Hello,

Push notifications are useful when there's an algorithm continually checking the market for a certain event. However, MetaForecast version 1.6 does not have such an algorithm, but I consider it to be included in subsequent versions.

Thank you

😊👍👌😁

No, this is not a bug. When you receive an output like this without a fitted yellow line, it indicates that the model is unable to learn from the data. Please ensure that the degree input is greater than zero. However, it's important to note that in the current version, the neural network method does not support degrees greater than one. Any degree greater than one will be internally changed back to one.

Dear Vahidreza,

lets go step by step.

Step 1:

You say: " without a fitted yellow line, it indicates that the model is unable to learn from the data"

The point is that the yellow line is NOT missing. It is there but on the screenshot you can not see it because of the chart scale. Please check the next screenshot where i used an extreme chart scale to make everything visible. As you see on the screenshot below in the middle we have the "squeezed" price data (because of the chart scale) and you also see the pink forecast data with HUGE swings up and down and you can also see the blue regression line and the yellow line. So everything is there but the data is useless obviously.

Step 2:

You say:

So what is my mistake? I want to use large value for the past size and therefore i have set degree = 0 as recommended on the product page. I want to build a neural network based on the last 10 years on M15 TF. So i set degree = 0. Because as you say degree = 1 is for smaller past size! right? So i have set degree = 0

Step 3: Here is the log


Step 4:

Here are my settings




I am a very experienced user regarding Metatrader. Please tell me detailed how to create a usable neural network model with your indicator with degree = 0.  I am setting degree = 0 only because i am following your instructions:

Thank you very much! 👍😊

😊👍😁👌❤️

Please be so kind and answer one more question:

I do not understand the relationship between:

a) 

Past Size

Specifies the number of bars that MetaForecast uses to create a model for generating future predictions. The model is represented by a yellow line drawn over the selected bars.

and:

b)

Layers

This input configures a brain for your AI model. For example, if we enter [1000, 2000, 100], this indicates that the model accepts 1000 past bars as input, utilizes 2000 neurons to forecast 100 bars into the future. If we enter [1000, 2000, 500, 200, 100], this signifies that the model accepts 1000 past bars as input and employs a deep neural network with 3 layers containing 2000, 500, and 200 neurons, respectively, to forecast 100 bars into the future. As you can see, you can create a network with any desired depth. For the examples above, a minimum past size of 1000 is recommended. However, a past size of 3000 is generally better. As a rule of thumb, the larger the size of the layers in your model, the larger the past size should be.

For example if i set "Past Size" = 10.000 bars  and then i set "Layers" to for example:  [1000, 2000, 100]

Now the indicator will train based on 10.000 bars (past size) but the neural network only "accepts" the last 1000 past bars as input.

What does this really mean?

Will the training now be based on the last 10.000 bars (past size setting) OR: based on the last 1000 bars because of the layers setting  [1000, 2000, 100] ?

Thanks 

MetaForecast doesn't know the future of the last 1000 bars among the 10,000 selected bars. That's why it feeds the network with other chunks of data from these 10,000 bars, whose futures are known, to train your network.

Encountering these errors indicates that your layer sizes are excessively large for your current settings, leading to the generation of tensors that exceed MT5's capacity to handle.

Please note that upon encountering these errors, closing the chart and restarting your terminal is necessary to restore proper functionality.

Implementing neural networks within MT5 indicators presents several limitations. MT5's array indexing operates on 32-bit integers. Using large sizes alongside multiple symbols and the multi-modle feature result in index overflows or negative indexing, triggering system errors. For instance, the mentioned layer sizes could produce matrices as large as 32000 x 10000, potentially overwhelming your GPU that makes it to be not a monster any more. However, smaller models remain valuable. With MetaForecast, you can construct adequately big models relative to the available data for a given asset.

Using a past size smaller than the network's input size is acceptable because MetaForecast uses PastSize + InputSize amount of data, and it will inform you in case this amount of data is not available.

If you don't want to wait for the training process and want to break it, simply remove the indicator from your chart.

Subject: Request for Detailed Clarification on Data Utilization in MetaForecast

Dear Vahidreza,

I hope this message finds you well. I've been diligently trying to optimize the use of the MetaForecast indicator, particularly understanding how the training data is structured and utilized, especially in relation to the Past Size and Layers settings.

I have configured the indicator with a Past Size of 10,000 bars and a Layers setting of [1000, 2000, 100]. From our previous discussions, it is my understanding that MetaForecast uses Past Size plus Input Size amount of data, totaling 11,000 bars in this setup. Could you please clarify exactly how these bars are utilized for training the neural network?

To ensure clarity in our discussion, let's define each bar on the chart as follows:

• The current live bar (which has not yet closed) is labeled as shift = 0 .
• The most recently closed bar is shift = 1 .
• The bar before that is shift = 2 , and so on, increasing the shift count as we go further back in history.

Given this naming convention, could you specify which exact bars are used for training? For instance, does the training use data from shift = 1 to shift = 10,000 ?

Additionally, you mentioned that MetaForecast does not know the future of the last 1,000 bars. Could you specify which bars these are? Would they be from shift = 10,000 to shift = 11,000 ? How are these bars handled during the training process? Are they used as a form of out-of-sample testing to gauge the model’s predictive accuracy, or do they serve another purpose?

Understanding these details will greatly enhance my ability to use MetaForecast effectively and contribute a well-informed positive review. I am genuinely enthusiastic about the capabilities of your tool and eager to fully grasp its operational mechanics.

Thank you very much for your time and assistance. 😊👍

Best regards,

Lev