View RSS Feed

mql5

Two forex trading newbies talking with each other

Rate this Entry
by , 08-19-2021 at 05:57 PM (595 Views)
      
   
Two traders are talking (taken from this mql5 thread):

Consider a linear regression model xi = a + b * i + ei in time i = 1, 2, ..., n, where the errors ei are white noise with the Laplace distribution. The error density then has the form p (x, c) = 0.5 * c * exp (-c * | x |), log (p (x, c)) = log (0.5) + log (c) -c * | x |

The likelihood function for the noise will have the form L = p (d1, c) * p (d2, c) * ... * p (dn, c), where di = xi-ab * i are the residuals of the model. Logarithm of the likelihood function LL = n * log (0.5) + n * log (c) -c * S, where S = | d1 | + | d2 | + ... + | dn |. S does not depend on the parameter c, therefore the problem of maximizing LL is solved in two stages
and it is the reply:

This is all true. The question is what exactly to take for the sliding between the two rows. For example, there is a traditional opinion that the length of the perpendicular to the regression line. But it seems to me that this is not quite the right way. For it gives a separation not relative to the previous values, but relative to a certain midpoint of them. Such a substance as the "asymmetry" of the opening is lost, and I would like to feel it.
And what do you think?
  • minimization of S (since c> 0) with respect to a and b ?
    or
  • maximization of LL with respect to the parameter c, with the found value of S ?

Submit "Two forex trading newbies talking with each other" to Google Submit "Two forex trading newbies talking with each other" to del.icio.us Submit "Two forex trading newbies talking with each other" to Digg Submit "Two forex trading newbies talking with each other" to reddit

Comments