Vectorization

Hi,

I have a while loop that takes quite a long time to be calculated (especially when used in combination with ga optimization...I am not sure why) and that I would like to replace with some vectorization.

Here it goes:

 while i< end_iteration

    if G(i,3)/G(i-lookback_period,6) -1 > threshold_1
        s(i: i+ hold_length) = 1;
        i = i + hold_length;

    elseif  G(i,3)/G(i-lookback_period,6) -1 < -threshold_2
        s(i: i+ hold_length) = -1;
        i = i + hold_length;
    else
        s(i) = 0;
        i = i+1;
    end
end
s(1:hold_length)=0;

G is a 7 columns matrix filled with doubles, with around 1million rows. Part of the problem can be solved by doing:

lookback_vector = [zeros(lookback_period,size(G,2));G(lookback_period+1:end,:)];
s(G(:,3)/lookback_vector(:,6)-1 > threshold_1)=1;
s(G(:,3)/lookback_vector(:,6)-1 < -threshold_2)=-1;
s(1:hold_length)=0;

The problem I have then is how to fill with 1s and -1s the value from "i" to "i+hold", also because if say "i_1" < "i_2" <= "i_1 + hold" (with "i_1" and "i_2" being two indices satisfying the if or elseif conditions), then "i_2" shouldn't be a triggering condition.

Any help is appreciated.

Thanks

[Jan]

What is "hold"? It is recommended not to overwrite the names of built-in functions.

Did you pre-allocate "s" sufficiently? This is essential.

[Andrea]

Thanks for the reply. "hold" is a scalar with positive values. I take your advice and let's call it "hold_length" from now on. "s" has been preallocated to be zeros(end_iteration + hold_length, 1).

And sorry, I don't think your solution solves the problem. If s(index) has two consecutive 1s (or, as written above,"i_1" < "i_2" <= "i_1 + hold_length", with "i_1" and "i_2" being two indices satisfying the if or elseif conditions), it becomes a mess.

Answer 1

[Jan]

cumsum help for such cases usually:

index    = find((G(:,3) / lookback_vector(:,6)-1) > threshold_1);
s(index) = 1;
s(index + hold_length) = -1;
s = cumsum(s);

Care for overlapping sections.

[EDITED]: If there are overlapping sections, this might be helpful: FEX: mcolon.

[Andrea]

Sorry, I don't think your solution solves the problem. If s(index) has two consecutive 1s (or, as written above,"i_1" < "i_2" <= "i_1 + hold_length", with "i_1" and "i_2" being two indices satisfying the if or elseif conditions), it becomes a mess.

[Andrea]

Is there a way to go for a second pass and zeroing out all the "fake" triggers?

Something like s(index : index + hold_length-1) = 0 would work?

[Jan]

This was meant by "overlapping sections". See [EDITED]