Store result of three nested for loops

I have a matrix X consisting out of variables of 50 period (loop 1). I want to do a forecast for different time horizons: 2,4,6,8 periods (loop 2). Due to the class imbalance problem of X I want to repeat a subsample (the small group) several times (variable i)(loop 3). I want to store the period, the forecasting horizon, the number of repetitions and the accuracy.

i = 10:10:100;
for t = 1:max(period)-8
for z = 1:4 %lag order
lag_order = 2*z
h = t + lag_order;
for n = 1:numel(i); % repetitions of the a submatrix of X
accuracy
end
end
end

Maybe there is even a better way to do it with a matrix instead of a for loop.

[Stef]

accuracy is just a simple calculation of other variables included, but it varies depending on the different variables I gave you

[Stef]

It´s 200 lines of code. So basically in the third loop there is a support vector machine whichs accuracy is evaluated for each combination of the loop parameters

Answer 1

[Rik]

You could also use ndgrid to generate all combinations of loop parameters, and then use arrayfun to calculate your outputs.

[Stef]

Does it work like this?

lag = 2:2:8
period = 1:42
repeat = 10:10:100
grid = ndgrid(period, repeat, lag)

[Rik]

No, like this:

lag = 2:2:8;
period = 1:42;
repeat = 10:10:100;
[grid_period, grid_repeat, grid_lag]= ndgrid(period, repeat, lag);
[grid_sum,grid_product]=arrayfun(@example_function,...
    grid_period, grid_repeat, grid_lag);

function [out1,out2]=example_function(period,repeat,lag)
out1=period+repeat+lag;
out2=period*repeat*lag;
end

[Stef]

Thank you. If I make one big loop over the whole grid, what index does the loop then have?

for k = ?
...
end

[Rik]

You'll have to pre-allocate the outputs, and then use something like the code below. But why do that? arrayfun will either be a lot faster, or it will not make any difference.

lag = 2:2:8;
period = 1:42;
repeat = 10:10:100;
[grid_period, grid_repeat, grid_lag]= ...
    ndgrid(period, repeat, lag);
[grid_sum,grid_product]=arrayfun(@example_function,...
    grid_period, grid_repeat, grid_lag);

for k=1:numel(grid_period)
    period=grid_period(k);%optional
    repeat=grid_repeat(k);%optional
    lag=grid_lag(k);%optional

      %rest of your code

end

[Stef]

Because I have an SVM I want to test with different parameters. And I don´t know how to apply the arrayfun to such a complex procedure

[Rik]

You simply need to wrap your code in a function that takes a few inputs and returns a few outputs. The code complexity itself doesn't matter. My mini-example shows how to implement such a function. The example I used is of course very simple, but you can use the structure. Just make a function out of it.

[Stef]

Ok so the problem is that I have some variables depending on the grid like

X = [repmat(y grid_repeat(t),1)];

Then I have functions like fitcsvm that are included. So I think I can´t just make one line of arrayfun including 100 lines of code

[Stephen23]

@Stef: put the code into a function. Call that function in arrayfun. Just like Rik Wisselink showed.

There is no limit to how many lines of code your functions can have, whether local function, nested function, or main function ... all of these would allow you to have 100 lines or 10,000 lines of code (not that I would recommend doing that many in one function!).

[Stef]

Ok, say I put the whole for loop in an .m file. But in that file I have something like X(period==t) where t is any period. But sometimes I refer to period and sometimes I need the index of the lag order. I just don´t know how to bring that all in one file

[Rik]

That is just a question of inputs. What inputs does your function need? Does it need all three vectors and the current indices to them?

If memory is not an issue, you can wrap the vectors in a cell, and use repmat to duplicate the data. If that is too much overhead, you can set the vectors in a separate function. Both are shown below.

lag = 2:2:8;
period = 1:42;
repeat = 10:10:100;
[grid_period, grid_repeat, grid_lag]= ...
    ndgrid(period, repeat, lag);
grid_period_vector=repmat({period},size(grid_period));
%just use grid_period_vector as another input to your arrayfun function

Option 2: call the function below to set the three vectors. You can call this to make your grids, and inside your arrayfun function.

function [lag,period,repeat]=set_lag_period_repeat
persistent lag_ period_ repeat_
if isempty(lag_)
    lag_ = 2:2:8;
    period_ = 1:42;
    repeat_ = 10:10:100;
end
%persistent variables cannot be outputs themselves
[lag,period,repeat]=deal(lag_,period_,repeat_);
end

[Rik]

Did my suggestions help you? If so, please consider marking it as accepted answer. If not, feel free to comment with your remaining issues.