ForecastWorx uses a pattern-aware, rules-driven, backtest-scored forecasting engine. It classifies demand patterns, narrows eligible algorithms, runs tournament scoring, and produces ensemble-weighted forecasts — all at the item-location level.
This feature is part of the ForecastWorx connected planning platform — an AI-native system where demand forecasting, inventory optimization, replenishment automation, and supply planning share the same data layer and work together in real time. Every capability described below is designed to reduce manual effort, improve decision quality, and deliver measurable business outcomes for planning teams across retail, wholesale, CPG, and manufacturing.
Perspectives
The ForecastWorx engine doesn't apply one model to everything. It classifies, narrows, tests, and ensembles — automatically.
The engine detects demand patterns in a strict hierarchy: data sufficiency → intermittency → declining → seasonality → trend → spiky overlay → stable vs erratic residual. This prevents misclassification.
Multiple algorithms (ARIMA, ETS, Prophet, Croston, TSB, gradient boosting, and more) compete through backtesting. Top performers are ensemble-weighted using inverse-error scoring, so no single model carries the entire forecast.
ForecastWorx backtests over the item's actual lead time using rolling origin cross-validation, not arbitrary windows. This evaluates models where accuracy matters most — the replenishment horizon.
Each eligible algorithm is scored using WMAPE across backtest folds. The tournament selects the best-fit models per item-location automatically.
Every forecast receives a governance score (0–100) measuring data quality, model fit, and volatility. Planners focus review on high-risk items, not every SKU.
Dampening and banding rules prevent period-to-period forecast whiplash, ensuring smooth signals for downstream inventory and replenishment systems.
Learns promotional response curves from historical data and integrates planned promotions into the baseline forecast.
Incorporate weather, economic indicators, and market data as regression features to improve forecast accuracy.
Every forecast comes with decomposition showing the contribution of trend, seasonality, promotions, and anomalies — so planners can validate with confidence.
Step 01
Historical demand is analyzed at the item-location grain. The engine classifies each series into one of seven primary patterns — new/sparse, intermittent, seasonal, seasonal+trend, trending, level/stable, or erratic — plus overlays like declining or spiky.
Step 02
Based on the detected pattern, the engine narrows which algorithms are eligible. Intermittent items get Croston/SBA methods. Seasonal items get Holt-Winters and seasonal decomposition. Trending items get regression models. No algorithm runs where it doesn't belong.
Step 03
Every eligible model is backtested across the item's lead time using rolling origin cross-validation. Models are scored by WMAPE. The tournament ranks them and selects top performers.
Step 04
In AutoML Ensemble mode, top-scoring models are combined using inverse-error weighting. Better backtest performers get more influence. The ensemble outperforms any single model.
Step 05
Final forecasts pass through stability controls (dampening, banding) to prevent whiplash. Each forecast receives a Forecast Risk Score for planner prioritization and governance.
ForecastWorx uses a pattern-aware, rules-driven engine. It first classifies each item-location's demand pattern (intermittent, seasonal, trending, etc.), then narrows eligible algorithms, runs backtesting over the item's lead time, and produces an ensemble-weighted forecast from the top performers.
Rather than forcing a single model, ForecastWorx runs multiple algorithms through a tournament. The best performers (by WMAPE) are ensemble-weighted using inverse-error scoring. This consistently outperforms any single-model approach.
Seven primary patterns: New/Sparse, Intermittent, Seasonal, Seasonal+Trend, Trending, Level/Stable, and Erratic. Plus overlays for Declining, Spiky/Lumpy, and High Uncertainty. The hierarchy prevents misclassification.
Every forecast gets an FRS from 0–100 measuring data quality, model fit, and volatility. High-risk items surface to planners for review, while low-risk items flow through automatically. This eliminates the need to manually review every SKU.
ForecastWorx uses lead-time-oriented backtesting with rolling origin cross-validation. Models are evaluated over the actual replenishment lead time, not arbitrary windows, ensuring accuracy is measured where it matters most.
No. ForecastWorx handles model selection, training, and evaluation automatically. Planners interact with the results, not the algorithms.
The ensemble includes ARIMA, ETS, Prophet, Croston, TSB, gradient boosting, and several proprietary models. The system automatically selects the best performer for each item-location.
Yes. Weather data, economic indicators, market indices, and custom external signals can be integrated as regression features to improve forecast accuracy.
Book a demo and see how ForecastWorx classifies demand, runs tournaments, and produces ensemble forecasts — all automatically.
"The most expensive software is shelfware. AI eliminates the excuse — your software should now work how you work, look how you want, and get used by your team."— ForecastWorx