The eight forecasting model in v8.1 are all exponential smoothing models. There is no model available for forecasting such as the linear regression model used with Trend Lines to weigh data the same. Also, Tableau does NOT allow you to force smoothing parameter values to zero to maximize smoothing and minimize the extra weight of recent periods. Instead, Tableau uses the parameter values which minimize the error of the model.
With regard to forecasting a nonlinear process, v8.1 now supports multiplicative models which do exactly that. If the optimal parameterization of a multiplicative model produces the lowest Akaike Information Criteria of all models and does not have excessively large relative residuals, Tableau automatically selects that model. If the model you want is not selected by default, you can force it by Choosing "Custom" in the "Forecast Model" section of the "Forecast Options" dialog and then choosing a Multiplicative trend or season or both trend and season.
If you send me your data or a packaged workbook, I might be able to make further recommendations to improve your forecast.
If you are still available to reply to this post can i offer to post a sample workbook that I have on pricing forecast and you can use your expertise to improve it? I would really like to know the difference in big scale forecast models and a tableau forecast which uses exponential smoothing using time series data.
I'd be happy to take a look at your forecast or answer questions about differences between exponential smoothing and other types of forecast. Of course, you can learn even more by reading the work of Prof. Rob Hyndman on which I depended heavily while implementing forecasting for Tableau. See his online book here: https://www.otexts.org/fpp
Thank you tons for the link, this will help me understand the details behind a forecasting model without going into mathematical details and formulas. Really appreciate your guidance in this matter, attached you will find the forecast workbook with sample data. I have tried using the forecast model which as you know uses ES method and used multiplicative model for trend and season. As I trying to be more aggressive with the forecast and multiplicative works best. Any notes or tips on making this better would be great help. When looking at forecast description it gives me a 'GOOD' indicator which i dont know is really good forecast or the data is good for the forecast.
But also is it possible to look at the forecast model after a specific period has passed and look at the actual and forecast. This will give an idea on how different or how accurate the forecast was. I don't know if its possible to see actual numbers for example April 2014 on May 1st and say my april numbers turned out to be X$ but my forecast was Y$.
let me know what you think of the model.
Forecast.twbx 13.4 KB
Your forecast is really good already. I categorize a "Good" forecast as one with a mean absolute scaled error (MASE) of less than 0.4. Your models are well below this threshold with MASE values indicating your forecasts are likely to have about a quarter the error of a naive (tomorrow will be the same as today) forecast. The reason automatic model selection chose MAM models rather than your choice of MMM is that the AIC is slightly better for the MAM models. But, the difference is small and either choice should work well. However, keep in mind that automatic mode can select different models for different time series in your worksheet so that each has the highest quality (lowest AIC) model. When you use a custom model, all time series are forecast with the model you choose. In your case, this is not a problem because the same models work well for both time series.
With regard to saving and comparing your past forecasts with future results, Tableau doesn't yet make this easy.
You will need to export your forecasts and then join your saved forecast datasource with a datasource containing your actual data in the future to make comparisons.
If you want to have more flexibility in your forecast models, you can always use the new R Integration feature to access the huge number of models available in R. In particular, you could use Prof. Hyndman's forecast package which has more types of exponential smoothing models and ARIMA as well.
Bora Beran wrote a nice explanation of how to forecast using the R integration feature here: Using R forecasting packages from Tableau « Bora Beran
good luck, Scott
Thank you so much for your expertise and validation of the model. Too bad I cannot rate this post, but it was really awesome of you to share your knowledge.
I will try some basic modelling in R once i get the localhost set up.
Have a good one,
My pleasure. Hearing about real world use of this feature will help me improve it.