A forecast is an input to an inventory decision. The decision is the deliverable, and it depends on the error rather than the estimate.
Forecasting programmes are measured on accuracy and inventory policy is set from variability. Those are different quantities. A forecast that is seventy per cent accurate with a stable and well-understood error pattern lets you size safety stock properly. One that is eighty per cent accurate and erratic does not, because you cannot tell how much cover you need.
This table is the whole argument of the page in one place. None of the six is a forecast accuracy figure, and most planning functions hold the right-hand column rather than the middle one.
Closing the middle column is a measurement exercise. It is usually a few weeks of work and it changes the inventory position more than a better forecasting model does.
Inaccurate demand forecasts
The stated problem, and usually not the binding one. A plant with a mediocre forecast and a characterised error can hold less inventory than one with a good forecast and no error model.
Demand volatility and seasonal fluctuations
Seasonality is tractable when there is enough history. Volatility that is genuinely random is not reducible by modelling and has to be absorbed by inventory or capacity, which is a cost decision.
Excess inventory and working capital constraints
Excess is usually buffering against something: unreliable lead times, an untrusted forecast, or inaccurate stock records. Treating the symptom reduces the buffer and increases the stockouts.
Material shortages affecting production
Frequently present at the same time as excess, in the same plant. That combination points at policy and accuracy rather than at the forecast.
Inconsistent inventory planning policies
Parameters set at different times by different people, some for a demand pattern that no longer exists, and nobody owns the review.
Limited visibility into demand drivers
Which matters where demand is driven by something knowable: a customer programme, a build schedule, a promotion. Where it is, that driver is worth more than any statistical method.
Spreadsheet-based demand forecasting
The spreadsheet frequently contains real judgement that the historical data does not support. That judgement is worth extracting before it is replaced.
Difficulty managing slow-moving and obsolete inventory
Slow movers need a different method, a different policy and sometimes a decision to stop stocking. They are where the write-offs come from and where one-method-for-everything fails.
Supplier lead-time variability
Measured rather than contracted. This is one of the two largest inputs to safety stock and it is the one almost nobody has measured.
Limited coordination between demand, supply and inventory planning
Three functions working from three numbers. One agreed demand signal is a governance outcome rather than a modelling one.
Three stages in that order. Segmentation decides which items deserve which method. Error measurement decides the cover. Policy setting is where the value lands, and a programme that stops before it is the forecast nobody acted on.
Step 1
Demand and inventory planning process assessment
How the plan is made, who adjusts it, what the ordering system actually does with it, and which parameters were last reviewed when.
How the plan is made, who adjusts it, what the ordering system actually does with it, and which parameters were last reviewed when.
Step 2
Historical demand and inventory data analysis
Cleaned of the things that distort it: stockouts suppressing recorded demand, one-off orders, and a change of part number partway through.
Cleaned of the things that distort it: stockouts suppressing recorded demand, one-off orders, and a change of part number partway through.
Step 3
Demand pattern and seasonality identification
Per segment. Pattern detection on the whole catalogue at once averages away the structure that matters for the items that have it.
Per segment. Pattern detection on the whole catalogue at once averages away the structure that matters for the items that have it.
Step 4
Forecasting requirements and planning horizon definition
The horizon set by the longest lead time plus the review period, because forecasting beyond what you can act on produces numbers that look useful and are not.
The horizon set by the longest lead time plus the review period, because forecasting beyond what you can act on produces numbers that look useful and are not.
Step 5
Data preparation and feature engineering
Including the known drivers where they exist. A customer build schedule beats any statistical method and is frequently available and unused.
Including the known drivers where they exist. A customer build schedule beats any statistical method and is frequently available and unused.
Step 6
Statistical and machine learning model development
Method per segment. Statistical methods are the right answer for most manufacturing catalogues, and machine learning earns its place where there are genuine drivers and enough history.
Method per segment. Statistical methods are the right answer for most manufacturing catalogues, and machine learning earns its place where there are genuine drivers and enough history.
Step 7
Inventory optimisation model design
Safety stock, reorder point and review period per segment, derived from the measured error and lead time variability from section 2.
Safety stock, reorder point and review period per segment, derived from the measured error and lead time variability from section 2.
Step 8
Forecast validation and scenario testing
Validated on held-out periods, with the validation discipline described on the Predictive Analytics & Forecasting page, and tested against the stockout and excess consequences rather than only on accuracy.
Validated on held-out periods, with the validation discipline described on the Predictive Analytics & Forecasting page, and tested against the stockout and excess consequences rather than only on accuracy.
Step 9
Planning system and dashboard integration
The policy loaded into the system that places the orders, expressed in its own parameters. A recommendation that lives in a report is not a policy.
The policy loaded into the system that places the orders, expressed in its own parameters. A recommendation that lives in a report is not a policy.
Step 10
Continuous model monitoring and recalibration
With the policy parameters reviewed on a cadence and an owner, because demand patterns change and parameters do not review themselves.
With the policy parameters reviewed on a cadence and an owner, because demand patterns change and parameters do not review themselves.
Each is named with its precondition, because a model deployed without its precondition produces a number that planners will override and then ignore.
Statistical demand forecasting
Established time-series methods per segment. Needs enough clean history, and needs stockout-suppressed demand corrected first or it will forecast the shortage rather than the demand.
Machine learning-based demand prediction
Where real drivers exist and there is enough history to learn from. Needs a genuine driver. Without one it will reproduce a statistical method at greater cost and less transparency.
Demand segmentation and classification
The segmentation in section 3, maintained rather than produced once. Needs an owner, because items move between segments.
Item-level demand forecasting
Forecasts at the level the ordering decision is made. Needs the aggregation question answered: a forecast at item level and a plan at family level will disagree.
Safety stock optimisation
Cover sized from measured error and lead time variability. Needs an agreed service level per segment, which is a commercial decision rather than a parameter.
Reorder point and replenishment analytics
The trigger and the quantity, per item. Needs the review period and the ordering constraints: minimum order quantity, pack size, and the supplier’s own cycle.
Inventory policy optimisation
The whole policy set reviewed together. Needs the holding and stockout cost positions, which finance owns and which are frequently undocumented.
Lead-time variability analysis
Measured lead times with their distribution per supplier and item. Needs receipt history joined to order history, which is usually straightforward and rarely done.
Slow-moving and obsolete inventory analytics
Identification, ageing, and the recommendation to write down, use up or stop stocking. Needs a disposal decision owner, or the analysis produces a list nobody acts on.
Inventory scenario modelling
What a service level change, a lead time change or a supplier change does to the inventory position and the stockout risk. Often the most persuasive output on this page.
Demand and inventory planning dashboards
The planner view and the management view, on one model, with the policy parameters visible alongside the forecast.
The difficulty is stated for each. In this category it is almost always a data or a decision problem rather than a method problem.
Raw material demand forecasting
Derived from the production plan rather than forecast independently.
Finished goods demand prediction
Forecasting what customers will order.
Safety stock recommendation
Cover per item from measured variability and an agreed service level.
Supplier lead-time variability assessment
The distribution of actual lead times per supplier and item.
Item-level inventory optimisation
Policy parameters set per item within a segment.
Seasonal demand planning
Building the seasonal pattern into the plan and the cover.
Inventory shortage risk identification
Which items are at risk, and when.
Excess and obsolete inventory reduction
Identifying, ageing and recommending action.
Multi-location inventory allocation analysis
Where stock should sit across sites.
Demand and inventory what-if scenarios
The effect of a service level, lead time or supplier change.
In manufacturing catalogues, established statistical methods outperform more complex ones on most items, because most items do not have enough signal for a complex method to find. Complexity earns its place where there is a genuine driver and enough history, and it should be introduced there rather than everywhere.
Demand forecasting reference architecture
Data preparation, segmentation, model execution, policy derivation and delivery into the planning system, with the policy as the output rather than the forecast.
ERP, inventory and sales data integration
Demand, receipts, orders and stock history, with the receipt-to-order join that makes lead time measurement possible.
Time-series and statistical forecasting techniques
The workhorses, applied per segment, and transparent enough that a planner can see why a number moved.
Machine learning and predictive analytics
Where a driver exists. Introduced against a statistical benchmark, per the baseline discipline on the Predictive Analytics & Forecasting page.
Value, variability and intermittency segmentation
The segmentation from section 3, maintained as a product with an owner rather than produced as a one-off analysis.
Safety stock and reorder point models
Standard inventory models, applied with measured inputs. The arithmetic is not the hard part; the measured inputs are.
Optimisation algorithms and scenario simulation
For the multi-item, constrained cases and for the what-if scenarios that persuade finance.
Python and cloud analytics platforms
Chosen against your existing platform and what your team can maintain after handover.
Microsoft Fabric and Power BI
Named in the brief. We work with these where the client already runs them. The choice follows the estate and the operating model, and we state no partner or certification status.
Model monitoring and lifecycle management
Versioning, drift monitoring and a retraining trigger, plus a separate review cadence for the policy parameters, which drift independently of the model.
Governance defines who may change a model, a parameter or an assumption, how exceptions get resolved, and how the planning process continues when something fails. Everything below serves one of those three.
Model governance gets attention and parameter governance does not. A safety stock parameter set correctly two years ago for a lead time that has since doubled is a more likely cause of a stockout than any model problem, and nothing is monitoring it.
Inventory policy governance
Every policy parameter with an owner, a review cadence and a record of when it was last set and from what. A parameter nobody has reviewed is a decision nobody has made recently.
Demand data quality and completeness
Including the stockout censoring from section 5, part number changes, and one-off orders flagged rather than averaged in.
Forecast model ownership and accountability
A named owner per model family, responsible for its performance and for retiring it. A model with no owner persists and degrades.
Forecast assumptions and business rules
Recorded and visible with the output: what was assumed, what was adjusted and by whom. An adjusted forecast with no record of the adjustment cannot be learned from.
Model validation and backtesting
On held-out periods, with the validation discipline set out on the Predictive Analytics & Forecasting page. Validated against inventory and service consequences as well as against accuracy.
Forecast override and approval controls
Overriding is expected. The override is recorded with a reason from a short list, and the reasons are reviewed, because a reason used constantly is a missing input to the model.
Model explainability and transparency
A planner must be able to see why a number moved. An unexplainable forecast gets overridden, and once it is routinely overridden the model has no route to the business.
Model versioning and change management
Versioned with the record of which version produced which published forecast, so a past decision can be reconstructed.
Forecast bias and drift monitoring
Bias monitored separately from error, because persistent bias accumulates into inventory in one direction and a symmetric error measure conceals it.
Planning exception management
Exceptions with owners and ageing, and a route that reaches somebody who can authorise an expedite, a substitution or a customer conversation.
Either one alone can be improved by making the other worse, which is why neither is a target on its own. The pair, at an agreed service level, is the only honest statement of whether the programme worked.
Service level and fill rate
Against the agreed target per segment. The primary measure, paired with the one below.
Days of inventory on hand
Per segment and in total. The second half of the pair, and the figure finance will quote.
Forecast accuracy
Reported per segment rather than as one number, since a catalogue figure is dominated by the fast movers and says nothing about the tail.
Forecast bias
Reported separately and given equal weight, because bias is what turns into inventory or into shortage and an error measure hides its direction.
Weighted absolute percentage error
Weighted by value or volume so the measure reflects what matters commercially rather than treating every item equally.
Mean absolute error
In units, which is what inventory policy consumes. A percentage measure is for comparison; the absolute figure is what sizes the cover.
Stockout frequency
By item class and by cause: a forecast miss, a lead time miss, a policy parameter, or a stock record error. The cause split is what makes this actionable.
Inventory turnover
By segment. A finance measure that becomes operationally meaningful once the segments are defined.
Excess and obsolete inventory
Value and ageing, with the share that has a disposal decision pending, since identification without a decision owner produces a list.
Safety stock effectiveness
Whether the cover absorbed the variation it was sized for. If an item stocked out while holding its full safety stock, the inputs were wrong rather than the policy.
Model drift and forecast stability
Monitored against input distribution and output. The planning consequence of an unstable forecast is covered on the Predictive Analytics & Forecasting page.
More accurate and reliable demand forecasts on the segments where the data supports a model. On the intermittent tail, a better policy rather than a better forecast
Improved inventory planning decisions, most of the gain coming from measured lead time variability and an agreed service level rather than from the forecast
Reduced excess and obsolete inventory, which requires somebody to accept a write-down and will not happen without that authority
Lower material shortage risk at a stated service level. Lower risk at the same inventory requires better inputs, not a better model
Better working capital utilisation, reported as the service and inventory pair rather than as an inventory reduction on its own
Improved coordination between demand and supply planning, which follows from one agreed demand signal with an owner
If you hold too much inventory and still run short, a better forecast is unlikely to be the answer. Start with the measured lead time variability and the service level nobody has chosen.
CaliberFocus builds inventory policy from forecast error rather than forecast accuracy. We segment the catalogue before modelling, measure demand error and lead time variability, correct stockout-suppressed history, ask finance and operations to choose a service level per segment, and load a policy your ordering system can actually execute.
- Segmentation by value, variability, intermittency and lead time
- Safety stock sized from measured error and lead time variability
- A service level chosen per segment as a commercial decision
- Service and inventory reported together as a pair
Security & Compliance
