Menu
Sister Sites

Options

Bookmark Thread

First Page First Unread Last Page Last Post

Print Thread

Similar Threads

Neural Networks: Myths And Reality 19 replies

Extreme scalping using neural networks - EA/Manual 646 replies

Cognitive indicators and neural-networks 19 replies

Neural Networks in Trading (Java, Algodeal, Technical Analysis) 3 replies

Price Analysis with Neural Networks 41 replies

Attachments: Any success with neural networks in trading
Exit Attachments

Any success with neural networks in trading

Quoting synicz
{quote} Thats kind of a tall order dont you think? My opinion is that it is possible to teach a nn to either mirror your trading method. But to optimize it takes some creativity. Irregardless, the first step is to define, and then quantify the signals/situation whereby you will enter/exit a trade. How to do that is the golden question aint it? Thats where the creativity aspect comes in.
Right now, it's a nearly impossible task to develop a truly intelligent AI in a "not closed system" like trading.
In my personal opinion... the whole idea of forex trading is to 'predict future using the past data'. There are probably just two mechanical systems to automate that:

  1. Human Idea[captured from the screen pattern] -> Indicator[parameters] + fix trading rules + Fresh Data = trading decision
  2. Past Data + Algorithm[based on some idea] -> Model[hyperparameters] + Fresh Data = trading decision [via predicted probability]

The first approach relies on finding suitable parameters via optimization [bruteforcing many trades scenarios to find the best] and testing to confirm. Second approach can use for example neural network model. In a latter case CPU is not 'bruteforcing' trades execution but follows the algorithm rules trying to find parameters of the model. Technically use of NN models is more efficient as modern algorithms can use all CPU cores. As apposed to MT4 platform that can only use 1 CPU. At the same time training NN is more challenging due to reproducibility, overfitting issues, etc.

In short these both approaches are both possible, although for me, a hobby-researcher trader, use of NN is preferred as it is allowing to select core idea and automate it. There is much more freedom and it's totally possible to design AI trading system which would show what it is doing so the trader could understand it's behavior and start trust it .

Quoting vladi1979
In my personal opinion... the whole idea of forex trading is to 'predict future using the past data'. There are probably just two mechanical systems to automate that: Human Idea[captured from the screen pattern] -> Indicator[parameters] + fix trading rules + Fresh Data = trading decision Past Data + Algorithm[based on some idea] -> Model[hyperparameters] + Fresh Data = trading decision [via predicted probability] The first approach relies on finding suitable parameters via optimization [bruteforcing many trades scenarios to find the best] and testing...
Exactly, brutforcing the market to detect dominant patterns (not the usual sh*t) or algorithms.
It seems here is a good place to post my experiment with neural networks in my EAs.

Have made a simple EA based on open prices (https://www.forexfactory.com/showthr...0#post12630390) - to use open prices modeling in MT5 Strategy Tester.

Initial balance of 1000 USD. Leverage 1:500.

Step 1: First optimize without a neural network - to find out the best setup for years 2000 - 2009 for USDJPY. Best result: recovery factor 6.80, drawdown 417 USD, profit 2842 USD

Step 2: Then simulate the "real world" and run Strategy Tester backtest with the Step1 setup for years 2010 - 2019. Result: recovery factor 1.98, drawdown 563 USD profit of 1116 USD

Here we can make the first observation that our test is according to this statement: "Backtest and when go to real-world, be ready that drawdown can be 2x deeper and profits will be 2x less". Here we see this is true. (but in real "real life" we will re-optimize EA at least once every year)

Ok, now implement the neural network based on Relative Strenght Index indicator previous N (from 3 to 10) bars values (that's input for neural) - to filter out "bad deals". We now turn on the neural option in Strategy Tester and remove testing checkboxes for all the rest Step1 optimized EA parameters - so we will just add additional logic and find out if the neural network can help to improve results found in Step1.

Step3: Optimize with neural network years 2000 - 2009. Best result: recovery factor 12.96, drawdown 190 USD, profit 2468 USD

Step4: Simulate the "real world" now with neural network added for years 2010 - 2019: recovery factor 3.51, drawdown 253 USD, profit 889 USD

See attached backtesting Excels for each step. And MT5 EA to play with.

Further steps: need to emulate neural re-training each year (2010, 2011, 2012 etc). Then we need to do all this to 8 major currencies pairs and some top indexes. And we need to optimize each weekday separately too. This way we can emulate\test "real world" trading\re-training process and in case of positive results go with real money. Yes or no?
Attached File(s)
File Type: ex5 CandlesTrend_V08_OC_neural.ex5   95 KB | 192 downloads
File Type: xlsx USDJPY_2010_2019_all_days_recovery_3.51_dd_253_profit_889_neural.xlsx   123 KB | 154 downloads
File Type: xlsx USDJPY_2000_2009_all_days_recovery_12.96_dd_190_profit_2468_neural.xlsx   110 KB | 114 downloads
File Type: xlsx USDJPY_2010_2019_all_days_recovery_1.98_dd_563_profit_1116.xlsx   319 KB | 122 downloads
File Type: xlsx USDJPY_2000_2009_all_days_recovery_6.80_dd_417_profit_2842.xlsx   282 KB | 122 downloads
Neural networks are one of the most popular classes of algorithms for machine learning. In financial analysis they are most often used for forecasting, creating custom indicators, algorithmic trading and risk modeling. Despite all this, the reputation of neural networks is damaged, because the results of their use can be called unstable. One of the main problems why a neural network may not work is that data is often poorly prepared before it is loaded into the system. Normalization, removal of redundant information, sharply deviating values should be done before working with the network to improve its capabilities. Even if you have properly configured your neural network, and it trades successfully in and out of the sample, it doesn't mean that after a while it won't stop working. This isn't about neural networks, this is about how the financial market behaves. Financial markets are complex adaptive systems. What works today may not work tomorrow. This characteristic of them is called non-stationarity or dynamic optimization. Neural networks aren't yet able to cope with it. So neural networks are complex in their application and this significantly limits their possibilities.
Quoting vladi1979
In my personal opinion... the whole idea of forex trading is to 'predict future using the past data'. There are probably just two mechanical systems to automate that: Human Idea[captured from the screen pattern] -> Indicator[parameters] + fix trading rules + Fresh Data = trading decision Past Data + Algorithm[based on some idea] -> Model[hyperparameters] + Fresh Data = trading decision [via predicted probability] The first approach relies on finding suitable parameters via optimization [bruteforcing many trades scenarios to find the best] and testing...
There's also:
Dynamic past data -> No idea or algorithm -> All models on average = trading decision (against the models' predicted probablity)
I describe this theory a little: https://www.forexfactory.com/thread/...twork-strategy

For neural networks, GPUs (not CPUs) are faster by a huge factor.
Quoting joyny
It seems here is a good place to post my experiment with neural networks in my EAs. Have made a simple EA based on open prices (https://www.forexfactory.com/showthr...0#post12630390) - to use open prices modeling in MT5 Strategy Tester. Initial balance of 1000 USD. Leverage 1:500. Step 1: First optimize without a neural network - to find out the best setup for years 2000 - 2009 for USDJPY. Best result: recovery factor 6.80, drawdown 417 USD, profit 2842 USD Step 2: Then simulate the "real world" and run Strategy Tester backtest...
Joyny, what's the lastest neural EA version? This one from 2019?
Quoting costos
{quote} Joyny, what's the lastest neural EA version? This one from 2019?
yes
Quoting joyny
{quote} yes
How did further development go?

I wanted to test this but I get Unsupported filling mode errors. Is this source availabale so I can alter it?


Thanks!
  • Post #30
  • Quote
  • Last Post: Jan 7, 2022 11:59am Jan 7, 2022 11:59am
There are controversial opinions concerning neural networks in trading as far as I concerned, so you have to be really careful with them. Neural networks do not make any forecasts. Instead, they analyze price data and uncover opportunities. Using a neural network, you can make a trade decision based on thoroughly examined data, which is not necessarily the case when using traditional technical analysis methods. Well, if you understand all the risks and you are ready to rely on the data which is created due to neural networks, then of course you can try, but it's tough to find reliable software.
0 traders viewing now