Machine Learning with algoTraderJo | Page 38 | 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

  • #741
  • Oct 31, 2015 8:06am Oct 31, 2015 8:06am

  1. For the outputs evolve a long and short trade forward a (D) number of bars. This means moving the virtual SL for the trade as specified in the system rules above. Record the trade exit for both and assign each to a separate output. Each example has C inputs and 2 outputs.


In my understanding, the specified rule implied to move on each hour your SL to keep the same distance, this within the horizon D.

In the training set (output), i dont think you will hit often the SL.. The pair need to move at least ATR20 in the hour. This is relatively rare.

This is not more easy to only use the close at the end of the horizon ?

I think this will not much afecte your learning function (linear regression).

  • #742
  • Oct 31, 2015 8:30am Oct 31, 2015 8:30am
Quoting klubow
{quote} I tested following combinations: Number of samples: 20 .. 240 (step 20) Inputs: past 5 daily bars as - continues values (Close-Open)/Open scaled - binary: -1 for bearish +1 for bullish - binary: -1 for bearish +1 for bullish with some threshold for small movements (those are set as 0) Targets: next day - binary: -1 for bearish +1 for bullish - binary: -1 for bearish +1 for bullish with some threshold for small movements (those are set as 0) The best system I got for FW20WS (future contract for polish WIG20 index): CAR = 4% MaxDD = -13% Sharpe...
That's a very lazy proposition man. If you want to learn the best system that has been posted then read the thread and compare system statistics. The thread establishes clear steps to allow you to create your own systems and get the best out of any ML algo you use. If you want to know what the best system is, I suggest you put the work necessary to find it. The thread guides you in how to do this but I won't be doing the hard work for you. Also consider that getting the "most profitable" is meaningless if you do not measure mining bias. The mining bias of the strategies is not posted on the thread on purpose. Doing this analysis is what will ensure most will never do ML successfully in Forex using the posted guidance, only a few will be willing to put the hard work and effort to perform such an analysis.
  • #743
  • Oct 31, 2015 8:31am Oct 31, 2015 8:31am
Quoting sahara85
Good morning, I am going to join you. I am going to start (just for fun) a project by developing a TS based on NN. Of course I am already worried about few issues that my choices are going to face: In particular I would like to build a Dynamic Neural network based on outputs (feedback) data. The main problem for me will be to avoid the overfitting. For that I would like some advices from you? I hope within this week to set up something in a demo account to start to evaluate this kind of solution. I ll keep you updated and I ll follow closely this...
To avoid over fitting do two things. One, constantly retrain your algorithm before each trading decision using a moving window as shown by the systems on the thread. Two, use as much data as you can.
  • #744
  • Oct 31, 2015 8:33am Oct 31, 2015 8:33am
Quoting sahara85
One simple question to the expert of this thread: Inputs? This is one of the fundamental aspects of building a NN. Which according to your experience are the Inputs necessary for good NN perfomance? I know it's a very generic question. Although I would like to know for instance if technical indicators are useful? Thanks for your answer
Depends on what your outputs are. Different inputs can lead to better predictions on different outputs. Generally you will find it very hard to get any inputs with better performance than simple bar returns. In my own research technical indicators have provided very little values, they always lead to worse results IMHO.
  • #745
  • Oct 31, 2015 8:40am Oct 31, 2015 8:40am
Quoting ForexSamsam
{quote} Another question, if only one hit the target.. what value are you using for the other one. i.e We hit the SL for the long and nothing for the short, then we have [TP,?]
If you don't hit a target you skip that example, it is not a valid example if you don't hit both targets within the used horizon.
  • #746
  • Oct 31, 2015 8:42am Oct 31, 2015 8:42am
Quoting ForexSamsam
For the outputs evolve a long and short trade forward a (D) number of bars. This means moving the virtual SL for the trade as specified in the system rules above. Record the trade exit for both and assign each to a separate output. Each example has C inputs and 2 outputs. In my understanding, the specified rule implied to move on each hour your SL to keep the same distance, this within the horizon D. In the training set (output), i dont think you will hit often the SL.. The pair need to move at least ATR20 in the hour. This is relatively rare. This...
That's a wrong understanding. The SL is moved so that it is always the same but it is only moved if it leads to a more favorable SL than the current one. So if you enter a trade with an SL of 100% of the ATR20 then you will only move the SL if at the start of a bar placing an SL 100% of the ATR20 away from that bar's open price is more favorable than the currently active SL. If on the second bar price moves against you the SL remains unmoved. You do not need price to move an ATR20 in one bar and therefore targets are mostly always hit within a 100 bar horizon.
  • #747
  • Oct 31, 2015 8:57am Oct 31, 2015 8:57am
Quoting fajst_k
{quote} Can you generate the trades of this system and post it together with the currency data ?? I will try to reproduce it. Krzysztof
Sorry I have deleted those trading results (I do tons of research and cannot keep everything that I post stored). If you want to ask questions about things I post please do so rather quickly so that I can answer while the data is fresh and had not been deleted. However you can implement any of the suggested 1H or the 30M strategies and obtain the same or similar results. I will post some 30M systems with trading results soon so that you can analyze them deeper.

As with previous people that have posted if you want to reproduce I strongly recommend you join the Asirikuy community so that you can get the same programming framework and the same currency data. Of course you can reproduce if you can code everything yourself, but it will be much harder and you will potentially be unable to reproduce due to bugs within your own software (very difficult to make sure you have software without bugs when working alone!).
  • #748
  • Oct 31, 2015 9:00am Oct 31, 2015 9:00am
Quoting sahara85
I am sorry to bother you. But among these AI techniques: -SVM -NN -ANFIS which according your experience is the best for short term forex prediction Thanks in advance
Ask yourself if you can perform a thorough data mining bias analysis using any of these techniques. Can you? If you can't, does it make sense to use them? If you could perform the bias measurement with any technique you wanted then your choice of ML technique will, IMHO depend on what specific inputs and outputs you want to use. You can be successful using simple linear regression or you can use neural networks or any other ML method that achieves the results you want, it's not as simple as saying "A is the best ML method". As long as you can adequately measure mining bias and can perform constant retraining through your entire data then you can choose whichever gives you better results on your precisely intended input and output structure.
  • #749
  • Edited 1:30pm Oct 31, 2015 1:01pm | Edited 1:30pm
Quoting algoTraderJo
{quote} That's a wrong understanding. The SL is moved so that it is always the same but it is only moved if it leads to a more favorable SL than the current one. So if you enter a trade with an SL of 100% of the ATR20 then you will only move the SL if at the start of a bar placing an SL 100% of the ATR20 away from that bar's open price is more favorable than the currently active SL. If on the second bar price moves against you the SL remains unmoved. You do not need price to move an ATR20 in one bar and therefore targets are mostly always hit within...
What you mean by "more favorable" ? less loss ?

Let use a long example for understanding..
D = 5
Start with H0 : open 1.0000 SL = 100pip (0.9900)
next hour H1 : 1.0050 SL = (0.9950) change
next hour H2 : 0.9980 SL = (0.9950) keep
next hour H3 : 0.9960 SL = (0.9950) keep
next hour H4 : 0.9990 SL = (0.9950) keep
next hour H5 : 1.0100 SL = (1.0000) change

that right ?
  • #750
  • Oct 31, 2015 1:21pm Oct 31, 2015 1:21pm
Quoting ForexSamsam
{quote} What you mean by "more favorable" ? less loss ? Let use a long example for understanding.. D = 5 Start with H0 : open 1.0000 SL = 100pip (0.9900) next hour H1 : 1.0050 SL = (0.9950) change next hour H2 : 0.9980 SL = (0.9950) keep next hour H3 : 0.9950 SL = (0.9950) keep next
hour H4 : 0.9990 SL = (0.9950) keep next hour H5 : 1.0100 SL = (1.0000) change that right ?
Yes, that's correct.
  • #751
  • Oct 31, 2015 4:34pm Oct 31, 2015 4:34pm
Quoting algoTraderJo
{quote} If you don't hit a target you skip that example, it is not a valid example if you don't hit both targets within the used horizon.
If you dont reach the SL within the horizon you skip the day ?The training set could have very different date , is not a problem i guess ?
  • #752
  • Oct 31, 2015 6:23pm Oct 31, 2015 6:23pm
Quoting ForexSamsam
{quote} If you dont reach the SL within the horizon you skip the day ?The training set could have very different date , is not a problem i guess ?
You just need to have a given number of examples in the training set, there is no rule indicating that such examples must be fully contiguous in the past. You can have 100 examples but you might have had to skip a few because they did not reach one of the targets (making them invalid).
  • #753
  • Oct 31, 2015 6:46pm Oct 31, 2015 6:46pm
Take a look at this example on the EURUSD 30M timeframe. I am using the last described trailing stop that involves no additional parameters. The machine learning algo is a linear regression which is trained after every trading decision, as I explained in past posts. This one uses the same A,B,C,D parameter meaning I used in the past too. To the advantage of those who are Asirikuy members this trailing stop has now been implemented in the official ML system release in F4, so I also attach the configuration and parameters I used to run this test as well as the trade results from the Asirikuy tester (it should be a piece of cake to reproduce using the "F4 stable" library). Exit times indicate the candle where the exit happened, not the time of the precise exit. These are the parameters (I put the meaning there as a reminder for those that want a quick reference):

Inserted Code
SL = 90% of the ATR20
A = 10 GMT+1/+2 (hour for trading)
B = 200 (examples used for training before each trading decision)
C = 25 (number of past bar returns used as inputs per example)
D= 90 (horizon to derive outputs)

Attached Image


This system is both highly efficient and highly linear. It has fairly decent characteristics. You will find (at least in my experience) that the lower you go the more impressive statistics you are in general able to mine from ML but the more your system will depend on your data and the more that it will suffer from changes in the feed and inaccuracies in your simulation (i.e not taking into account slippage, constant spreads, etc). Computer power costs also increase because lower TF simulations are more expensive (more bars to process). It is indeed a tricky problem to know what TF you may want to use. What to trade then? Give me your thoughts!
Attached File(s)
File Type: zip eu_m30_s1_results.zip   221 KB | 395 downloads
  • #754
  • Oct 31, 2015 7:25pm Oct 31, 2015 7:25pm
Quoting algoTraderJo
{quote} You just need to have a given number of examples in the training set, there is no rule indicating that such examples must be fully contiguous in the past. You can have 100 examples but you might have had to skip a few because they did not reach one of the targets (making them invalid).
ok thanks

l'm creating my dataset now... 2000-2015

half of the work is done..
  • #755
  • Oct 31, 2015 7:35pm Oct 31, 2015 7:35pm
Quoting ForexSamsam
{quote} ok thanks l'm creating my dataset now... 2000-2015 half of the work is done..
One important thing IMHO is that you do things when backtesting exactly as you will do when live trading. I like F4 because you use the exact same code when doing both so it's impossible to introduce data snooping bias. By creating a dataset for backtesting you run the risk of falling into problems like this since your strategy is different for simulations and live trading. It's perfectly possible to do it if you're super careful but in my experience the less layers of potential bugs/problems you have the better.
  • #756
  • Edited 5:46am Nov 1, 2015 4:19am | Edited 5:46am
Quoting algoTraderJo
Take a look at this example on the EURUSD 30M timeframe. I am using the last described trailing stop that involves no additional parameters. The machine learning algo is a linear regression which is trained after every trading decision, as I explained in past posts. This one uses the same A,B,C,D parameter meaning I used in the past too. To the advantage of those who are Asirikuy members this trailing stop has now been implemented in the official ML system release in F4, so I also attach the configuration and parameters I used to run this test as...
I think that the 1H timeframe gives a good compromise between computation time/cost whilst still having the ability to take advantage of the daily volatility cycles in the Forex market (unlike 1D). I suppose that going down to the 30M TF may help to achieve a low DMB in pairs such as GBPUSD (as you said before), but I'm not sure whether it becomes easier to find systems on pairs such as AUDCAD or AUDUSD etc as we go further down.

Do you think that simulation inaccuracies when using 30M are too prevalent to consider live trading, or does the possibility of finding low DMB systems for a greater number of pairs outweigh this negative? I suppose the only way(s) to test this would be to verify backtests across different historical data, and of course to obtain lots of real OOS results.

Another thing that makes the situation even more complicated is whether or not there is any benefit from the use of more data on the 30M (in the sense that there are twice as many bars)? Does this mean that curve-fitting bias is reduced due to using more data, or is it increased due to the tendency to fit to the data and not a general market characteristic?

Unfortunately it is probably the amount of processing power that a trader has access to that determines which timeframe(s) they can trade, and so I would probably stick to 1H in general, but perhaps go down to the 30M for either a.) Pairs that have a high DMB on the 1H, or b.) Pairs for which it is difficult to find systems on the 1H.
  • #757
  • Nov 1, 2015 6:53am Nov 1, 2015 6:53am
Quoting algoTraderJo
{quote} One important thing IMHO is that you do things when backtesting exactly as you will do when live trading. I like F4 because you use the exact same code when doing both so it's impossible to introduce data snooping bias. By creating a dataset for backtesting you run the risk of falling into problems like this since your strategy is different for simulations and live trading. It's perfectly possible to do it if you're super careful but in my experience the less layers of potential bugs/problems you have the better.
I agree.. bugs can be very difficult to find and can introduce bias. In my case my code in live trading is the same as my backest. My vm machine is use for both.You also need to run your code in practice account during a couple of month/year for validation.
  • #758
  • Nov 1, 2015 7:03am Nov 1, 2015 7:03am
Quoting algoTraderJo
{quote} One important thing IMHO is that you do things when backtesting exactly as you will do when live trading. I like F4 because you use the exact same code when doing both so it's impossible to introduce data snooping bias. By creating a dataset for backtesting you run the risk of falling into problems like this since your strategy is different for simulations and live trading. It's perfectly possible to do it if you're super careful but in my experience the less layers of potential bugs/problems you have the better.
F4 is c++ style framework, i respect coder who use this but i prefer 4 generation language like python or ruby. With python pandas dont need any framework juste use build it function and avoid many errors with array manipulation. I also prefer to control everything in ma code instead of use framework. Framework hide many function call to made it simple, you can do many thing with that but never exactly what you want.

Here creating ohlc candle from tick, with A hour.

tick=pd.read_pickle(pair +'_tick') #(read tick from hard drive)
tick.index = tick.index + pd.offsets.Hour(6) #GMT +1 EST TO GMT + 1
day = ((tick['Ask']+tick['Bid'])/2).resample('24h',how='ohlc',base=A).dropna() #resample tick to ohlc

doing this in C++ could be very perilous
  • #759
  • Nov 1, 2015 8:10am Nov 1, 2015 8:10am
Thanks for your answer. Thinking about that I agree that likely technical analysis don't add values to a neural network. At the end of the day you are adding multiple view from the same input. About retraining every time, well I think agree with you. Surely it's not going to be easy.

Thanks again
  • #760
  • Nov 2, 2015 4:57am Nov 2, 2015 4:57am
Quoting beachboy44
{quote} I think that the 1H timeframe gives a good compromise between computation time/cost whilst still having the ability to take advantage of the daily volatility cycles in the Forex market (unlike 1D). I suppose that going down to the 30M TF may help to achieve a low DMB in pairs such as GBPUSD (as you said before), but I'm not sure whether it becomes easier to find systems on pairs such as AUDCAD or AUDUSD etc as we go further down. Do you think that simulation inaccuracies when using 30M are too prevalent to consider live trading, or does...
You can probably go down to 15M before any of these problems become too prevalent with the Asirikuy data. Providing your SL and TP (if any) values are large enough then you should not face any significant issues until you go to something like 5M, where it becomes hellish to develop something that works well across multiple brokers because of these problems. Also using more data in terms of "more bars" does not reduce curve fitting bias. The amount of market conditions inside your test depends on the number of ticks that were needed to construct the data, the tick is the fundamental piece of market information (not the number of bars, which are a construction from the tick data). So if you have 5 years of data in 1H format or in 30M format both give you to the same degree of curve fitting because the number of ticks needed to make both sets of data was the same. Your systems know how to work across the same market conditions, only viewed from two different stand points (1H and 30M).

Since you mention it, coming up with profitable and low DMB systems for the AUD containing symbols is haaaaard. Have you attempted this? What results have you obtained?
2 traders viewing now