Inventory and demand forecasting that closes the loop to procurement

Most demand forecasts never reach the procurement system

Walk into a typical operations review and the demand forecast is a tab in a planning spreadsheet, updated monthly by a small team, presented in a deck, and then ignored by the procurement team because the buyer has been buying from this supplier for eleven years and knows what to order. The forecast is socialized; it is not operationalized.

The disconnect is structural. The forecast lives in one tool. The reorder points live in the ERP. The buyer's heuristic lives in the buyer's head. There is no machinery that translates a forecast into a purchase recommendation that the buyer sees in their daily queue. So the forecast informs strategic conversations and never touches operational decisions.

Probabilistic forecasts give you the inputs reorder math actually needs

Reorder-point calculation is statistical: reorder when projected demand during the lead-time window plus safety stock exceeds on-hand inventory minus pipeline. The math requires a distribution of expected demand, not a single point estimate. A point forecast of 1,000 units monthly with 200-unit standard deviation is a fundamentally different reorder calculation than a point forecast of 1,000 with 50-unit standard deviation.

We deploy probabilistic forecasts that produce demand distributions per SKU per location, updated daily as new sales and shipment data arrives. The reorder-point engine reads the distribution, computes the buffer required to hit the configured service level (typically 95% or 98%), and emits the reorder quantity and timing. The buyer receives a queue of recommendations with reasoning, not a deck of forecasts they have to interpret.

Stockout reduction

~38% first 12 months

Excess inventory reduction

22–30% on slow-moving SKUs

Forecast refresh

Daily distributions, per SKU per location

Buyer recommendation acceptance

~78% sent unmodified

Lead-time variability is the biggest hidden driver of safety stock

Reorder math typically assumes a fixed supplier lead time. Reality is that lead times vary by supplier, by season, by mode (ocean vs air), and by external shock — port congestion, weather, geopolitics. A safety stock calculated against a fixed 28-day lead time is wildly under-provisioned when actual lead times bounce between 21 and 49 days.

We model lead times as distributions per supplier per route, updated from actual receipt history. The safety stock calculation uses both the demand distribution and the lead-time distribution, and surfaces lead-time degradation before it causes a stockout. When a supplier's lead times drift wider, the buyer sees the trend in the recommendation queue along with a suggested action — qualify a backup, increase the buffer, or accept the risk.

ABC segmentation should drive forecast investment, not the other way around

Not every SKU deserves the same forecast investment. The top 5–10% of SKUs by margin and volume drive the majority of P&L and need sophisticated probabilistic forecasts. The middle tail benefits from simpler statistical methods. The long tail of slow-moving SKUs needs a 'reorder when stocked-out, plus a small buffer' rule that is honest about its imprecision.

The mistake is applying the same forecast methodology to every SKU. The result is excess investment in long-tail forecasting that doesn't move the needle and under-investment in the head SKUs that do. The ERP's segmentation drives the forecast model selection, not vice versa.

Stockout and excess risk should be visible before either happens

Operations leaders historically learn about stockouts from a customer escalation and learn about excess from a year-end inventory writedown. The ERP should surface both as forward-looking risk signals: 'SKU 3812 will hit zero on-hand in 9 days at current draw rate, current pipeline arrives in 13 days, mitigation suggested.' Same for excess: 'SKU 4127 has 240 days of cover at current demand, recommend price action or supplier deferral.'

These signals show up in the operations review, not in the year-end review. The conversation moves from 'why did we stock out' to 'how should we handle this risk.' The shift is structural: the system surfaces the risk, the operator decides the action, and the action is captured in the system as a documented decision.

Demand-sensing closes the gap between weekly forecasts and daily reality

Traditional forecasts update monthly or weekly. Real demand spikes and dips daily — promotions, weather, competitor actions, social-media moments, distribution-partner orders. Demand sensing layers a short-horizon model on top of the strategic forecast that incorporates yesterday's sales, in-flight orders, and external signals to adjust the next 7–14 days of expected demand.

The reorder engine reads the sensed demand for the lead-time window, not just the strategic forecast. The result is fewer surprises during promotional periods and tighter inventory through normal operations. Demand sensing is most valuable for fast-moving SKUs where the strategic forecast is most easily wrong about the next two weeks.

The buyer's queue is the last mile, and it's where most ERP projects fail

Sophisticated forecasts and reorder math fail when the buyer's queue isn't designed for the buyer's day. The recommendations have to surface in the buyer's primary tool — typically the ERP itself, sometimes a supplier portal — with the supporting reasoning visible without a click-through, and an interaction model that lets the buyer accept, edit, or reject in seconds.

We design the queue with the procurement team for two weeks before any model goes live. The acceptance rate of recommendations — what percentage are sent unmodified — is the metric that decides whether the closed loop is real. Below 60%, the recommendations are noise. Above 75%, the system is paying for itself in buyer time alone.

We had probabilistic forecasts for two years before we had reorder-point automation. The forecasts were beautiful. The procurement team didn't use them. Once the recommendations landed in the buyer's queue with the reasoning attached, the forecasts started actually shaping what we ordered. That was the year stockouts dropped 38%.

— VP Supply Chain, distribution client

Frequently asked

Why are probabilistic forecasts better than point forecasts for inventory? +

Reorder-point math requires a distribution of expected demand, not a single number. A forecast of 1,000 units with 200-unit standard deviation produces a fundamentally different safety stock than 1,000 with 50-unit standard deviation. Probabilistic forecasts give the math the inputs it actually needs and let the team set service-level targets — say 95% or 98% — and have the system compute the buffer that hits the target.

How does lead-time variability affect inventory levels? +

Materially. Most reorder math assumes fixed lead times, which are wrong in practice — supplier lead times bounce by season, mode, and external shock. We model lead times as distributions per supplier per route, updated from actual receipts, and use both demand and lead-time distributions in the safety stock calculation. This catches lead-time degradation before it causes a stockout.

Should every SKU get the same forecasting treatment? +

No. ABC segmentation drives forecast model selection. The top 5–10% by margin and volume need sophisticated probabilistic models. The middle tail benefits from simpler statistical methods. The long tail uses pragmatic 'reorder when stocked, plus a small buffer' rules. Applying the same methodology everywhere over-invests in long-tail SKUs that don't move the needle and under-invests in the head SKUs that do.

What does 'closed loop' between forecast and procurement mean? +

The forecast writes back to reorder points in the ERP. The reorder engine produces purchase recommendations with quantity, timing, and risk reasoning. The buyer sees them in their daily queue, reviews and accepts or edits, and the PO drafts generate automatically into supplier flow. Forecasts that live in a planning spreadsheet without reaching the procurement system are decorative; the loop has to close in the ERP.

What is demand sensing and when does it matter? +

Demand sensing layers a short-horizon model on the strategic forecast that incorporates yesterday's sales, in-flight orders, and external signals to adjust the next 7–14 days. It matters most for fast-moving SKUs where the strategic forecast is most easily wrong about the immediate window. For slow-moving SKUs, demand sensing adds noise without value. The reorder engine reads sensed demand for the lead-time window, not just the strategic forecast.

How do you measure whether the closed loop is working? +

Stockout rate (typically down ~38% in year one), excess inventory on slow-moving SKUs (down 22–30%), and buyer recommendation acceptance rate — what percent of system recommendations are sent unmodified. Below 60% acceptance, the recommendations are noise and the model needs work. Above 75%, the system is paying for itself in buyer time alone, before counting the inventory savings.