Prediction relying on future information

[leobasilio]

Hi. I've been studying the prediction code for a while now and unfortunately there's a problem with it. The wavelet denoising effectiveness rely on future information. From what I get, given the price up to time t, we want to predict it at time t+5.

┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐
│  t  │ │ t+1 │ │ t+2 │ │ t+3 │ │ t+4 │ │ t+5 │
└─────┘ └─────┘ └─────┘ └─────┘ └─────┘ └─────┘

I'm actually in interested in trend prediction, but I believe that a good price predictor might also be a good trend predictor. And, in fact, when running this code, we a get a 79% accuracy in trend prediction.

However, when preprocessing the data, it is denoised all at once. In a real scenario, denoising would be "incrementally" done. I mean, it could only be applied until time t. So I've modified the code for the preprocessing to consider only up to the last candle of each window. After doing that, the accuracy drops to around 50%. No better than a coin tossing.

After some tests, I've realized that the more points I add after time t, the better the prediction. So if my windows end at time t but I use prices up to t+4 while denoising, I get back to the 79% accuracy. Now I'm 5 minutes away from the predicted price, but 20 minutes past the prediction's time of reference. It's too late to use that information for any sensible action.

I'm just point this out so it can be taken into account on future researches.

Thank you for your work.

[hereisreza]

Hi. I've been studying the prediction code for a while now and unfortunately there's a problem with it. The wavelet denoising effectiveness rely on future information. From what I get, given the price up to time t, we want to predict it at time t+5.

┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐
│  t  │ │ t+1 │ │ t+2 │ │ t+3 │ │ t+4 │ │ t+5 │
└─────┘ └─────┘ └─────┘ └─────┘ └─────┘ └─────┘

I'm actually in interested in trend prediction, but I believe that a good price predictor might also be a good trend predictor. And, in fact, when running this code, we a get a 79% accuracy in trend prediction.

However, when preprocessing the data, it is denoised all at once. In a real scenario, denoising would be "incrementally" done. I mean, it could only be applied until time t. So I've modified the code for the preprocessing to consider only up to the last candle of each window. After doing that, the accuracy drops to around 50%. No better than a coin tossing.

After some tests, I've realized that the more points I add after time t, the better the prediction. So if my windows end at time t but I use prices up to t+4 while denoising, I get back to the 79% accuracy. Now I'm 5 minutes away from the predicted price, but 20 minutes past the prediction's time of reference. It's too late to use that information for any sensible action.

I'm just point this out so it can be taken into account on future researches.

Thank you for your work.

Hi
Which part of the code has some information from the future???
I Don't see anything wrong in the Data_process.py file.
Thanks for your attention.

[leobasilio]

Hi. It's here:

https://github.com/mkhushi/code-with-paper/blob/main/Forex%20Price%20Prediction%20by%20Wavelet%20Denoised-ResNet%20CNN%20and%20LightGBM/pre_processing/DataProcess.py#L29

It's using all the data in the denoising process.

Let me try to exemplify. Suppose this is the whole dataset:

/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\

Denoising gives you this:

/~~~~~~~~~~~~~~~~~~~~~~~~~~~~\

If we take any data point, the denoised price at that point was calculated using points that came before and after the one we're considering. Notice that the edges are still noisy, because they lack neighbor data points.

When we're doing real time prediction, our latest data point is always at the edge.

/~~~~~~~~\ 
 /~~~~~~~~\ 
  /~~~~~~~~\ 
   /~~~~~~~~\ 
-------------------> time

The price we get at the edge after denoising is not the same value we would get if that point were in the middle of the window.

The right way to do this in my option would be to denoise every window individually, not the whole dataset at once. If we take that approach, the accuracy drops to 50%.

[hereisreza]

Hi. It's here:

https://github.com/mkhushi/code-with-paper/blob/main/Forex%20Price%20Prediction%20by%20Wavelet%20Denoised-ResNet%20CNN%20and%20LightGBM/pre_processing/DataProcess.py#L29

It's using all the data in the denoising process.

Let me try to exemplify. Suppose this is the whole dataset:

/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\

Denoising gives you this:

/~~~~~~~~~~~~~~~~~~~~~~~~~~~~\

If we take any data point, the denoised price at that point was calculated using points that came before and after the one we're considering. Notice that the edges are still noisy, because they lack neighbor data points.

When we're doing real time prediction, our latest data point is always at the edge.

/~~~~~~~~\ 
 /~~~~~~~~\ 
  /~~~~~~~~\ 
   /~~~~~~~~\ 
-------------------> time

The price we get at the edge after denoising is not the same value we would get if that point were in the middle of the window.

The right way to do this in my option would be to denoise every window individually, not the whole dataset at once. If we take that approach, the accuracy drops to 50%.

Yeah that's right.
Or we can use a denoising method that works with [t-n : t] for every index that is denoising;
Maybe This approach gives us more than 50% accuracy!
Thanks for your opinion.

[zainabtareen]

How to compute techinal indicator in case when denoising is performed on each window individually? should they be extracted from noisy data, since, computing them from a small window won't make much sense? what would you suggest @leobasilio

[leobasilio]

I couldn't find a way around this. Sorry. 😕