How to Build a Demand Forecasting System (Playbook)

Demand forecasting is one of the oldest applications of statistics in business and one of the most common places where machine learning disappoints, not because the models are poor but because they never connect to decisions. This playbook covers building a demand forecasting system that does: defined decisions, clean data with drivers, probabilistic models, honest backtesting, planner integration, and monitoring. It follows FISTA's AI enablement practice; context is in ai demand forecasting and the AI for supply chain resilience whitepaper.

What does the system do?

Component	Function
Decision definition	Which decisions, at which level, over which horizon
Data pipeline	History, drivers, hierarchy, calendar, cleaning
Models	Probabilistic forecasts across the hierarchy
Reconciliation	Coherent forecasts across levels
Backtesting	Time-based evaluation against baselines and the current process
Integration	Forecasts into planning tools with override capture
Monitoring	Accuracy, bias, drift, forecast value added

Step 1: Define the decisions

Forecasts serve decisions: replenishment, production scheduling, purchasing, staffing, promotions. For each, define the level (SKU by store, product family by region), the horizon, the frequency, and the cost of over- and under-forecasting. These define the hierarchy, the loss function, and the evaluation. See how to write acceptance criteria for ai.

Step 2: Assemble history and drivers

Collect sales or demand history (correcting for stockouts, which censor demand), prices, promotions, marketing, calendar and holidays, weather where relevant, and product attributes. Build the hierarchy and handle product transitions and substitutions. Clean outliers deliberately, distinguishing data errors from real events. Data practice is in how to prepare data for ai.

Step 3: Train probabilistic models

Establish statistical baselines per series. Then train global models, gradient-boosted models on engineered features or neural forecasting models, that learn across series and incorporate drivers. Produce quantile or distributional forecasts, because inventory policy needs the distribution, not the mean. Handle new products with attribute similarity and launch curves. Model practice is in how to build a predictive model.

Step 4: Reconcile across the hierarchy

Forecasts at SKU, store, region, and total levels must be coherent. Apply hierarchical reconciliation so that aggregations match and information from stable aggregate levels improves noisy detail levels. Planners lose trust quickly when levels disagree.

Step 5: Backtest honestly

Evaluate with rolling-origin time splits that mimic production: train on data up to a point, forecast the horizon, roll forward. Compare against naive and seasonal baselines and against the current process's historical forecasts. Report error by level, horizon, and product velocity; measure bias separately from accuracy; and evaluate distributional forecasts with quantile loss and coverage. Method is in the AI evaluation and testing whitepaper.

Step 6: Integrate with planning and capture overrides

Deliver forecasts into the planning system planners already use, with uncertainty bands and driver explanations. Capture every planner override with its reason. Overrides are legitimate where planners know things the model cannot, and harmful where they encode bias; the data shows which. Integration patterns are in how to build an erp ai integration.

Step 7: Measure forecast value added

Compare each stage of the process against the naive baseline: statistical forecast, model forecast, planner override, consensus adjustment. Forecast value added shows which stages improve accuracy and which degrade it, and it directs effort: improving the model, coaching planners, or removing a consensus step that adds error. Measurement design connects to the AI ROI measurement framework whitepaper.

Step 8: Monitor and adapt

Monitor accuracy and bias by segment, forecast value added, driver data freshness, and drift in demand patterns. Retrain on schedule and on drift triggers; after disruptions, when history stops predicting, widen uncertainty and increase planner review. See what is model drift and the AI observability whitepaper.

How do forecasts connect to inventory decisions?

Distributional forecasts feed inventory policy: safety stock from the demand distribution and lead-time uncertainty, reorder points, and allocation. Connecting the forecasting system to optimization is where the value is realized; see ai inventory optimization and how to build a dynamic pricing engine for the price-demand interaction.

Where do language models help?

At the edges: extracting promotion and event information from unstructured plans and communications into driver data, summarizing forecast changes and their causes for planners, and answering planner questions about forecast drivers. Core forecasting is a time-series modeling problem.

What does it cost to run?

Cost is dominated by data engineering and integration; training and inference at scale are moderate and predictable. Value is measured in service level, inventory, expedite cost, and planner time against baseline. Drivers are in predictive analytics cost.

What are the common mistakes?

• Forecasting at the wrong level for the decision.
• Point forecasts feeding safety-stock formulas that need distributions.
• Ignoring stockout censoring and promotions.
• Random train-test splits that leak the future.
• Forecasts in a separate tool planners do not use.
• Never measuring whether overrides help.

Worked example: a multi-store retailer

A retailer forecasts weekly demand by SKU and store to drive replenishment, with a thirteen-week horizon for purchasing. History is corrected for stockouts using availability data; drivers include promotions, prices, local events, and weather. A global gradient-boosted model with quantile outputs replaces per-SKU statistical models for high-volume items, while statistical baselines remain for sparse series, and forecasts are reconciled across store, region, and total levels. Rolling-origin backtests show improvement over the planners' historical forecasts at SKU-store level and larger gains during promotions. Forecasts flow into the replenishment system with uncertainty bands, planner overrides are captured with reasons, and forecast value added shows overrides help on new products and hurt on stable items, which shapes planner guidance and the next model iteration.

How FISTA Solutions builds forecasting systems

FISTA Solutions builds demand forecasting systems to this playbook: decision-defined specifications, driver-rich data pipelines, probabilistic global models with hierarchical reconciliation, rolling-origin backtesting against the current process, planner-tool integration with override capture, forecast value added measurement, and drift monitoring. The AI enablement practice delivers the platform, AI agents handle exception workflows built on the forecasts, and forward deployed engineers embed with your planning team. The record behind the work is 150+ projects with 47% average efficiency gains.

To scope a demand forecasting system, message FISTA on WhatsApp, or read ai in supply chain for the sector context.