Message Boards Message Boards


Optimal exponential smoothing: Why is NMinimize so slow?

Posted 9 years ago
4 Replies
3 Total Likes

Hello, I apply the exponential smoothing technique to some time series and would like to calculate the optimum smoothing constant (a) that minimize the in-sample sum-of-squared forecast errors with Mathematica. Unfortunately, this takes forever in Mathematica. For the exponential smoothing technique please refer to e.g.

Let's use some random GBM data

data = RandomFunction[
    GeometricBrownianMotionProcess[0.01, 0.1, 100], {1, 100, 1}][

Calculate simple returns

returns = Differences[data]/Drop[data, -1];

Apply the exponential smoothing technique (with constant a)

smooth = ExponentialMovingAverage[returns, a];

And use NMinimize to search for the a that minimizes the "forecast error":

NMinimize[Total[(Drop[returns, 1] - Drop[smooth, -1])^2], 

With my machine and Mathematica 10.0.1 this takes approximately 16 seconds, i.e. an eternity. If I do the same in R or Excel this takes a miniscule amount of time. I have the feeling that NMinimize is slow because it takes much time to evaluate the symbolic smooth expression. What am I doing wrong? Is there a way to speed things up? Any hints would be appreciated.

Thanks, Michael

POSTED BY: Michael Puhle
4 Replies

Define the objective so that it only exists for explicit numeric values of a. Can be done like this.

data = RandomFunction[
     GeometricBrownianMotionProcess[0.01, 0.1, 100], {1, 100, 1}][
returns = Differences[data]/Drop[data, -1]
smooth[a_] := ExponentialMovingAverage[returns, a]
ff[a_?NumberQ] := Total[(Drop[returns, 1] - Drop[smooth[a], -1])^2]

Now NMinimize is notably faster.

Timing[NMinimize[ff[a], a]]

Out[659]= {1.824000, {0.892557135365, {a -> 0.085192215858}}}

FindMinimum is faster still, by nearly two orders of magnitude.

Timing[FindMinimum[ff[a], {a, 0}]]

Out[660]= {0.040000, {0.892557135365, {a -> 0.0851922052551}}}
POSTED BY: Daniel Lichtblau
Posted 9 years ago

Thanks a lot Daniel! This seems to work like a charm (also in more difficult contexts).

POSTED BY: Michael Puhle
Posted 9 years ago

Thanks for looking into this Frank! I also thought about Simplify before, but unfortunately it doesn't scale. If you generate a path of length 1,000 then


takes forever (I aborted the evaluation). And a time series sample of size 1,000 isn't even large (approx. 1 day of 1 minute bars).... Do you have any other ideas?

POSTED BY: Michael Puhle
smooths = Simplify[smooth];

NMinimize[Total[(Drop[returns, 1] - Drop[smooths, -1])^2], a]

takes a little less than two seconds on my computer, compared to 15 seconds using smooth.

POSTED BY: Frank Kampas
Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Reply Preview
or Discard

Group Abstract Group Abstract