Hello, I was proposed a new challenge from my research process and the truth is that I have not been able to face it due to my lack of knowledge.

I have a time series that indicates the number of visits to a post, on a daily basis, the idea is to create an alert system that notifies me when the numbers of visits are above and below the historical number, I have thought of implement time series but I don’t know the process to evaluate and create alerts
visitors.xlsx (15.8 KB)
this is the base in case someone wants to help me

Hi @Jalvear , hopefully you are learning something the answers to all of your posts. What have you tried so far for this one?

And what kind of alerts are you thinking of here? There are few possibilities in Knime, but what do you have in mind?

Hello, at first I thought of taking the data as a normal distribution and generating alerts if your results were two deviations above the average, however the data is not normally distributed.

Researching the forums I found a flow of series that could help me, so I came across this graph that I liked but I don’t know if it suits what I need

I could deduce that if it breaks the confidence band it can be considered an alert, I don’t know if it’s wrong and the alerts just by sending an email would already be an alert for me

Hi @bruno29a I was talking to a professor who handles the subject better than me and he told me that he could look for a model that would best adapt in this case a SARIMA after that forecast one day and create a confidence interval and if the real value came out of that interval could generate an alert, is this possible with knime?

Hi @Jalvear

This is a very common problem in signal processing which can be solved in many different ways, e.g. using a Moving Average linear Model (or ARIMA, SARIMA, etc.).
I’m adding here below a different and very simple approach based on the -Moving Aggregation- node using two different configurations. One is based on Min-Max absolute values from beginning of signal. The second solution is using a sliding window of size 21 (which could be changed to what you want):

In other words, the first solution is telling you which are the historical Min-Max from the beginning of the signal history. The second solution only remembers the Min-Max historical value from the last 21 recorded samples of the signal.

Hope it helps.

Best

Ael

1 Like

Hello, I will put it to the test and I will tell you, thank you very much for the help

1 Like

Hello @Jalvear

Did you solve this question ? If so, what kind of solution did you implement ?

Best

Ael

I haven’t. However I think based on Scott’s idea, I could create a prediction interval and if the value comes from there it´ll create an alert.

Although this feature isn’t active yet. I’ve been told that this is going to be available in future updates.

When I get this feature, I’ll share with you the implementation of it. It has been an interesting challenge.

For context, Jalvear is referring to a separate discussion in this thread:

1 Like

I saw the thread you mention @ScottF and I guessed both threads were related. I’ll be glad to have feedback from you if you reach a solution to this problem based on SARIMA.