Concept card

SAC Smart Predict — Classification, Regression, Time Series

As of 2026-09-25

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is SAC Smart Predict?

Smart Predict is SAC's no-code machine-learning tool across three scenario types — classification, regression, time series — trained and read through KI (predictive power) and KR (prediction confidence) indicators. SAP documents that on HANA data it delegates modeling to APL (C344), the same automated engine behind Smart Discovery and Predictive Planning, not PAL and not a SAC-proprietary algorithm.

The three scenario types

SAP Analytics Cloud's Smart Predict trains machine-learning models directly inside the SAC UI, without writing code, across three scenario types: Classification (predict a categorical outcome, such as which customers will churn), Regression (predict a numeric value, such as expected deal size), and Time Series / Forecast (predict future values of a metric from its own history). Each scenario type follows the same overall workflow — pick a dataset and a target variable, let Smart Predict train candidate models, review a training report with quality indicators, then apply the chosen model to new data to generate predictions written back into a dataset a story or a planning model can consume.

Reading the quality indicators

For classification and regression, SAC's training report centers on two indicators consultants should be able to name precisely in front of a client: KI (Predictive Power), which measures how well the model's predictions rank against a hypothetical perfect model, and KR (Prediction Confidence), which measures how stable the model's quality is expected to be when applied to new, unseen data rather than the training sample. A model can show strong predictive power and weak confidence — a sign it may be overfit to the training data — and that combination is exactly the kind of nuance a business stakeholder needs explained, not just a single accuracy percentage. Time-series training reports carry their own quality view, centered on how well the model reconstructs the historical pattern and how wide its forecast confidence interval is going forward.

Why it matters

SAP documents that Smart Predict delegates to APL (C344) on HANA data — the correct answer to 'what trains this model' is APL's automated functions, not a proprietary SAC algorithm and not PAL, and getting this right shapes how a consultant explains the tool's ceiling and its bridge to direct SQL/hana-ml work.
KI (predictive power) and KR (prediction confidence) can diverge — strong power with weak confidence signals overfitting — and being able to explain that combination precisely is what separates a consultant from someone reading a single accuracy number off a screen.
Smart Predict, Smart Insights and Smart Discovery are classical ML/statistics, explicitly distinct from Just Ask's generative layer (C335, C342) even though SAP documents them in the same SAC help chapter — conflating the two misrepresents both the tool and the roadmap.

Key points

Three scenario types: Classification, Regression, Time Series (Forecast) — no code, trained inside the SAC UI.
Classification/regression quality indicators: KI (Predictive Power) and KR (Prediction Confidence) — strong KI with weak KR signals overfitting.
Time-series training reports evaluate historical-pattern reconstruction and forward confidence-interval width.
SAP documents that Smart Predict delegates to APL (C344), not PAL (C343), when the data source is HANA — the same engine behind Smart Discovery and Predictive Planning (C342, C340).
Time-series forecasting needs enough historical periods to detect seasonality reliably — commonly at least two full cycles; check row count and time span before proposing a pilot.
Smart Predict, Smart Insights and Smart Discovery are classical ML/statistics — SAP documents this explicitly as distinct from Joule's generative Just Ask layer (C335, C342), despite sharing the same help chapter.
Escalation point: multi-feature interactions, custom architectures, or large-scale real-time scoring belong in Databricks-grade ML or a custom SAP AI Core build, not a stretched Smart Predict scenario.
A trained scenario applies to new data and writes predictions back into a consumable dataset for a story or planning model.

Terms used on this page

KI (Predictive Power): Smart Predict's indicator of how well a classification/regression model's predictions rank versus a hypothetical perfect model.
KR (Prediction Confidence): Smart Predict's indicator of how stable a model's quality is expected to be on new, unseen data versus the training sample.
Overfitting: A model that fits training data closely but generalizes poorly — flagged in Smart Predict by strong KI paired with weak KR.
Time-series seasonality bar: The minimum historical span (commonly ~2 full cycles) needed for Smart Predict's forecast to detect a seasonal pattern reliably.
APL engine: The Automated Predictive Library (C344) that SAP documents as the actual modeling engine behind Smart Predict on HANA data.

Sources

Full card available to members. What the full card adds: the full decision framework · the common pitfalls and their fix · the cheat sheet · the facts worth quoting.

Open in the app →