Page 2 of 502 FirstFirst 12
Results 11 to 20 of 20

Thread: Price Analysis with Neural Networks

  1. #11
    Quote Originally Posted by ;
    Yes



    The ideal trade info you mentioned is really what I spoke about in my next article in this journal. What you are talking about is targeting the near future high and low for each input pattern. The buy/sell signal is really redundant, provided that you may calculate it by the saying (highgt;low)*2-1. After training on a large set of those patterns, but the neural net will learn to output the mean of those values over the set of similar patterns.
    You are right, it is the same thing you started with. That saying threw me for a minute. (hgt;nonetheless ) is Boolean, 1 or 0, got it. Anyway, as you stated the NN will just spit out the way so you've got to utilize a distribution.

    Quote Originally Posted by ;
    Instead, I would rather have the community learn the distribution. Knowing the past leaflet distribution permits you to create strategies based upon the calculated probability of hitting different price points. For example, we are utilized to the idea of accepting trades with a 2:1 or 3:1 TP to SL ratio, but we can not be sure that they'll really increase our winnings, as without a predictive edge, a nearer stop is generally more likely to be hit than the further take profit. If we knew the statistical distribution of price points during the next hour (or day, or week... whatever), we can set stops which, historically, are now less likely to be struck than our take profit level.

    Ideally, we'd have the ability to choose trades which are extremely high-probability and high-reward, where the end is not likely to be struck, the take profit is very likely to be hit, and also the take profit is significantly farther than the stop. Warren Buffet always attributes his success to creating low-risk, high-reward investments. Our aim ought to be the same - which is, to maximize the probability of winning while minimizing the probability of losing or in trader's conditions: to have an edge on the market.
    You are talking my language here. I have a closed thread called Expectancy is always zero where this is only one of those topics. A TP/SL of 3:1 means that win rate would be about 25 percent in a strategy with no edge. Generally, the mandatory gain rate (RWR) for almost any strategy is (Avg. Loss)/(Avg. Gain Avg. Loss).

    The thing is, I've tried to wrap my head around trading out of a supply in the past and had trouble envisioning the way to take action, especially in the case of a separate supply for highs and lows. Actually, I guess what you would really have is a distribution over all pairings of highs and lows, meaning over a place with high on the X axis and reduced on the Y for example. At any given point theoretically you would have a worth although I know you are going to bin them.

    Finally, in order to create a trading tool, there has to be some process for getting in the inputs at one end to the 3 big decisions (buy/sell, stop, TP) at the other. This may prove to be a far thornier problem than the structure and training of the net itself.

    I am bringing this up today as it's essential to know at this stage what you would like the output to look like. It's a pain to find something all built and then realize at the conclusion that you would like it to do something different. Hopefully you'll avoid that problem, and maybe you've already got it covered.

    I am out of here for the night, but looking forward to your updates. This thread will rule!

  2. #12
    Drummer,

    Awesome thread, man!

    You have already upped my comprehension about Neural Networks and programming them.

    I am supposed to do different things!

    . . .Damn you!!! . . .kidding. LOL

    I'm going to enjoy this!

  3. #13
    Quote Originally Posted by ;
    The thing is, I have tried to wrap my head around trading from a supply previously and had difficulty visualizing how to do it, particularly in the instance of a separate supply for highs and lows. Actually, I guess what you would have is a distribution over all pairings of highs and lows, meaning over a place with large on the X axis and reduced on the Y such as. At any given point theoretically you would have a value although I understand you are going to bin them.

    Finally, in order to create a trading application, there needs to be some process for getting from the inputs at the same end to the 3 large decisions (buy/sell, stop, TP) at the other. This may prove to be a far thornier problem than the construction and instruction of the internet itself.

    I'm bringing this up today as it's key to understand at this point what you would like the output to look like. It is a pain to find something all constructed and then realize at the conclusion that you would like it to do something different. Hopefully you will avoid that issue, and maybe you've already got it covered.
    I really have given a bit of thought to this issue. I understand from one of your PMs which you're interested in genetic algorithms. This might be a case where GAs might be rather helpful in determining good approaches to make use of the outputs of the neural network. On the other hand, there are quite a few other machine learning or classification techniques that could be applied to the problem of trading decision using the supply data.

    We may also have the ability to produce a trading strategy through simple theory, however. So long as we understand what the supply means, we can make intelligent use of it. Essentially, the supply will inform us how probable it is that a certain price point will be reached in a certain timeframe. We know that, to reach a high of 50 pips above the current price, price also (typically) must pass through the 10 pip, 20 pip, 30 pip, and 40 pip levels. So, knowing the approximate supply, we could sum the probabilities of more extreme levels to calculate the expected probability of hitting any amount during a future interval. We can then produce trade standards which choose only the highest-probability trades.

    We can also divide the take profit and stop loss orders into quite a few stratified orders across the supply according to the probabilities of different degrees. Additionally, this is a set up that could be calculated in order to facilitate the highest reward with the minimal risk.

    Anyway, this is all conjecture which will need to be reconsidered after the final network is finished. I believe I will start developing the community with the construction described, and when we realize later that the probability distribution is less useful than another target, we could alter them out fairly easily. The actual difficulty of programming is simply setting up and debugging the framework for training, testing, and viewing the output.

    It is going to be quite a while until I have something helpful developed, however I shall keep this thread updated as I make progress. Have a fantastic weekend, or whatever's left of it

  4. #14
    Very very good strategy! I am doing something similar.

    What do you think about integrating very large scale tendency to the equation by normalizing the bars.

    What I mean is say 10000 bars ago price was 3000 pips reduced. That means that on average we are going up 3 pips each 10 bars.

    Could this be useful if you normalize the OHLC bars to reflect that on your distribution equation?

  5. #15
    Quote Originally Posted by ;
    Really excellent approach! I am doing something similar.

    What do you consider incorporating very large scale tendency into the equation by normalizing the bars.

    What I mean is state 10000 bars past price was 3000 pips lower. Meaning that on average we're going up 3 pips per 10 bars.

    Could this be useful if you normalize the OHLC bars to signify that in your supply equation?
    Interesting idea. Without normalization, the supply would be skewed towards anything the long-term tendency is. This would be OK if we were in the exact same long-term trend, but not fine if the trend had changed lately. I suppose you could do some type of normlization, but it's also not clear exactly how you should normalize appropriately. It may be better to just include a set of moving averages on different phases on the inputs. This would help the system to separate the reason for the skew, and be in a position to replicate the right amount of distribution skew for the current market.

    Thank you, charlinks... good point!

  6. #16
    Quote Originally Posted by ;
    I understand from one of your PMs which you are interested in genetic algorithms.
    Genetic programs (GP) really, which was a theory designed after GA by a researcher named Koza I think. I am just nit-picking on a Sunday night. :

  7. #17
    Here is an indicator I created that may prove to be useful.

    It's a easy price distribution. Fiddling with the parameters could create quite different outcomes.

    Cell_width
    how many bars to use for your price distribution per column

    cell_rows
    how many rows to divide the price distribution into

    cell_cols
    how many columns to generate... for some heavy computations you can lessen this number if functionality suffers as you most likely need the recent price distribution

    cell_step
    the higher the number the more blocks are skipped for every step it is still calculating cell_width number of pubs. So if cell_step and cell_width are exactly the same you will find a crystal clear picture. If cell_width is higher you will observe data fuzzy in.

    Cut_off
    all cells are normalized with a value from 0 to 1 in the event that you draw all them it will be very resource intensive. So everything bigger than cut_off is not drawn. (I put 0.8 sometimes to identify powerful support/resistance areas)

    count_inside_bars
    count_support
    count_resistance
    and this is some magic... instead of counting the entire pub I've divided it into
    inside bar = ( Open - Close )
    support = ( Low - Min(Open, Close )
    resistance = ( Max(Open, Close) - High )

    That way you can concentrate your research only on support or only on resistance areas.

    Happy Neural Networking!!!
    https://www.cliqforex.com/attachment...3427100106.ex4

  8. #18
    Quote Originally Posted by ;
    Interesting idea. Without normalization, the distribution would be skewed towards whatever the long-term trend is. This would be fine if we were in the exact same long-term trend, but not fine if the trend had changed recently. I guess you could do some type of normlization, but it's also not clear exactly how you ought to normalize appropriately. It might be better to just incorporate a set of moving averages on different phases on the inputs. This would assist the system to separate the cause of the skew, and be in a position to replicate the right amount of distribution skew for the current market.

    Thanks,... good purpose!
    You know... it just hit me...

    All you have to do is correct the price for interest exchange rates!!!
    In fact some brokers do precisely that... they fix your entering price to signify the exchange so if you held GJ for 100 days you would have like a purchase price of 200 pips below your initial commerce in there...

  9. #19
    Quote Originally Posted by ;
    Genetic programs (GP) actually, which was a concept developed after GA with a researcher named Koza I believe. I'm only nit-picking on a Sunday night. :
    Yeah, I figure GP is a sensible application of GA theory. The machine learning novel I'm most familiar with (Machine Learning by Tom Mitchell) treats every type of learning algorithm as a search approach. So GP is your application of the GA search method to programming tasks. Taxonomy aside, employing a genetic algorithm search method to the issue of optimal trade choice could definitely create some decent solutions that we would never think of ourselves.

    Quote Originally Posted by ;
    This is a indicator I made that may turn out to be useful.

    It is a simple price supply. Fiddling with the parameters can create results that are quite different.
    ...

    This way you'll be able to focus your research only on support or only on resistance areas.
    Very trendy. I made something like this once that attracted a range of s/r lines using different colours based on the previous high/low distributions, but that is way more configurable. Well done!

  10. #20
    What are the inputs?

    Purely price? COT? Interest rates? CPI?

    Second, what is the criteria for entering/exiting? A majority wins predicated on NNs in timeframes?


    I'm quite educated on NNs, and if you introduce noncorrelated (data with no strong linear dependency to price) inputs you will introduce error into your learning. I bet you already knew that.

    So the question is, what inputs are strongly connected? I'd wager that Currency Market is very efficient, so new external data such as COT and CPI are likely fully realized from the price in a few minutes of this data being released, thus VERY compact lag and inefficiency.


    Nicely, NNs are not dinner table conversation. Either you get what I am saying or you do not. Hope it helps.

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •  
This website uses cookies
We use cookies to store session information to facilitate remembering your login information, to allow you to save website preferences, to personalise content and ads, to provide social media features and to analyse our traffic. We also share information about your use of our site with our social media, advertising and analytics partners. Read more.