AI Demand Forecasting & Planning

The problem

Master planning in Dynamics 365 Supply Chain Management does exactly what it is told. The problem is what it is told.

Most planning teams generate a baseline statistical forecast, then spend the following week arguing with it in Excel. The model does not know about the promotion starting in three weeks. It treats last year's supply outage as genuine low demand and forecasts the shortage forward. It gives one number per item per period with no indication of confidence, so a planner cannot tell the difference between a stable runner they should trust and an intermittent item where the forecast is close to a guess. Safety stock is a fixed number somebody set two years ago and nobody has revisited. And when a forecast turns out to be badly wrong, there is no way to ask why — so nothing is learned and the same error repeats next quarter.

The result is familiar: stockouts on the items that matter, cash tied up in the items that do not, and a planning team whose expertise is spent on data cleanup instead of decisions.

What ECOSIRE builds

ECOSIRE builds a forecasting and planning layer that sits alongside your existing F&O master planning rather than replacing it. Demand forecasts are generated by models we build and tune against your history, then written into F&O demand forecast lines so master planning, planned orders and the whole downstream supply process work exactly as your team already knows them. Nothing about how planners release planned orders changes. What changes is the quality of the number going in.

Forecast models fitted to your demand patterns

Different items behave differently, and one model cannot serve all of them. We segment your catalogue — fast-moving runners, seasonal lines, intermittent and slow-moving items, new products with no history — and fit appropriate methods to each: seasonal decomposition and time-series models for items with stable patterns, intermittent-demand methods for sparse items, and attribute-based analogue forecasting for new items that have no history of their own. Model selection is automated against a held-out backtest so the choice is evidence-based, not a preference.

History cleansing before modelling

A forecast fitted to unclean history is a forecast of your past problems. We build configurable history cleansing: outlier detection and treatment, stockout periods flagged so censored demand is not learned as low demand, one-off large orders separated from base demand, and promotional periods identified so their uplift is modelled rather than baked into the baseline. Every adjustment is visible and reversible — planners can see what was changed and why.

External demand signals

Where they genuinely improve accuracy for your business, external signals are brought into the model: promotional calendars, price change plans, customer forecasts and open sales agreements from F&O, weather where it drives demand, trading-day and holiday calendars, and web or storefront demand indicators. Each candidate signal is tested for whether it actually improves backtested accuracy before it goes into production, and signals that do not earn their place are dropped rather than kept for appearances.

Explainability

Every forecast comes with an explanation: the contribution of baseline level, trend, seasonality, and each external driver, plus the model chosen and why. A planner reviewing a number that looks wrong can see what is driving it in one screen. This is the difference between a planning team that adopts a forecast and one that overrides it by reflex.

Confidence intervals and safety stock optimisation

Forecasts are produced with prediction intervals, not just point values. Those intervals feed service-level-driven safety stock calculation by item, site and warehouse — so safety stock reflects actual demand variability and supply lead time variability at a target service level, instead of a number someone typed once. Results are written to item coverage settings, so master planning consumes them natively. Where you run advanced warehousing, coverage is calculated at the granularity your warehouse configuration actually uses.

Scenario comparison

Planners can create and compare scenarios side by side: a promotion at two different depths, a supply constraint on a key component, an optimistic and a conservative demand case. Each scenario can be run through master planning as a separate plan version so the inventory, capacity and purchasing consequences are visible before a commitment is made, and the chosen scenario promoted to the working forecast.

Forecast accuracy measurement

Accuracy is tracked continuously by item, product group, customer, site and planner, using measures agreed with you — typically MAPE, weighted MAPE and bias — with the baseline statistical forecast retained as a comparison so you can see whether the new models are actually earning their keep. Persistent bias by planner or by product group is surfaced, because a consistently over-forecast group is a fixable process problem, not a modelling one.

How it is delivered technically

The F&O-side components are an X++ extension model with no overlayering: forecast staging tables, the demand forecast write-back, planner review and approval forms, coverage settings update, and the accuracy reporting objects. Forecast generation runs on the standard batch framework with its own batch group and a schedule you control. Where model execution runs outside F&O, data moves through data entities and OData, and where you already have a Power Platform footprint we can surface planner review and scenario approval as a model-driven app instead of a second place to log in.

Who this is for

Supply chain and demand planning teams running F&O master planning who have outgrown a single baseline statistical forecast: businesses with seasonal or promotional demand, long or variable supply lead times, a mix of fast-moving and intermittent items, multiple sites and warehouses, or a regular new-product introduction cycle.

How delivery works

1. Scoping call. We review your current forecasting process, your item portfolio and demand patterns, how master planning is configured, what external signals you have access to, and which decisions you want the forecast to improve.

2. Fixed quote. You receive a written scope covering item segments, models, signals, screens and reports, with a fixed price before any development starts. The listed price reflects a defined scope; broader portfolios and additional external signal integrations change it and we say so before you commit.

3. Build. ECOSIRE builds the X++ extension model, the forecasting layer and the planner screens, and fits and backtests models against your own demand history. Lead time is typically one working week from signed scope, longer where several external data sources have to be integrated.

4. Install in test. We deploy the deployable package into your sandbox through your LCS pipeline and run a parallel period: our forecast alongside your existing one, measured on the same accuracy definitions, on your real data. You see the comparison before you commit to it.

5. Production. After sign-off, the package deploys to production through your normal release process, forecast batch jobs are scheduled, and the first live planning cycle runs with us available.

6. Support. A support window covers defect fixes, model behaviour questions and configuration help. Model retuning as your demand patterns change is quoted as ongoing work.

This is engineered against your data after a quote. There is no pre-built package and no free trial — a forecasting model that has not been fitted to your history would not be worth trialling.