Time Series Forecast - No Seasonality in Forecast

[HarryBrunning]

Hi All,

I am trying to predict the number of days holiday taken per employee per month, based on historic data (approx three years worth). I am trying to forecast six months forward, and would expect to see some seasonality in my forecast (more holiday in December than January for example). However, my model is only returning a linear forecast i.e. a flat number of holidays across the forecasted periods.

Is this forecast expected or have I done something wrong with the configuration of my workflow? Any help greatly appreciated.

Thanks

[lukas_1234567890]

I am a noob when it comes to Alteryx, but the Arima model is the same as the R ARIMA model.

You need to play around with your values of pdq and PDQ in order to get a good result.

See the attached screenshots as an example. I'm not saying these results are good, but you can see how playing around makes a difference. start with 1,1,1
Also it would help if you had more data, or if you could sample on weekly fre

[martinding]

Hi @HarryBrunning ,

I am not an expert in time series forecasting, but looking at the data and the decomposition plot, it seems that the dataset contains too much noise and the noise (random ups and downs) is making it really hard for the model to make future predictions. The forecasted number is simply the simple average across all month and seems like this is the best the model can do here.

I feel that the amount of gaps (imputed as 0s) are causing a big problem here! 16 zeros out of 46 observations, that's 35% of data missing.

You mentioned that you would expect to see seasonal effect (e.g. Dec being higher).

In the sample data, there are four observation for December: 2019-12-01: 0; 2020: 41.25; 2021: 30, 2022:30.

However, none of these Decembers were the highest in the year, and among the December month observations, there are big dispersions as well (ranging from 0 - 41.25).

Intuitively, if you had give this data to a human expert, he/she probably wouldn't be able to come up with a good prediction, simply due to the noise in the data.