Menu
Sister Sites

Options

Bookmark Thread

First Page First Unread Last Page Last Post

Print Thread

Similar Threads

Trend Bias and Weighted Trade Signals Trading Tools 57 replies

Weighted Coin Toss Trading 118 replies

Weighted Escalator - a weighted stoch system 51 replies

Linear weighted T3? 0 replies

Linear Weighted MA's 0 replies

How to add weighted conditions in trading formula

  • Post #1
  • Quote
  • First Post: Feb 20, 2015 4:10pm Feb 20, 2015 4:10pm
Consider two formula's A and B. For example, A checks if RSI (for example >50) and B checks if CCI (for example >50).

In a normal trade setup we could have (pseudo-code) IF (A>50 && B>50) THEN {....trade....}

How to weight the formula's?

Using a % factor may help. Consider X and Y which are factors;

IF ((X/100 * A + Y/100 * B) / 2) > 50) THEN {....trade....}

But, there is clearly a problem here; a result of 0.10 on the first bit (X/100 * A) and a result of 0.60 on the second bit (Y/100 * B) have the same outcome as 0.60 on the first and 0.10 on the second, while they are very different values.

Another problem is that if check A is >50 and check B > 70 we now have to do something like " > (50+70/2=60) " - all very messy.

Instead, how does one write a formula with proper weights (and && statements) which produces a single yes/no trade outcome?
Not financial advice. For testing purposes only. Do not use on live account
Quoting roel13
Consider two formula's A and B. For example, A checks if RSI (for example >50) and B checks if CCI (for example >50). In a normal trade setup we could have (pseudo-code) IF (A>50 && B>50) THEN {....trade....} How to weight the formula's? Using a % factor may help. Consider X and Y which are factors; IF ((X/100 * A + Y/100 * B) / 2) > 50) THEN {....trade....} But, there is clearly a problem here; a result of 0.10 on the first bit (X/100 * A) and a result of 0.60 on the second bit (Y/100 * B) have the same outcome as 0.60 on the first and 0.10 on...
If (A[1]>50&&B[2]<70&&B[1]>70) return(-1);

Assuming that A is the weaker of the signals, and you want it to be above 50, and want to wait until your B is greater than 70, you use these as your conditions.

Inside of the [] are your periods. [1] is the last candle. [2] is 2 candles ago.

If [2] candles ago B was less than 70 and in the [1] current candle we are greater than 70, we are looking for a buy.

Previous and post conditions aren't typically what folk that don't work with code think about...they say shit like, if (a>50 and b>50), and then from here will set a take profit and stop loss...issue being, if this condition is still present after the trade closes in profit or loss, another trade immediately opens up.

You need your conditioning to read, that if the condition wasn't present, but is now, take trade.

Does this help??
Money Can't Buy Happiness. Poverty Can't Buy SHIT! You Choose!
Quoting Magix
You need your conditioning to read, that if the condition wasn't present, but is now, take trade.
Yes, agreed.

So, is there a way to give a weight (0-100) for each condition? For example, when optimizing with a neural network one could want to make some conditions (auto-optimized) more or less important then another in different cycles of the market.

So, the formula could be something like

factor1 (0-100) * indicator value 1 condition
factor2 (0-100) * indicator value 2 condition
factor3 (0-100) * indicator value 3 condition
factor4 (0-100) * indicator value 4 condition

Where "if factor 1 is 0, indicator value 1 does not matter" and "if factor 2 is 100, indicator value 2 matters the most" and "if factor 3 is 70 and factor 4 is 20 then factor 3 weights more heavily then factor 4"

Can this be mathimatically/code wise easily done? Bool (on/off) is too limited; the neural network sees only yes/no, tries a few times with one indicator, finds a good value and turns off the other ones. The problem is that the other ones are able to fine tune that first one.

Call it "weighted entry" or something.
Not financial advice. For testing purposes only. Do not use on live account
Quoting roel13
{quote} Yes, agreed. So, is there a way to give a weight (0-100) for each condition? For example, when optimizing with a neural network one could want to make some conditions (auto-optimized) more or less important then another in different cycles of the market. So, the formula could be something like factor1 (0-100) * indicator value 1 condition factor2 (0-100) * indicator value 2 condition factor3 (0-100) * indicator value 3 condition factor4 (0-100) * indicator value 4 condition Where "if factor 1 is 0, indicator value 1 does not matter" and...

Not really getting you here...

Still not sure why this would be needed, unless you are creating a blue moon trader, ie only looking for conditions that happen, once every blue moon...

Almost anything can be mathematically coded, sure there's some work to it, but it's not like you have to do this for every trade...

When you check the top left hand corner of your indicator, it gives you the name, values and a number...everything else in that window is a visual representation by line or histogram...they are all math.

We see the pictures, the code sees the numbers.

But, without being absolutely clear on what it is you are trying to accomplish, this is about all I can add...
Money Can't Buy Happiness. Poverty Can't Buy SHIT! You Choose!
This seems to work and examplifies what I am thinking of;

Inserted Code
double w1=50/100.00;
double w2=50/100.00;
aa[i]=iRSI(NULL,0,3,PRICE_OPEN,i);
bb[i]=iADX(NULL,0,3,PRICE_OPEN,MODE_MINUSDI,i);
cc[i]=((aa[i]*w1)+(bb[i]*w2));

w1 =weight for the RSI (auto-optimized), w2=weight for the iADX (idem). cc[i]=can be used for trading decision, i.e. cc[i]>50, for example.

The challenge here is that all of the values are merged into one chart, and so the "overall condition" (for example cc[i]>50) is just one "overall" condition.

How to mathematically word a condition on each item aa[i] and bb[i] while maintaining the importance of the weights.
Not financial advice. For testing purposes only. Do not use on live account
Quoting roel13
This seems to work and examplifies what I am thinking of; double w1=50/100.00; double w2=50/100.00; aa=iRSI(NULL,0,3,PRICE_OPEN,i); bb=iADX(NULL,0,3,PRICE_OPEN,MODE_MINUSDI,i); cc=((aa*w1)+(bb*w2)); w1 =weight for the RSI (auto-optimized), w2=weight for the iADX (idem). cc=can be used for trading decision, i.e. cc>50, for example. The challenge here is that all of the values are merged into one chart, and so the "overall condition" (for example cc>50) is just one "overall" condition. How to mathematically word a condition...
Inserted Code
cc[1]=((aa[i]*w1)+(bb[i]*w2));
cc[2]=((aa[i]*w1)+(bb[i+1]*w2));

??
Money Can't Buy Happiness. Poverty Can't Buy SHIT! You Choose!
Quoting Magix
{quote} cc[1]=((aa[i]*w1)+(bb[i]*w2)); cc[2]=((aa[i]*w1)+(bb[i+1]*w2)); ??

Scratch that...I have no idea....



I'd send a pm off to This Guy!
Money Can't Buy Happiness. Poverty Can't Buy SHIT! You Choose!
Lets say that you have 5 rules that you wish to evaluate and they each have different weights assigned to them.

Rule 1..Rsi > 50 weight 25
Rule 2 Close[1] > Close[2] weight 30
Rules 3 Close[1] > EMA55 weight 15
Rule 4 CCI > 100 weight 15
Rule 5 Close[1] > Prev Week High 15

The weights total 100. Let's pretend that you used a minimum threshold of 75 to evaluate as True.

So if Rules 2,3,4,5 were true then the weight total is 75 and the signal is fired. If Rules 1,2,3,4 is true then the signal is fired also.

Use the number of rules you need and apply the weights as needed then just decided on a threshold. Just keep accumulating the total score and then compare to your threshold.

Alternatively use Fuzzy Logic. There are a couple of fuzzy logic function sets that you can google for MT4.
  • Post #9
  • Quote
  • Edited 2:18am Feb 21, 2015 12:00am | Edited 2:18am
@roel13

As an alternative approach, you can weight multiple indicator values something like:

weighted_result = (indicator1_value x weight1 + indicator2_value x weight2 + indicator3_value x weight3 + ...) / (weight1 + weight2 + weight3 + ...)

Then you could use the weighted_result to make your trading decisions.

So, for example, if you want to apply weights of 1, 2 and 3 to different indicator values of 20, 40 and 60, respectively, the result would be:
(1x20 + 2x40 + 3x60) / (1+2+3)
= 280 / 6
= ~ 46.7

That makes sense, because the highest value of 60 gets the highest weighting of 3, thus skewing the final result above the median value of 40.


However, to be mathematically sensible, the above assumes that all indicators have similar upper/lower limits (e.g. RSI ranges between 0 and 100).

You can normalize anything with different limits to a value between 0 and 100 using something like:

indicator1_value = 100 x (indicator_value – low_limit) / (high_limit – low_limit)

and then plug the result indicator1_value into the initial formula.

For example, if an indicator ranges between a low value of –200 and a high value of +200, and is currently at +70, then its normalized value would be:
100 x (70–(-200)) / (200–(-200))
= 100 x 270 / 400
= 67.5


But it all depends on what you're trying to achieve, and what your rules are.
Hi roel13,

Yes it is possible to weight formulas. But it sounds like you're still trying to decide what to do. Basically if you are dealing with boolean conditions: true/false you can use a simple voting scheme where 1 means your indicator indicates a long signal and -1 means a short signal and 0 means no signal. Do this for all your indicators.

Then if the sum of your {-1, 0, 1} signals is greater than zero (or some other threshold) you are net long, and if the sum of the signals is less than zero (or some threshold) you are net short. That's one simple way to do it (that I use) and may work fine for your application. Then it becomes more about picking good (complementary) features (indicator inputs) rather than focusing so much on getting the weights perfect. From my experience, overemphasizing the weights can lead to overfitting and all the problems that go along with that.

If you are feeding these numbers into a neural network then the inputs need to be scaled properly (all on the same scale). You could use the same scheme described above, and just allow the NN to come up with the weights on its own. It does this naturally. If you think you have an idea of the strength of a signal, you could try weighting signals from weak long (0.1) to strong long (1.0) and likewise for weak short (-0.1) to strong short (-1.0).

If you are going off an indicator reading and attempting to use that as a weight I think you need to ask yourself whether a 60 RSI is a weaker signal than an 80 RSI, because that's what rating by indicator values implies and this is what the NN will use. So define what you're going after, make an hypothesis and then test to see if it has merit or not. Then move on to an alternative hypothesis and keep testing. The only way to know for sure for your specific application is to get your hands dirty in the data and do lots of testing to see what works best and what fits your needs. Good luck!
Quoting roel13
Consider two formula's A and B. For example, A checks if RSI (for example >50) and B checks if CCI (for example >50). In a normal trade setup we could have (pseudo-code) IF (A>50 && B>50) THEN {....trade....} How to weight the formula's? Using a % factor may help. Consider X and Y which are factors; IF ((X/100 * A + Y/100 * B) / 2) > 50) THEN {....trade....} But, there is clearly a problem here; a result of 0.10 on the first bit (X/100 * A) and a result of 0.60 on the second bit (Y/100 * B) have the same outcome as 0.60 on the first and 0.10 on...
Did you have a look at Sugeno fuzzy logic models? To me it looks like this is what you're looking for. In fuzzy logic the boolean operators AND, OR and NOT become functions. For example x OR y may becomes x+y-x*y. There are several formulas. Note x and y must be normalized between 0 and 1 (they are seen as probabilities).

Here is an intro http://www.bindichen.co.uk/post/AI/t...zzy-model.html

In the first formula on the page "if x is A" can be translated into "if RSI is oversold". The "then z=f(x,y)" translates into open a trade of f(RSI, ADX) lots. If f is negative you're short. This isn't perfectly a "if X then go long" but something that continuously increases or decreases your market exposure. You can easily change that if you don't like it. Note that you don't have a "RSI>0.9" but "RSI is strongly overbought". Those conditions are generated using fuzzy set functions.

A limitation I couldn't perfectly overcome with the fuzzy sets is that someone can be too tall to be "tall" but not enough to be "very tall" and he's classified neither of the two classes.
No greed. No fear. Just maths.
  • Post #12
  • Quote
  • Last Post: Feb 23, 2015 3:05am Feb 23, 2015 3:05am
Quoting flyer415
Use the number of rules you need and apply the weights as needed then just decided on a threshold. Just keep accumulating the total score and then compare to your threshold. Alternatively use Fuzzy Logic. There are a couple of fuzzy logic function sets that you can google for MT4.
Quoting PipMeUp
Did you have a look at Sugeno fuzzy logic models? To me it looks like this is what you're looking for. In fuzzy logic the boolean operators AND, OR and NOT become functions. For example x OR y may becomes x+y-x*y. There are several formulas. Note x and y must be normalized between 0 and 1 (they are seen as probabilities).
Thanks flyer415 & PipMeUp, started looking into Fuzzy Logic

Quoting FXEZ
Then if the sum of your {-1, 0, 1} signals is greater than zero (or some other threshold) you are net long, and if the sum of the signals is less than zero (or some threshold) you are net short. That's one simple way to do it (that I use) and may work fine for your application.
Thanks FXEZ, this is also a great idea. One thing I like about it is that it "pushes out" values to their max. For example, RSI>val1 and RSI<val2 - val1 and val2 can be optimized. Then a range (20 wide, step 0.1) can be created between val1 and val2 and the more it leans towards val1 (==1.0) it's a long or towards val2 (==0.0) it's a short or vice versa, etc. Almost-fuzzy?

Quoting hanover
weighted_result = (indicator1_value x weight1 + indicator2_value x weight2 + indicator3_value x weight3 + ...) / (weight1 + weight2 + weight3 + ...)
Cool alternative, thanks Hanover.

Also used the normalize range before. It's handy, but normalizing fails a bit when indicators have odd/irregular values - have you found a good solution for this - ranges maybe?
Not financial advice. For testing purposes only. Do not use on live account
0 traders viewing now