getting significant decrease in a time series array

Hi there, I have time series array and I need to determine one thing: the starting index of a possible significant decrease. However, my problem is a little different from the others: For example

arr[100]=500 and arr[101]=100

is not a significant decrease, it should continue a while, lets say n times consecutively. For n=5 :

arr[100]=500
arr[101]=450
arr[102]=430
arr[103]=400
arr[104]=300

Finally, I need something like this:

C-like code:

int find_significant_decrease(*arr,n)
{
....
return idx;
}

How can I do that? Thanks in advance...

[Walter Roberson]

We can code for N consecutive decreases easily enough, but you have not given us enough information about what kind of decrease is significant enough, and you also have not indicated whether the decrease must be strict... For example if it went from 450 to 400 to 401 to 300 at the tail would that modest temporary increase (e.g. noise) be enough to block the trend?

[Emrah]

Actually, we may use mean and std.dev here (like in the box-plot) to determine a significant decreasing trend. And yes, I am looking for the decrease between two local peaks.

[Walter Roberson]

You need to tell us the formula. Mean and standard deviation of all of the prior points? Of all of the points in the complete time series? The points in a window? Once the mean and standard deviation are measured, how does that translate into determining whether points are a sufficient drop?

Once you have specific criteria, determining whether there are n points in a row is not difficult at all.

[the cyclist]

To add a bit to what Walter is saying ...

Before one starts to write code, it is best to have firmly in mind the algorithm that one is coding. You haven't specified the algorithm yet, and it sounds like it is not yet clear in your mind.

So, the first step is to decide on the algorithm. Then we can help you with the code.

[Image Analyst]

Why don't you eliminate a lot of uncertainty and simply attach a screenshot of your data with the sections of it that you want identified? Also attach a data file with the paper clip icon so that we can use your actual data. Otherwise this could go on for days and delay your solution.