Machine Learning with algoTraderJo | Page 29 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Algorithmic Quant Trading (Machine Learning + Stat-Arb) 25 replies

Forex with Machine Learning Software Project 1 reply

Machine Learning + Retail Forex = Profitable? (Quant) 1 reply

Potential new machine learning style software. 79 replies

My most recent advancements into machine learning 16 replies

Attachments: Machine Learning with algoTraderJo
Exit Attachments

Machine Learning with algoTraderJo

  • #561
  • Jul 7, 2015 11:34am Jul 7, 2015 11:34am
Here my first demo account result.... im just lucky..

http://www.forexfactory.com/forexsamsam#55

I open short in svc- and long in svc+. I can't hegde on oanda..
  • #562
  • Jul 7, 2015 11:53pm Jul 7, 2015 11:53pm
Hi ForexSamsam

What time frame do you use?
What is your SVC output?
do you use stop loss or you close at end of forecast period?

Thanks
Lev
  • #563
  • Jul 9, 2015 12:03pm Jul 9, 2015 12:03pm
Quoting levkly
Hi ForexSamsam What time frame do you use? What is your SVC output? do you use stop loss or you close at end of forecast period? Thanks Lev

I use Daily data only.

My target is +ATR or - ATR [0,1] , the first reach (ATR 20 day same as algotraderjo)

Every day, i open a new trade according to SVC prediction.

This with ATR20 as stroploss or takeprofit.

The real difficulty of this setting is to get good data source... I'm using every tick to know witch of +/-ATR is reach first
  • #564
  • Aug 4, 2015 4:31pm Aug 4, 2015 4:31pm
Machine learning enthusiasts!

Algotraderjo is back with more machine learning fun for all the family. I cannot post as frequently as I would want due to my trading business (that takes up a lot of my time) but I really try to give you some juicy stuff whenever I can come back to Forexfactory to have some good discussions with like-minded folk.

About all the PMs I have received, I believe most of you must have already found a solution to the issues you were attempting to solve, if you still need my help then send me another one (a copy of the last one is just fine) and I'll reply. I would happily comb through everything if I had the time, but that's not the case.

Some of you have asked about the use of take profits in machine learning algos and today I want to address this point a bit.

Take for example this generic trading system scheme:

  1. Uses a takeprofit and a stoploss expressed as a % of the daily ATR.
  2. Attempts to predict the outcome of a long or short taken at the current bar using machine learning
  3. If it predicts that a buy would have a positive result and a sell a negative result it then enters a long and vice versa
  4. Opposite signals close previously open trades
  5. Signals in the same direction reset the SL/TP values to whatever they would be if we had just entered a new trade
  6. The system only generates signals at a specific hour of the day (A)
  7. System trades on the 1H chart

Then we create a machine learning algo that:

  1. Builds a database on each bar using the past B examples where each example is built from a bar in the past with the same hour, C bar return values, a horizon D that determines how far into the future we look to see if the SL/TP is hit and a predictor that is the trade outcome in absolute price units.
  2. The algorithm is retrained on each bar where a decision is made. The algorithm uses a linear regression approach with C inputs and 2 predictors.
  3. The algorithm then uses the past C returns to create a decision for the current bar. We get two predictors, one for a long term entered and another for a short trade entered.
  4. System risks 0.5% per trade. I used 1% on most other posts on this thread but TP taken systems trade more frequently and hence need a risk reduction to remain in overall acceptable risk adjusted return values.

We can use the following parameters on the EURUSD. As stated above, the ML is retrained on every bar:

A=10
B=50
C=19
SL=40% of the ATR20
TP=210% of the ATR20

To generate these results for different horizons (D):

Attached Image (click to enlarge)
Click to Enlarge Name: machinelearning_sltp_sample1.jpg Size: 116 KB


I want you to reflect on the effect that using a TP has compared to the trading results I posted before (pay special attention to the monthly returns in the middle of the graph). I also want you to think about why the changes in the horizon (shorter or longer) cause such a significant difference in trading results. Why is the horizon so critical if database values should be the same at long horizon values? Thinking caps on!
  • #565
  • Aug 4, 2015 4:39pm Aug 4, 2015 4:39pm
I also recently found out that the guys at Asirikuy have now started a cloud mining effort to calculate data-mining bias for machine learning strategies (this means that many members contribute to the calculation of mining bias, which is by nature very computationally expensive) . I have just emailed Daniel to ask if I can use some of the graphs generated by them for illustrative purposes so that you guys can see some of the output graphs from applying machine learning generation procedures on random data and real data. I will however not go into the specific algos they are using for this as I both respect the privacy of their community effort and want to focus on some other things here.
  • #566
  • Aug 4, 2015 4:49pm Aug 4, 2015 4:49pm
Watch the effect of changing the TP:

Inserted Code
A=10
B=50
C=19
D=100
SL=40% of the ATR20
Attached Image (click to enlarge)
Click to Enlarge Name: machinelearning_sltp_sample2.png Size: 96 KB
  • #567
  • Aug 5, 2015 12:04am Aug 5, 2015 12:04am
Hi AlgotraderJo,

Good to see you back,

I have 2 general questions,

1. How do you deal with different feeds (different brokers) for same currency?

2. What you suggest to change in ML strategies on currencies like USDCAD ,AUSCAD (more MR currencies )?

- I think TP depend on the horizon (D) , in average after D bars price change X%.
  • #568
  • Aug 5, 2015 6:02am Aug 5, 2015 6:02am
Quoting levkly
Hi AlgotraderJo, Good to see you back, I have 2 general questions, 1. How do you deal with different feeds (different brokers) for same currency? 2. What you suggest to change in ML strategies on currencies like USDCAD ,AUSCAD (more MR currencies )? - I think TP depend on the horizon (D) , in average after D bars price change X%.
1. Depends, I only ensure that results do not change very significantly on the broker where I want to live trade. I will generally perform a back-test in Oanda Vs 1987-2015 data using Oanda data from 2004 and in this way I'll check that results do not vary on the feed where I intend to use them. I have never found it necessary to check across other feeds different than my live broker.

2. This is tricky business! It took me years to get decent results on pairs like these. I won't give you the answer but I can tell you one thing: you do not need more data.

3. This is true for small values of D but when D grows large whether the TP is hit or not is unaffected by the value of D and yet the results still change. Why?
  • #569
  • Aug 5, 2015 6:12am Aug 5, 2015 6:12am
We can also get highly linear results (correlation coefficient > 0.94) using this type of strategy.

Inserted Code
SL=40%
TP=280%
A=10
B=75
C=11
D=150

Attached Image
  • #570
  • Aug 5, 2015 6:29am Aug 5, 2015 6:29am
Quoting algoTraderJo
3. This is true for small values of D but when D grows large whether the TP is hit or not is unaffected by the value of D and yet the results still change. Why?
The maximum likelihood of the TP is smaller for a small horizon but its variance is also smaller. At longer horizon the maximum likelhood of TP is greater but the variance is much greater. The variance grows linearly with time but the TP only grows as a square root. The slope of sqrt(.) is greater at the small values: waiting more is less and less efficient and more and more uncertain. Additionally the longer you wait the less confident you are about which of the SL or the TP comes first.
No greed. No fear. Just maths.
  • #571
  • Aug 5, 2015 2:47pm Aug 5, 2015 2:47pm
Quoting PipMeUp
{quote} The maximum likelihood of the TP is smaller for a small horizon but its variance is also smaller. At longer horizon the maximum likelhood of TP is greater but the variance is much greater. The variance grows linearly with time but the TP only grows as a square root. The slope of sqrt(.) is greater at the small values: waiting more is less and less efficient and more and more uncertain. Additionally the longer you wait the less confident you are about which of the SL or the TP comes first.
I don't think I agree with "the longer you wait the less confident you are about which of the SL or the TP comes first". When I create the examples used for the building of the training dataset on each bar I only care that the SL or TP is hit within the chosen horizon (D). If the SL or TP is not hit then I have a problem because I cannot asign the value for the predictor (there is no defined exit for the example).

Suppose I am building an example where the TP is hit before the SL and the TP is hit on bar 50 after entry. If I choose a horizon of 40 then I am in trouble because none of the two have been hit but if I choose 50, 60, 100 or 200, then the predictor value is always that the TP is hit with a profit of X because, no matter what, the TP was always hit 50 bars after entry no matter how much further I could have kept looking.

The question is: Why do results change if I set the horizon to a larger value if the value of the predictors remains unchanged after some point? Why would 500 and 1000 give different results even though the construction of the examples has the exact same predictor values?

Tip: It has nothing to do with the SL/TP.
  • #572
  • Aug 5, 2015 11:33pm Aug 5, 2015 11:33pm
Quoting algoTraderJo
{quote} The question is: Why do results change if I set the horizon to a larger value if the value of the predictors remains unchanged after some point? Why would 500 and 1000 give different results even though the construction of the examples has the exact same predictor values? Tip: It has nothing to do with the SL/TP.
As you described, having a larger value for horizon also means that we are excluding the past samples. So a value of 1.000 means we are excluding 1000/24 = 42 days. The model would lack recent information.

A way to get around this issue is to have the finding of TP or SL dynamic and include the samples that already found a target value. This however is computationally expensive.
  • #573
  • Aug 6, 2015 2:28am Aug 6, 2015 2:28am
Quoting algoTraderJo
Suppose I am building an example where the TP is hit before the SL and the TP is hit on bar 50 after entry. If I choose a horizon of 40 then I am in trouble because none of the two have been hit but if I choose 50, 60, 100 or 200, then the predictor value is always that the TP is hit with a profit of X because, no matter what, the TP was always hit 50 bars after entry no matter how much further I could have kept looking.
I thought you were still using the maximum positive and negative excusion to estimate the direction; what you call MLE/MSE. In which post I missed did you changed this?
No greed. No fear. Just maths.
  • #574
  • Aug 6, 2015 6:54am Aug 6, 2015 6:54am
Quoting PipMeUp
{quote} I thought you were still using the maximum positive and negative excusion to estimate the direction; what you call MLE/MSE. In which post I missed did you changed this?
No I changed the technique since I restarted posting, read this post for more details about the last ML algos http://www.forexfactory.com/showthread.php?p=8417689#post8417689.
  • #575
  • Aug 6, 2015 6:56am Aug 6, 2015 6:56am
Quoting KaBo
{quote} As you described, having a larger value for horizon also means that we are excluding the past samples. So a value of 1.000 means we are excluding 1000/24 = 42 days. The model would lack recent information. A way to get around this issue is to have the finding of TP or SL dynamic and include the samples that already found a target value. This however is computationally expensive.
Precisely the point. Another question, is including data up-to-present always the best thing to do?
  • #576
  • Aug 6, 2015 7:14am Aug 6, 2015 7:14am
Quoting algoTraderJo
{quote} Precisely the point. Another question, is including data up-to-present always the best thing to do?
I tried and it wasn't improving the results (much), but added a lot of computation time as you need to loop through all the bars that don't have a target/output yet. I thought being able to include the past couple of samples would add value, but it seems marginal. Since i only did a couple of experiments, this is just my subjective finding of course. And i did not pursue this road any further.
  • #577
  • Aug 6, 2015 7:40am Aug 6, 2015 7:40am
We can also get good linear results (correlation coefficient = 0.96) on the GBPUSD:

Inserted Code
SL=70%
TP=220%
A=14
B=225
C=9
D=70

Attached Image
  • #578
  • Aug 6, 2015 11:19am Aug 6, 2015 11:19am
I will start a new backtesting engine this week with you A,B,C,D setting and GBPUSD.
Your result seem very goog !
  • #579
  • Aug 6, 2015 11:39am Aug 6, 2015 11:39am
Quoting ForexSamsam
I will start a new backtesting engine this week with you A,B,C,D setting and GBPUSD. Your result seem very goog !
Remember that these systems work on the 1H chart and also remember that I am not using MSE/MLE here but a new algo (read this post carefully http://www.forexfactory.com/showthre...89#post8417689). This is also a linear regression algorithm, if you want to reproduce avoid using classifiers.
  • #580
  • Aug 6, 2015 12:05pm Aug 6, 2015 12:05pm
in your last example i understand :

at 14h GMT compute return for the last 9 day bar return at the same hour (14h gmt return in pip or percent) -- >this is the input like [a,b,c,d,e,f,g,h,i]
Output is the return from now + 70 hours (delta from now) and your add 1 is TP is reach first or 0 is SL is reach first (2 output vector) like [X,Y]

this for 225day.
2 traders viewing now