Help: Time Series Predictive Model (Beginner)

Question

Hello!

I learnt to build this model from a YouTube video. I adjusted it according to my practice dataset. However, the predicting ability of both models is very bad. And the second TS Plot tool has an error which I couldn't to figure out how to fix. Can someone please help?

Thank you in advance!

S2 sales forecast model.yxmd

(S2)Dataset A_Digital Marketing_AY2425.xlsx

KGT · Answer

With that formula, I'm sure it would be [Row-1:Sum_Sales..], not [Row-1:Sub Category]. You can not add Timberland to $289.35.

I'm not sure why the action tool is there. It converts the model to a macro, which you don't want. You can delete it and go to Workflow Configuration, Workflow Tab to change back to a standard workflow.

With the TS Filler, it is designed to fill in gaps, not create the bulk of the data. On daily, you would have generated data for 90% of the data, which would then be not using your data for the model. By rolling it up to Month, you will have data in most options. Think about what the result should indicate. Does it really make sense that you think you can predict how many Timberlands are sold each day and is that data relevant as to whether the sale was on Tuesday or next Wednesday? And are those sales reported to this dataset? Even if you rolled up to weeks, there would be a lot of zeros, so monthly would be more ideal. The better your data, the better the prediction.

Add a formula before the summarise with DateTimeTrim([Order Date],"month"). Then summarise by that instead of Order Date etc. Looking at it now, it looks like the Time Series model tools do the aggregation to period anyhow, but by averaging to get that in-between data, we would be re-enforcing bias in the model.

For the TS Plot error, I think it might be because the field is a straight line forecast... I would worry about getting some variation in the predictions before worrying about the TS Plot. I'm not exactly sure what's wrong with the Arima, and don't have time to look now, but it's not taking Seasonality into account, I think...

The average price per pair ranges from $23 in some months, to $438 in 4 months. I imagine the price per pair is set and has promotions etc. So, trying to predict the sales amount is always going to be sketchy. Predicting Quantity would be better.

I'm not sure the instruction that you watched, and so my apologies if I'm making it into a real situation instead of a training sample. I would be looking to predict the quantity rather than the sales amount and on a monthly or quarterly grouping. I don't want to get you too far from the training material though.

I've attached your workflow with what I mentioned (I think I got it all, but again, typing, running, doing my job at the same time).

Final S2 sales forecast model - KGT.yxmd

Jenny_Vu · Answer

Hello KGT,

Thank you for your advise! How would you suggest to fill in the null values instead of the formula: IF IsNull([Sum_Sales (£)]) THEN [Row-1:Sub-Category] + [Row+1:Sum_Sales (£)]/2 ELSE [Sub-Category] ENDIF?

In the YouTube video, the instructor used this formula so I am not sure why the plus doesn't work here.

Also, do you mean that at the TS Filler tool, I should change the interval to month instead of day?

Here is the correct file for my final workflow. Sorry I mixed them up. Please give it a second look. Many thanks for your time and support!

Do you suggest to keep predicting sub-category sales by quarter or by month? And if so, do you think setting the interval at the TS Filler tool to week, month or year?

Also, the YouTube instructor included the Action tool but did not explain it. May I ask if it is useful in anyway here and how I can make it work?

Many many thanks!

Final S2 sales forecast model.yxmd

KGT · Answer

You may have uploaded this before saving. That workflow only has a TS Filler from the Time Series palette.

The formula in the Multi-row needs to remove part that is incorrect.

IF IsNull([Sum_Sales (£)]) THEN [Row-1:Sub-Category] + [Row+1:Sum_Sales (£)]/2 ELSE [Sub-Category] ENDIF

Also, as you have many many imputed values, the results are not going to be good. I advise rolling the data up to monthly and then trying it on that granularity.