Machine Learning with algoTraderJo | Page 40 | 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

  • #781
  • Nov 8, 2015 10:26am Nov 8, 2015 10:26am
Real word vs simulation

The close day of the trade is not same in two figure but i think i still interesting

August to October this with the first version of the SVC
Attached Image

Attached Image
  • #782
  • Nov 8, 2015 11:47am Nov 8, 2015 11:47am
Quoting algoTraderJo
{quote} I tend to agree with that. Even if you have some improvement in Sharpe ratio with more training it's not worth the additional time because the DMB analysis would be impractical. With the NN training 50 iterations it takes about 2x the time of a linear regression algo, which makes it feasible for system searches and DMB analysis. To answer your other question, I can't imagine how expensive it must be to do the DMB of this image based systems. Such systems are by their very nature complex and slow, imagine having to search parameters that...
No I haven't done any research on them, I was just wondering whether or not to add it to my to-do-list. But I think for the reason you posted, and since there are many simpler things I haven't tried yet, I will put it on the back burner for a while
  • #783
  • Nov 8, 2015 1:08pm Nov 8, 2015 1:08pm
Another parameter is the initialization of the random weights when training the NN algo. If you're a good observer and you read the last code I posted you might have noticed that weights in the NN are initialized to random values between -0.5 and 0.5 before training (good rule of thumb) but nothing tells us not to try other initialization ranges. This is what happens when you change the initialization range (using the last posted EUR/USD system for reference):

Attached Image


It turns out that -0.5/0.5 is probably not the best initialization range. With lower initialization ranges we get better values while with larger ranges we start getting worse results. Lowering the initialization range to -0.1/0.1 or -0.2/0.2 actually works better and is faster because the values are closer to the final weights that we get to in each training. The world of NN can be complex but little by little we learn and learn...
  • #784
  • Edited 3:07pm Nov 9, 2015 12:11pm | Edited 3:07pm
For a Long trade, if you get a very straightforward upside tendency within you horizon,

You will not get any SL for your Long and you will get a SL for the short.

This example will be skipped by the algo because you need to have both target reach in your output for the training [longSL,shortSL] .

Why not use specific level for trying to keep this in your training process ?

You could use something like :
if you get short SL and none Long SL
and the close is more than XX pips than the open
then set the long SL as you XX pip

instead of nothing you will get : [LongXXpip,shortSL]
  • #785
  • Nov 9, 2015 3:22pm Nov 9, 2015 3:22pm
What is your opinion about which is the best input/output for NN.
For instance, i see that my test NN answer better if I consider close values than using yields values.
What do you think?
  • #786
  • Nov 9, 2015 3:57pm Nov 9, 2015 3:57pm
I am pleased that this thread has now moved on to NN as this is an area I am particular interested in. As a hopefully useful contribution I would draw readers' attention to https://cseweb.ucsd.edu/classes/wi08...98b.pdfhttp:// where a pdf file is available with some really useful tips with regard to improving backprop performance.

Readers might also be interested in http://www.ntu.edu.sg/home/egbhuang/elm_codes.html where Extreme Learning Machines are discussed, along with code. Extreme learning machines are essentially NNs that are trained without the the need for backprop and hence can be trained much quicker.
  • #787
  • Nov 9, 2015 8:28pm Nov 9, 2015 8:28pm
I get something not too far from your resul (trailing sl version).
I also force a issued for the output at the end of the horizon if i get only a SL or TP. Without that i dont have enough data (2 years of data before first prediction!)

But my result is less linear than yours.
I need to make a couple of check up again..
Attached Image (click to enlarge)
Click to Enlarge Name: figure_1.jpg Size: 128 KB
  • #788
  • Nov 10, 2015 6:21am Nov 10, 2015 6:21am
Quoting ForexSamsam
For a Long trade, if you get a very straightforward upside tendency within you horizon, You will not get any SL for your Long and you will get a SL for the short. This example will be skipped by the algo because you need to have both target reach in your output for the training [longSL,shortSL] . Why not use specific level for trying to keep this in your training process ? You could use something like : if you get short SL and none Long SL and the close is more than XX pips than the open then set the long SL as you XX pip instead of nothing you...
Yes you could try that but then you would introduce a difference in what is being predicted. The idea is to predict the trade outcome when using a trailing stop, if your output sometimes mismatches this and is instead some other value then you're injecting a distortion into your training. If you see this happening then you need to increase your horizon such that skipping becomes very rare.
  • #789
  • Nov 10, 2015 6:23am Nov 10, 2015 6:23am
Quoting sahara85
What is your opinion about which is the best input/output for NN. For instance, i see that my test NN answer better if I consider close values than using yields values. What do you think?
What does it mean that it "answers better" ? Be careful how you measure NN performance! If you are measuring NN performance by NN accuracy then evidently the close values will perform better because the series of close prices is heavily autocorrelated. You should measure your NN performance by actual trading profitability in a trading system and not by prediction accuracy otherwise you will be fooled by autocorrelations that increase apparent prediction accuracy but would be meaningless in actually generating more profits. What input/outputs work? Read the thread for many examples of inputs and outputs used in NN and other machine learning algos.
  • #790
  • Nov 10, 2015 6:27am Nov 10, 2015 6:27am
Quoting ForexSamsam
I get something not too far from your resul (trailing sl version). I also force a issued for the output at the end of the horizon if i get only a SL or TP. Without that i dont have enough data (2 years of data before first prediction!) But my result is less linear than yours. I need to make a couple of check up again.. {image}
Your results are bound to be worse because of the distortions you are introducing. When you train the NN all your training examples must be built equally if you have problems with the hitting of targets then increase the horizon further but don't introduce examples where targets are not hit. You are using a different NN algo implementation so it might also be that you have different optimizers that require different variable initialization ranges, training iterations, etc. I also advice you purchase some long term data if you have problems with data length. You can join a place like Asirikuy where data is included or you can buy it from a data seller.
  • #791
  • Nov 10, 2015 9:23am Nov 10, 2015 9:23am
Quoting algoTraderJo
{quote} Yes you could try that but then you would introduce a difference in what is being predicted. The idea is to predict the trade outcome when using a trailing stop, if your output sometimes mismatches this and is instead some other value then you're injecting a distortion into your training. If you see this happening then you need to increase your horizon such that skipping becomes very rare.
Every day on A hour, how your machine know to not entrer in a trade ?

In my understanding, to enter long, you use long>0 and short<0 but did you use minimun return value to enter ?

If the idea is to predict the outcome of a trailing stop, why you dont set the minimun value as the trailing stop itself (20day ATR) ?

You could get [1pip,-1pip], this case you will enter for long and you could never hit the trailing stop or close in reverse

If the data is not continues and you consdider every A hour for trading, then you introduce difference between the training and the trading.
  • #792
  • Nov 11, 2015 8:56pm Nov 11, 2015 8:56pm
when you say A=8 ..

This means A = 8PM in us system (20h in HH24)

or 8AM
?

thanks

sam
  • #793
  • Edited 7:58pm Nov 12, 2015 7:19pm | Edited 7:58pm
I have something very strange, i get same result than you but 12h later.
Your 8h is for me 20H.

8AM in GMT+1 is the end of Sydney session, this not a very decisive time for EURUSD.. i think. ?
My 20h seems to have more sens, this after the end of London and New York near to close
Attached Image (click to enlarge)
Click to Enlarge Name: EURUSD 20 180 14 200 110.0_BEST.png Size: 33 KB
  • #794
  • Nov 17, 2015 3:10pm Nov 17, 2015 3:10pm
Hi algoTrader

First of all thanks for such an interesting thread.

I've been able to reproduce some of your results with F4 but I've noticed this algos are very feed dependant as you can see in the chart.

Settings:

SL = 160% ATR20
A = 1 GMT +1/+2
B = 90
C = 5
D = 90

Attached Image (click to enlarge)
Click to Enlarge Name: SEL.jpg Size: 93 KB


Is it always like that with this kind of logic? Is there any way to reduce the feed dependancy?

Thanks!
  • #795
  • Nov 17, 2015 8:34pm Nov 17, 2015 8:34pm
3 pairs mix EURUSD, EURJPY and USDCHF
each have it's own parameter
Attached Image (click to enlarge)
Click to Enlarge Name: EURUSD_EURJPY_USDCHF.png Size: 49 KB
  • #796
  • Nov 26, 2015 1:24am Nov 26, 2015 1:24am
Quoting ForexSamsam
3 pairs mix EURUSD, EURJPY and USDCHF each have it's own parameter {image}
Very very impressing. I am think I am close to the first version of my trading strategy based on NN. Cause I have made during my bus travelling to get to work and also cause I was no really familiar with NN it took a lot. I am still skeptical on what I have done so far...cause R is close to 0,8. Of course I have followed the best practice and I have based everything on values return instead of simply values.

One question ... what R do you reach with your models?

Thanks for answers
Best regards
  • #797
  • Nov 27, 2015 8:19am Nov 27, 2015 8:19am
My stat with my 6 pairs version:

Formula: Y ~ <x> + <intercept>
Number of Observations: 8058
Number of Degrees of Freedom: 2
R-squared: 0.9731
Adj R-squared: 0.9731
Rmse: 4816.8535
F-stat (1, 8056): 291372.0210, p-value: 0.0000
Degrees of Freedom: model 1, resid 8056
  • #798
  • Nov 27, 2015 7:18pm Nov 27, 2015 7:18pm
Quoting ForexSamsam
My stat with my 6 pairs version: Formula: Y ~ <x> + <intercept> Number of Observations: 8058 Number of Degrees of Freedom: 2 R-squared: 0.9731 Adj R-squared: 0.9731 Rmse: 4816.8535 F-stat (1, 8056): 291372.0210, p-value: 0.0000 Degrees of Freedom: model 1, resid 8056
With 6 pairs mix R = .9731
Maybe we can improve again with half hour timeframe ?
Attached Image (click to enlarge)
Click to Enlarge Name: figure_1.png Size: 34 KB
  • #799
  • Nov 29, 2015 9:38am Nov 29, 2015 9:38am
I found something interesting..
The ratio between D and SL is constant.. D/SL = Y
Finding the good Y, you can move D and SL to optimize the slope.
Each pair have its own Y depending of the pair caracteristic.
  • #800
  • Dec 3, 2015 1:22pm Dec 3, 2015 1:22pm
Quoting algoTraderJo
These are some initial sample EUR/USD results using the before mentioned NN function for predictions, the trailing stop as explained before and the input and output structure used for previous linear regression 1H strategies. As in all cases the ML algo is retrained before making each trading decision. These are the parameters and results: SL = 160% ATR20 A = 8 GMT +1/+2 B = 130 C = 2 D = 115 {image} I have put together results for different numbers of training iterations for the NN on the graph above. This means that when the NN is trained before...
Hi algoTraderJo,

If I understand this correctly, the outputs of the NN's would be the MLE and MSE in the following D hours, where C is the number of most recent returns to use? In all the examples that follow, I see that D is much bigger (115 in this case) than those used in the linear regression examples (8 or 10 or 12 etc.). Why would one want to use such big D when the strategy trades on H1 data?

Also, beside looking at the final return achieved using the profit target / stop loss in trading, I think one fundamental benchmark is the accuracy of predicting whether MLE or MSE is larger, since trades are initiated based on these two predictions. Could you comment on the accuracy of your NN examples? I checked on one of the linear regression examples and found that the accuracy was around 51%. It is not very impressive, but shows that the profits of the simulated trading seem to come from a good stop loss control.

Finally, could you clarify if you initialize a different NN and train it for a certain number of iterations each time for a new bar, or use the same weights each time? I think you are using a different weight initialization for each bar but just want to make sure. In that case, I think if you run it twice, the result may be very different because of the random nature of the algorithm?

Thanks for running this thread. The results and discussion are very interesting.
2 traders viewing now