lot of correlation included and you need to find some way to diversify your portfolio (for example portfolio of currency pairs or trading strategies). instead to going the brute force way, i would suggest you take a look at the Modern Portfolio Theory.

there are statistical procedures for arriving at optimal portfolio composition (efficient frontier).
you either use past statistics of your trades, or generate the statistics using a backtest for each asset in your portfolio.
the time interval you choose for the calculations is dependent upon your particular conditions, but you want to have a statistically significant sample of trades.
for the rest you better use some tool, because you need to calculate thing like menan and variation of returns for each asset on several time intervals, also the covariance matrix created from covariance between each asset pair combination, then you need to solve some eqations... you get the idea i'm sure you find the exact procedure if you google for it, also maybe some excel sheet to do the calculations.

anyway... if you like this idea, we can discuss it further

cheers

RoBiK

it all depends on your frame of reference. for example what would be the "normal values"?

you mention eurjpy. which pair are you correlating against? how far do you go back to get the normal value etc?

short term fluctuations could reverse back to a long term value without changing it much or it could herald the start of a long term change. in either case you would have to back test your theories to find out...

i agree with much of what you say, however i must admit i didn't understand all the finer details of your points!

yes you can make this study as simple or complex as you like. striking that middle balance between effort and practicality of results will take some experimentation.

the two pair correlation i used was just by way of example. lt can be as many as you like really.

here are some more thoughts:

the look up tables weight all the data points equally. you could apply some weighting formula - exponential, linear etc - to weight in favour of newer data at the expense of older, like they do with moving averages. that way you could plot data from the earliest point available and just extend it to what ever testing period you want.

you could also use the correlation to weight the position size of two or more trades you might want to take.

e.g the eurusd & usdchf have +90% anti-correlation, so trade half each when you get your signals.

if you have two pairs that have 60-80% correlation then readjust your weightings to perhaps 60-40. which would get which side is a matter of experimentation.

and you could go on and on...

for discretionists like myself i don't really want to get into deep besides finding general values for T and t. but for auto traders you could go crazy on this topic!

hi iso,

maybe i written too much, and it is hard to understand my point... sorry for that english is not my mother tongue so that is a small problem too. i will try to reduce it somehow.

first let me ask, if i understand your goal.
in simplified terms, you are trying to use the correlation tables to determine some weights for your possitions in different pairs, right?
let's assume you are doing it only for two currency pairs (with more pairs involved concurrently, the problem gets exponentially bigger)
you want to achieve that by trying different combinations of weights - directly derived from the corresponding correlation between the two currency pairs.

i'm missing one thing... let's suppose you make such tests (ussing weights from 0 to 1 with step 0.1), what do you do with the results? what exactly are the results? what values are you interested in? how do you want to compare them? how do you select the best one?

ok, i will wait for your answers, so i'm sure i understand your goals.
thank you

RoBiK

Sorry, I forgot to quote your message (oanda fxlabs):
Correlation between EURUSD and EURJPY:

1 week 0.82

1 month -0.09

3 months 0.43

6 months 0.60

1 year 0.69

2 years 0.09

I seems to me that the -0.09 is different from "normal" values. And maybe those "abnormal" values are compensated in the future, like with these 2 pair, because in the last week the value was suddenly much higher then normal. But you are right, it is just some gamble, I should test it with some historical values.

Sorry, I forgot to quote your message (oanda fxlabs):
Correlation between EURUSD and EURJPY:

1 week 0.82

1 month -0.09

3 months 0.43

6 months 0.60

1 year 0.69

2 years 0.09

I seems to me that the -0.09 is different from "normal" values. And maybe those "abnormal" values are compensated in the future, like with these 2 pair, because in the last week the value was suddenly much higher then normal. But you are right, it is just some gamble, I should test it with some historical values.