Perimattic
Analytics dashboard

AI Forecasting Engine

Ensemble AI that picks the right model for every SKU

Demand OS evaluates dozens of statistical and ML model classes per SKU, selects the best-performing method automatically, and logs every choice with its accuracy score.

Book a Demo

20–50%

Forecast error reduction vs statistical baseline

Source: McKinsey & Company

30+

Model classes evaluated per SKU before selection

Source: Perimattic product benchmark

Week 1

When planners see production-ready model-selected forecasts

Source: Perimattic implementation benchmark

How does demand planning AI software select the right forecasting model?

Demand OS evaluates each SKU against dozens of statistical, machine learning, and specialized model classes — including ARIMA, ETS, Holt-Winters, gradient boosting, and Croston's method for intermittent demand — and selects whichever produces the lowest error on held-out historical data. The selection and its accuracy score are logged for every SKU, so planners can inspect exactly why a given model was chosen. Models are re-evaluated on a configurable schedule, with out-of-cycle retraining triggered if accuracy degrades. The result is forecasting accuracy tuned per SKU instead of one method forced across an entire catalog.

The Problem

Why one model for every SKU always fails

Manual model selection across thousands of SKUs
Choosing a single forecasting method for 10,000 SKUs means every SKU with an unusual pattern — seasonal, promotional, intermittent — is fit with a model that was not designed for it. The inaccuracy is systematic, not random.
Accuracy that degrades silently between reviews
Models trained on last year's demand patterns drift as behavior changes. Without continuous accuracy monitoring, planning teams often discover a degraded model at month-end when the variance from actuals is already large.
Black-box selection with no audit trail
When a planner asks why the forecast changed, the answer should not be "the model decided." Without logged model selection and contributing factors, overrides happen by instinct rather than information.
Retraining cycles that run on a calendar, not on signal
Quarterly or annual model reviews are too slow for demand environments where patterns shift with promotions, product launches, or supply disruptions. By the time the review runs, the model has been wrong for months.

The Capability

How the forecasting engine selects and explains every model

FORECASTING CAPABILITY
Per-SKU model evaluation
The engine tests each SKU against statistical methods (ARIMA, ETS, Holt-Winters), machine learning approaches (gradient boosting, neural networks), and specialized methods (Croston's and Syntetos-Boylan for intermittent demand). The method that minimizes held-out error is selected.
A single model class applied uniformly ignores the demand pattern of each SKU — seasonal, promotional, or intermittent.
Learn more →
FORECASTING CAPABILITY
Demand-signal-aware model updating
Demand Sensing feeds near-term POS and order signals into the model evaluation loop, so the short-horizon forecast reflects current reality rather than the historical baseline alone.
A model trained on historical baseline alone misses what near-term POS and order signals are already showing.
Learn more →
FORECASTING CAPABILITY
Accuracy-driven safety stock calibration
The accuracy score produced by the forecasting engine feeds directly into Inventory Optimization's safety stock calculations. Higher forecast confidence means lower required safety stock buffer — and lower working capital.
Safety stock set with a flat rule ignores how confident the forecast actually is for a given SKU.
Learn more →

Benchmark Results

What forecasting teams see in the first quarter

20–50%

Reduction in forecast error vs statistical baseline across SKU portfolio

Source: McKinsey & Company

9%

Average reduction in manual override rate when planners can inspect model reasoning

Source: Perimattic product benchmark

34pts

Accuracy improvement on seasonal SKUs when moved from generic to per-SKU model selection

Source: Perimattic customer benchmark

In Practice

A food manufacturer cuts forecast error by 34 points on seasonal SKUs

The Situation

A food and beverage manufacturer with 2,400 SKUs was running a single exponential smoothing model across the entire catalog. Seasonal ranges, promotional products, and slow-moving long-tail items were all treated identically. Forecast accuracy on seasonal SKUs was 58% — well below the 75% planning target. The team was manually overriding 30% of all SKUs every Monday.

What Demand OS Did

The AI Forecasting Engine evaluated each of the 2,400 SKUs independently, testing ARIMA, ETS, Holt-Winters, gradient boosting, and Croston's method. Seasonal SKUs were assigned ETS or Holt-Winters based on historical pattern scoring. Promotional SKUs received gradient boosting models trained on event history. Slow-moving long-tail items were assigned Croston's method. All model selections were logged with accuracy scores visible to planners.

The Outcome

Forecast accuracy on seasonal SKUs improved to 92% in the first quarter — a 34-point improvement. The manual override rate dropped from 30% to 8%. The planning team stopped spending Monday mornings correcting the baseline and started reviewing the exception queue instead.

Related

Explore more Demand OS capabilities

Demand OS for Manufacturing

BOM-level disaggregation and intermittent spare-parts demand handling.

Learn more

Demand OS for Retail

Seasonal forecasting, promotional lift modeling, and store-level replenishment.

Learn more

Demand OS vs Blue Yonder

Faster implementation and explainable AI, compared to Blue Yonder's 12–24 month rollouts.

Learn more

Compare AI forecasting approaches across leading vendors in our guide to the 5 best AI-powered demand forecasting tools.

Frequently Asked Questions

AI forecasting questions answered

How does Demand OS decide which model to use for each SKU?+
The AI Forecasting Engine evaluates each SKU against multiple model classes — statistical methods including ARIMA, ETS, and Holt-Winters, machine learning approaches including gradient boosting and neural networks, and specialized methods including Croston's for intermittent demand. It selects the model or model ensemble that produces the lowest error on held-out historical data. The selection and its accuracy score are logged for every SKU so planners can inspect any choice at any time.
How does the engine handle new products with no historical data?+
New SKUs without sales history are handled through attribute-based similarity matching. The system identifies historical SKUs with similar attributes — category, price point, channel, seasonality profile — and uses their demand patterns as a starting baseline. Planners review the reference SKU selection before the new SKU enters live planning.
Can planners override model selections?+
Yes. Planners can review the selected model, the accuracy score, and the contributing factors for any SKU. Overrides are logged with the reason and planner attribution. If a model's accuracy subsequently improves beyond the override threshold, the system flags the SKU for review rather than reverting automatically.
How often are models retrained?+
Models are evaluated and retrained on a configurable schedule — weekly by default — incorporating new demand history. The system also triggers an out-of-cycle retraining alert if a running model's accuracy degrades beyond a configured threshold. All retraining events are logged in the audit trail.
Does the engine handle intermittent demand on spare parts and MRO items?+
Yes. The engine automatically identifies intermittent demand patterns and applies appropriate methods including Croston's method and Syntetos-Boylan approximation. The method selected for each SKU is visible to planners alongside its historical accuracy score. No manual configuration is required to route intermittent SKUs to the right model class.

Get Started

Ready to plan with confidence?

Talk to a demand planning specialist and see the AI forecasting engine live on your own SKUs. You leave with accuracy benchmarks, a scoped pilot plan, and published pricing — all in one call.

Book a Demo