Testing real-time forecasting systems - page 55

[Evgeniy Logunov]

grasn писал(а) >>

and the version hasn't changed :o(

I'll finish the article today and send it in for checking.

[Сергей]

lea >> :

I'll finish the article today and send it in for checking.

Great, we'll follow your progress :o)

[Stanislav Korotky]

grasn >> :

the forecast is more or less the same (the one with the maximum entropy) :o) A small refinement, the following trajectories remain, with the one that is "channeled" being the most likely.

The file would have to be dated (at least "verbally") ;-).

[Andrey Khatimlianskii]

lea >> :

A word of advice. Do not increase the array by 1 element in the loop. Keep the number of used (filled) elements in an additional variable and increase the array by a dozen or so elements (estimate for your own task) when you run out of space. This way you can get a significant performance gain.

I assert, speed gain is significant (I checked it personally).

Example-illustration of use:

double array[];
int size = 1000;
int step = 1000;
int cur = 0;

ArrayResize( array, size );

for ( int i = 0; i < unknown_big_value; i ++ )
{
   if ( some_condition )
   {
      array[ cur] = calculate_value( i );
      cur ++;

      if ( cur >= size )
      {
         size += step;
         ArrayResize( array, size );
      }
   }
}

ps: sorry for the off-top )

[Сергей]

to komposter, lea, Yurixx

Ok, it's silly to argue with MQL-masters. Although my estimations gave me a monopenial result, but I must have made a mistake somewhere, or the car is just strong, or I've used a smaller pitch. I think it would be better.

to marketeer

Файличик бы, с датой (хотя б "устно") ;-)

My bad, I forgot. Here's a new forecast with 300 samples (3 days since the post was published), M15, 17 base realizations in total:

Entropies calculated for them:

In the file is a matrix, each column is a realization, the column numbers correspond to the identifiers in the figure above. ie are the "winners", the realizations with maximum entropy:

PS: Ose, holiday, flippers, snorkel for 3 weeks. Good luck to all! You'll tell me later if the prediction came true... :о)

[Evgeniy Logunov]

grasn писал(а) >>

The calculated entropies are for them:

Could you tell us how you calculate entropy? Do you use any price transformations to calculate entropy or do you calculate directly for price series (forecast)?

I tried to count by box-counting - to calculate frequencies I divided the vertical axis into intervals, counted frequencies of price hits (not transformed), and then determined the entropy using the formula entropy=sum(i=1...n) -probability[i]*log2(probability[i]). But for some reason I didn't get such values of entropy, they were much smaller.

[Сергей]

lea >> :

Could you tell us how you calculate entropy? Do you use any price transformations to calculate entropy or do you count directly for price series (forecast)?

I tried to count by box-counting - to calculate frequencies I divided the vertical axis into intervals, counted frequencies of price hits (not transformed), and then determined the entropy using the formula entropy=sum(i=1...n) -probability[i]*log2(probability[i]). But for some reason I didn't get such values of entropy, they were much smaller.

while waiting for the plane ... :о) The formula is of course the same, but using specific price range conversions invented. Hence these values

[Evgeniy Logunov]

grasn писал(а) >>

while waiting for the plane ... :о) The formula is of course the same, but using specific price range conversions invented. Hence these values.

>> I see, thank you :)

[OO]

Here, everything seems to be in line with the prediction theme:

A and B - optimization on different periods

There are templates in the archive.

[Stanislav Korotky]

grasn >> :
Here's the new forecast still the same, for 300 counts (3 days since the post was published), M15, Total - 17 base realizations:

PS: That's it, holiday, flippers, snorkel for 3 weeks. Good luck everyone! You'll tell me later if the forecast came true... :о)

Some kind of a discrepancy turns out: forecast for 3 days and a holiday for 3 weeks ;-).