Flat Line forecast using ARIMA

[akaur58]

I have data of one year and I want to forecast for next 3 months with weekly periods. I used ARIMA Model to the forecast because it showed less error compared to ETS. However, I am getting a flat line with cone shaped confidence intervals. The possible reason I can think of is insufficient data to see trend and no seasonality. I have attached the time series plot and auto correlation plot. Is there any way I can forecast this data. (The historic data is weekly)

[CharlieS]

Hi @akaur58 

I'm sure the tools have the capability to configure a model that better fits your needs, but what you're seeing now is just the default configuration. The default configuration of the ARIMA tool is to start with a simple naive model (https://otexts.com/fpp2/simple-methods.html), which the analyst can begin customizing for seasonality, drift, and other dimensions. The second tab of the ARIMA tool configuration is where you'll find many options to customize the model estimation to suit your particular scenario. 

I highly recommend the forecasting material linked in the Alteryx support page on Time Series: 
https://otexts.com/fpp2/index.html

[akaur58]

Thank you for the help 

[Garabujo7]

Hello @akaur58 , mi recommendation before working with time series is: investigate the data.

You should check the ranges of the data (lower and greater value), average and std deviation to make sure you don't have crazy ranges.

You could check this article for some help:

https://community.alteryx.com/t5/Engine-Works/Exploratory-Data-Analysis-in-Alteryx/ba-p/961861

You may have outliers as well, those could negatively influence your results.

The yellow squares are outliers in the data.

Gabriel

[leozhang2work]

The spikes are causing you trouble, you need to think about them. From your ACF you can see there is some pattern, then maybe you can do some differencing (I in the ARIMA).If the pattern is too erratic, then the default flat line may give you the best results.

[akaur58]

Isn't differencing is for non stationary time series but my time series doesn't have any trend or seasonality so i guessed it is stationary

[leozhang2work]

Without seeing your data I wouldn't know. Probably share the decomposition plot.

Given the ACF is above the dotted line, there is something to be done, my guess is you need to try different q value (hopefully I get it the right way round).

[martinding]

To debug / tune a model, it is generally good practice to test out potential factors one by one. In your case, I think it comes down to two factors, 1) the data, 2) the model.

We can further breakdown the problem into finer factors:

The data:

1. Not enough data

2. Quality is bad, e.g. outliers

The: Model:

1. Too naive

2. Not the right model

So I'd suggest you investigate these one by one, for example:

The data:

1. Not enough data: why not simulate some additional data - > 1 more year's worth, but slightly randomized, see if the model is giving you more sensible results.

2. Bad quality data: drop your outliers or do some appropriate cleaning, and then see if the model improves.

Remember to check these factors one at a time, so you can understand which factor is causing you the problem.

Hope this helps and good luck!