You may ask yourself, ‘What is the Watson in Planning Analytics with Watson’? One place PA leverages Watson’s technology and ‘know-how’ is Workspace Forecasting.
Recently (PAW 74), Workspace forecasting moved from a custom service (used in CA and PA) to the larger Watson Time Series Library to perform the predictive forecasting calculations. The new Watson Library is the same library used in many other IBM Watson products; and provides a greater ability to support new user features in the future (like the recently released Outlier Detection).
However, the PA transition to the Watson Library has required a lot of recalibrations. It has many more ‘pulleys and levers’ which needed fine tuning. It’s superior, yet different, in the way it handles the modeling of historic data – and we are sorry if these differences created additional questions about consistency of results.
In the upcoming release (PAW 79), Workspace Forecasting has recalibrated its use of the library to better account for most of the client observed differences.
Specifically, adjustments have been made to;
- confidence envelopes that exceeded expected bounds,
- accuracy calculation (MASE/MAPE)
- short historical periods (training and testing splits) – generally, a factor of 3x is still recommended for seasonal forecasts (ie: 1 year forecast requires 3 years of historical data).
- the distribution of trend & seasonality contributions
- automatic determination of seasonality (without a manual entry)
If you have further questions, please ask me in this Community and I’ll try to answer