Machine learning in trading: theory, models, practice and algo-trading - page 1889

New comment
 
mytarmailS:

Alexei, do you remember how much acuracy we managed to squeeze out of the zigzag when we were raping one date set together?

0,71? or how much?

Forum on trading, automated trading systems and trading strategies testing

Machine Learning in Trading: Theory, Practice, Trading and Beyond

Aleksey Vyazmikin, 2020.05.17 18:45

If I increase the sample to a year (2019), I can see the trend for 2018 at least.


From merging of two samples we got Accuracy=68.06261099940409

And, from the graph, it appears that the trend of the new trees went sideways.


Decided to train separately on the two samples, since I had to prune them, as you said, the result:

Ваша Accuracy=65.46896295378517


My Accuracy=68.03370090447281



 
Aleksey Vyazmikin:

thank you!

I managed to get

 Accuracy 
0.82

))) The code is very messy, there is a probability of error.

In the meantime I hear cries that the markets are random ))

 
Petros Shatakhtsyan:

Don't make this up!

I never mentioned how the Sportlotto organizers regularly changed the balls, speed of the lotto machine, number of spins, etc., to rule out a pattern.

The same thing happens in forex and no MO helps here. Price movement is a random process and depends on human factor.

And the market behavior and currency pairs are always changing depending on market and political conditions. It is impossible to predict the direction of the movement.

The birth of man is also accidental, the whole world is a sports lotto
 

Now the f-i's are parsed with all the statistics in the cment. You can look into the f-i and immediately understand which cluster is responsible for what.

Further there is an idea to form at once .mqh file with f-files for all necessary clocks. And in general to simplify the subsequent reference to them

double decision_tree(double &features[]) {
/*
      cl 0 mean  cl 0 std  cl 1 mean  cl 1 std  cl 2 mean  cl 2 std
0      0.000007  0.000195   0.000739  0.000732  -0.000528  0.000344
1      0.000009  0.000187   0.000789  0.000693  -0.000538  0.000326
2      0.000013  0.000179   0.000810  0.000651  -0.000583  0.000344
3      0.000010  0.000190   0.000829  0.000645  -0.000630  0.000392
4      0.000017  0.000187   0.000824  0.000625  -0.000654  0.000403
5      0.000026  0.000195   0.000832  0.000651  -0.000688  0.000438
6      0.000025  0.000196   0.000839  0.000602  -0.000703  0.000437
7      0.000023  0.000199   0.000865  0.000615  -0.000717  0.000459
8      0.000016  0.000202   0.000905  0.000630  -0.000730  0.000469
9      0.000007  0.000215   0.000931  0.000659  -0.000738  0.000478
10     0.000006  0.000232   0.000928  0.000716  -0.000765  0.000480
11     0.000007  0.000255   0.000934  0.000782  -0.000783  0.000549
mean   0.000014  0.000203   0.000852  0.000667  -0.000671  0.000426
*/
    if ( features[10] <= 0.00035 )  {
        if ( features[10] <= -0.000315 )  {
            if ( features[11] <= -0.000385 )  {
                if ( features[8] <= -6.5 e-05 )  {
                    if ( features[9] <= -0.000595 )  {
                        if ( features[11] <= -0.00054 )  {
                            if ( features[6] <= -5 e-05 )  {
                                return 2; }
                            if ( features[6] > -5 e-05 )  {
                                return 0; } }
                        if ( features[11] > -0.00054 )  {
                            return 0; } }
                    if ( features[9] > -0.000595 )  {
                        return 2; } }
                if ( features[8] > -6.5 e-05 )  {
                    return 0; } }
            if ( features[11] > -0.000385 )  {
                if ( features[5] <= -0.000265 )  {
                    return 0; }
                if ( features[5] > -0.000265 )  {
                    if ( features[3] <= -1 e-05 )  {
 
Maxim Dmitrievsky:

What topic? Nothing works, we've already checked it.

I thought you were going to learn python. It's a good thing. But do not rely on profits from neural networks.

Damn Max, you just do not know how to cook them and yes they increase the chats by 80% at most. For BOs this is the trigger threshold or slightly higher. Literally 5% but even that will be enough..... Here's the catch. You can't just buy any watch, you have to get specific ones.... Secesh????
 
mytarmailS:

Thank you!

managed to get

from the zigzag, but the tssss... it's not exact )))) The code is wildly cramped, there's a chance of an error.

In the meantime you can hear the cries that the markets are random ))

If there is no error it's a good result - it may show reversals earlier.

 
Aleksey Vyazmikin:

If there is no error, then this is a good result - maybe it shows reversals earlier.

The reversals show excellent 0.8 on the zigzag is absolutely enough for earnings...

This is how forecasts look like on new data with acurasi 0,83

But we should simulate real trading, when a bar comes to the system every 5 min. I really hope that no errors will be found. :) otherwise it will be the same as always.)

 
Mihail Marchukajtes:
Man Max, you just don't know how to cook them and yes they raise the chats by 80% at most. For BO this is the trigger threshold or slightly higher. Literally 5% but even that will be enough..... Here's the catch. You can't just buy any watch, you have to get specific ones.... Secesh????

well mytarmails wrote you, he can do it.

learn python...
 
mytarmailS:

The reversals show perfectly 0.8 on the zigzag is absolutely enough to make money...

this is how forecasts look like on new data with acurasi 0,83

But we should simulate real trading, when one bar is fed to the system every 5 min. I really hope that no errors will be found. :) otherwise it will be the same as always ))

Doesn't it make sense that it is NOT possible to determine the lows and highs dynamically?

If it is even done a little further, with a delay, there is no guarantee that it is not a false reversal.

 
mytarmailS:

The reversals show excellent 0.8 on the zigzag is absolutely enough to make money...

this is how forecasts look like on new data with acurasi 0,83

But we should simulate real trading, when one bar is fed to the system every 5 min. I really hope that no errors will be found. :) otherwise it will be the same as always ))

It looks nice, have they invented new predictors?

1...188218831884188518861887188818891890189118921893189418951896...3699
New comment