2

Let's say I have some data about the amounts of reports to Help Desk (about technical problems) which were monitored and registered every five minutes for the whole day (even during the night). Is there any way to efficiently calculate the "alarming" thresholds:

top - above this means there are too many reports in the system! Time to get to work! (most important). bottom - lower than this means there are too few of them (report system got broken?) The problems I face:

System is fairly new, so at the start of the registering process there were extreme outliers. I want to have thresholds for a more stable system in the future. People obviously don't work during the night, so there's a lot of observations (let's call them "moments" - those "five minutes") where the number of reports stays the same.

Ethan Bolker
  • 95,224
  • 7
  • 108
  • 199
Luxy
  • 21

2 Answers2

2

An "alarming" threshold indicates that something becomes "abnormally" high (or low). This could be interpreted in statistical terms as "the phenomenon under consideration exceeded its mean value plus its standard deviation" (or went under).

So you need to calculate sample means and standard deviations to begin with. Of what metrics? You can easily find all around the web the Key Performance Indicators (KPI's) for Support Call Centers -no need to re-invent the wheel. As I happen to have some professional experience in the field, I would especially pay close attention to how the following two metrics evolve : 1) evolution of backlog (the amount of pending/open/unresolved reports) 2) evolution of average time it takes to resolve and close a report.

But to calculate various statistics of the metrics you will eventually select, and do this "efficiently" and also not-misleadingly, you have to understand the special characteristics of your data (and study the corresponding areas of statistics). In your case, these special characteristics are : your data is time series, your data is high frequency, your data exhibits all sorts of seasonality (through the day, daily, weekly, monthly). These together create the following complex situation : your data depends on the immediate past (eg, backlog at beginning of each, say, hour, depends on what happened the previous hour (high frequency time series), but it also co-varies with what happened the same hour the previous day (high-frequency seasonality). You need to study these aspects to understand what kind of model you must set up in order for the various statistics like mean and standard deviation to eventually help you and not fool you.

Alecos Papadopoulos
  • 10,388
  • Thank for the answer. I know all of the terms/concepts/thing you listed, but you do think it's doable in Excel? Cause that's how I'm supposed to present the results. – Luxy Aug 10 '13 at 00:37
  • "Present in excel format" means that you will have to send to someone an electronic excel file? With the file containing the summary statistics and KPI's, as well as all the raw data from which the summary statistics were calculated? – Alecos Papadopoulos Aug 10 '13 at 00:42
  • Well, not exactly. I don't have to send it to anyone. I just have to explain how it was calculated and they are valid thresholds. – Luxy Aug 10 '13 at 10:20
  • But this has nothing to do with what software are you going to use to actually do the calculations. What you write is about having a convincing argument as to a) why, say, "mean + stdev" is a reasonable threshold and b) why taking, say, the sample mean, or rolling sample mean, or "seasonal" sample mean is a reliable estimation of the true mean (stdev etc). – Alecos Papadopoulos Aug 10 '13 at 18:09
-1

You should look into Statistical Process Control it uses values derived from a baseline period to determine whether a system is stable (experiencing normal variation) or if there is something that needs to be investigated.

Nick
  • 167