Does Forecaster have options to use various methods, how many options are available, and can users modify the values (top/bottom/mid)?
Answer:
Is Forecaster a black box system, or is it possible to tune/define parameters of algorithms?
Answer: Forecaster comes with built-in SHAP-based explainability for granular transparency. The model training process is not open for outside tuning today, but user-defined parameter tuning is under consideration as a roadmap item.
Does Forecaster self-learn, or have the ability to explain why its own forecast was inaccurate?
Answer: Forecasts are generated with a 4-view explainability framework. Each view provides a different angle on what impacted predictions, the dynamics between features, down to a single-prediction breakdown using precise values (not relative impact).
How much historical data can Forecaster handle (e.g., 2–4 years), and does it depend on storage?
Answer: Forecaster supports large datasets and up to 100,000 SKUs (forecast items) in each forecast run, with potentially very long history (up to 10 years). Forecaster reads and writes data directly from the Anaplan platform. In the near future, the Anaplan Data Orchestrator (ADO) will also be added as an additional read/write data source.
Can users update their numbers without requiring them to check the box? Is it dynamic that way?
Answer: Yes. Checking the box is one way to perform the override—it helps identify which value was overridden and keeps both values (previous and overridden) on record. Alternatively, users can choose to edit directly in the main forecast value cell without checking the box.
What is the product lineage between PlanIQ and Forecaster—is Forecaster a next-gen replacement, and what differentiates their predictive capabilities? Is the infrastructure still backed by AWS?
Answer: PlanIQ is no longer available; it was sunset about a year ago. Forecaster is the next-generation replacement and is fully native—the algorithms are owned and operated directly by Anaplan, and no data leaves the Anaplan environment.
What is the minimum historical data (in years) expected to use Forecaster?
Answer: It depends on the method selected. Forecaster has methods specifically designed to generate good forecasts even with very limited history, such as TimesFM (foundation model) and DeepAR (probabilistic neural network).
What is the computational overhead of the forecasting logic in terms of peak calculation engine runtime, formula dependency chain latency, and workspace memory utilization?
Answer: Forecaster runs on its own dedicated computation architecture. Training a model and generating a forecast does not lock the Anaplan model. Writing data into Anaplan does initiate a standard model write lock, with duration depending on the volume of data being written.
Can we fine-tune algorithm hyperparameters or import custom ML models, or are we limited to out-of-the-box algorithms and automated parameter optimization?
Answer: Currently, hyperparameter tuning and importing custom ML models are not supported, and users utilize the out-of-the-box algorithms. However, hyperparameter tuning and importing custom models are both on the roadmap.
How does Forecaster handle new product introductions (NPI) without historical data?
Answer: Several methods in Forecaster are designed to work in low-history scenarios, including TimesFM (pre-trained foundation model) and DeepAR (neural networks). Forecasting with no history at all (zero-shot forecasting) is not fully supported today, but is on the roadmap.
