Concept card

Time-Series Foundation Models vs IBP Classical Forecasting

As of 2026-07-23

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 Time-Series Foundation Models vs IBP Classical Forecasting?

IBP's classical forecasting (Croston, Holt-Winters, ARIMA) breaks on three frequent SAP scenarios — new product introductions, low-volume intermittent spare parts, and promotion-driven volatility — exactly where zero-shot models like Chronos and TimeGPT are designed to fit.

What it is

Time-series foundation models are large neural networks pre-trained on diverse temporal datasets; they produce zero-shot or few-shot forecasts without dataset-specific training. The two reference models are TimeGPT (Nixtla, 2023 — transformer encoder-decoder trained on 100B+ time-series observations from financial, energy, web-traffic and retail domains) and Amazon Chronos (2024 — probabilistic forecasting using T5-class architecture, treating time-series values as tokenised language). The structural claim: a single pre-trained model generalises across domains well enough to beat classical univariate benchmarks on unseen series without any fine-tuning.

The structural SAP pain point: SAP Integrated Business Planning (IBP) for Demand ships with classical forecasting algorithms — Croston's method for intermittent demand, triple exponential smoothing (Holt-Winters), ARIMA, and ML Random Forest (added Q4 2022). These work well on high-volume SKUs with regular demand patterns. They fail on three SAP-frequent scenarios: new product introductions (NPI, no historical data), intermittent spare-parts demand at low volumes (Croston handles it, but accuracy degrades below 12 months of history), and demand volatility spikes from promotions that are encoded in external variables not available in IBP's native feature set.

Why it matters

Chronos or TimeGPT can generate a zero-shot demand signal for new products with no historical data, loaded into IBP via OData as an external forecast blended with the statistical baseline.
Chronos's probabilistic quantile output maps naturally to IBP's safety-stock formula, improving spare-parts MAPE by 1-3% over Croston on sparse series.
TimeGPT accepts exogenous regressors (promotion flag, price discount, weather) that IBP's own ETS can't incorporate without custom ML work.

Key points

TimeGPT (Nixtla) and Chronos (Amazon) are the two reference time-series foundation models as of 2026.
IBP classical algorithms (Croston, ETS, ARIMA) handle 80 % of regular-demand SKUs; foundation models address the 20 % tail (NPI, intermittent, promo-sensitive).
Chronos outputs full quantile distributions — maps natively to IBP's safety-stock formula.
Integration pattern: Chronos/TimeGPT endpoint on SAP BTP AI Core → OData write to IBP forecast override.
Classical IBP wins on interpretability, governance, zero GPU cost, and CSRD regulatory traceability.
Time-Series Foundation Models vs IBP Classical Forecasting is mastered only when it changes a named buyer decision.
Start with the semantic contract and control model before demonstrating the tool.
Use current SAP, analyst, study, KG, and news signals as evidence, not decoration.
Separate verified facts from directional trends and modeled assumptions.
Define owner, metric, threshold, support path, and rollback before scaling.

Terms used on this page

TimeGPT: Nixtla 2023 time-series foundation model. Transformer encoder-decoder trained on 100B+ observations from diverse domains. API-accessible; supports exogenous regressors.
Chronos: Amazon 2024 probabilistic time-series forecasting model. Uses T5-class architecture treating time-series values as discrete language tokens. Produces full quantile distributions, not point forecasts.
IBP: SAP Integrated Business Planning — S/4HANA cloud-native supply-chain planning suite. Demand module includes Croston, ETS/Holt-Winters, ARIMA, and ML Random Forest forecasting algorithms.
NPI: New Product Introduction — a product with zero or minimal historical demand data; the scenario where classical statistical forecasting has no signal to fit.
Zero-shot forecasting: Producing a forecast for a series the model has never seen, relying on patterns learned from other series during pre-training.
Decision owner: The accountable person who accepts the trade-off and funds the next action.
Semantic contract: The shared definition of business terms, metrics, entities, and access rules used by tools and teams.
Control plane: The layer that applies policy, access, lineage, monitoring, and escalation across the operating model.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks.

Open in the app →