Predictive analytics and forecasting for MedTech, built on an honest account of what you can observe about demand and what you are inferring.
Consumable demand follows procedure volume. Procedure volume follows referral patterns, clinical practice, reimbursement and capacity at hospitals you do not operate. Service demand follows device utilization you can only measure on the connected portion of a fleet whose size is itself contested. The forecasting problem in a device company is rarely that the method is wrong. It is that the quantities the method depends on are not directly observable and nobody states which ones are estimates.
Capital equipment demand is a small number of large, slow, individually negotiated decisions driven by hospital budget cycles, which is a pipeline question rather than a statistical one. Consumable demand is high volume, driven by procedures, derived from an installed base and genuinely forecastable. Treating them with one method produces a capital forecast that is a sales opinion dressed in statistics and a consumable forecast that could have been much better.
The Installed Base Is the Shared Dependency
Service, consumable and capacity forecasts depend on an installed base that different functions count differently.
Demand Is Derived, Not Direct
Consumable demand follows procedures at hospitals the manufacturer cannot directly observe.
Capital Sales Are Few and Lumpy
Large equipment purchases depend on hospital budgets and sales pipelines, making statistical forecasting less useful.
Reliable Equipment Produces Few Failures to Learn From
Predictive maintenance needs failure history, which reliable devices rarely provide in sufficient quantities.
Some Predictions Cannot Be Acted On Immediately
Recommendations involving validated processes, product disposition or regulated activities may require approvals that outlast the forecast horizon.
Forecasts Need Continuous Evaluation
Without comparing predictions against actual outcomes, the same forecasting biases can persist for years.
Forecast improvement programmes default to better models, and in device companies the larger gains usually sit in the quantities the model consumes. An installed base with stated confidence, a utilization rate measured rather than assumed, and a consumption factor derived from data rather than from a product assumption together move accuracy further than any change of algorithm.
Step 1
Establish the decision lead time first, since a prediction arriving after the purchasing deadline or maintenance window has no value regardless of accuracy.
Step 2
Decompose the forecast into its inputs, so the uncertainty can be attributed rather than carried as a single error figure.
Step 3
Establish the installed base with stated confidence, covered on our installed base page and underneath most of these forecasts.
Â
Step 4
Measure utilization where connectivity allows, and state the extrapolation where it does not.
Step 5
Separate capital from consumable forecasting, because one is a pipeline process and the other is a statistical one.
Step 6
Score historical forecasts against outcomes, since without feedback no method improves and bias persists.
Step 7
Quantify the directional bias, which is usually present, usually consistent and more actionable than average error.
Step 8
Model the cost asymmetry, since stockout and overstock are not symmetric and neither are service capacity errors.
Step 9
Assess predictive feasibility honestly, particularly for equipment failure where the history may be too thin.
Step 10
Deliver forecasts as ranges into planning systems rather than point estimates into a spreadsheet.
Clinical and device-level prediction is covered on our Predictive and Statistical Modeling page. These are the operational forecasting capabilities and the work underneath them.
Ground
Forecast Decomposition
Breaking a forecast into its component inputs so uncertainty can be attributed and effort directed at the weakest one.
Installed Base Grounding
Establishing the device population with stated confidence, since it is the shared dependency under service, consumable and capacity forecasting.
Utilization Measurement
Actual usage from connected devices, with the extrapolation to the unconnected population made explicit rather than implied.
Input Monitoring
Tracking the quantities a forecast consumes, since degradation usually begins in an input rather than in the method.
Forecast
Consumable and Demand Forecasting
High-volume, procedure-driven demand where statistical forecasting genuinely applies and where most of the value is.
Service Demand Forecasting
Intervention volume by region, skill and part, derived from installed base, age, configuration and observed failure patterns.
Capacity and Supply Planning
Production and inventory requirements under scenarios, with the cost asymmetry between shortage and excess made explicit.
Reliability Forecasting
Expected failure patterns by cohort and age, using exposure-adjusted history, covered further on our installed base page.
Qualify and Operate
Forecast Scoring
Continuous comparison against outcomes with bias quantified by segment, which is the only mechanism by which a method improves.
Uncertainty Communication
Ranges rather than point estimates, since planning against false precision produces decisions the data does not support.
Model Deployment and Operations
Forecasts delivered into planning systems and refreshed reliably, covered in more depth on our platform pages.
Drift and Input Monitoring
Watching both model behaviour and the inputs it consumes, since an ageing installed base degrades a forecast silently.
What CaliberFocus does, and does not do?
Forecast the event and response window, fix input quality before changing models, and measure historical bias to identify improvements. Commercial and operations teams retain ownership of planning, inventory, staffing and risk decisions.
These differ enormously in how much usable history exists and how comparable it is. The third column is the honest constraint, which should determine whether the work is a model, a scenario or a structured conversation.
| Forecast | What It Drives | The Constraint That Governs It |
|---|---|---|
| Consumable Demand | Production, inventory and supply planning | Strongest statistical case with sufficient history and volume |
| Service Intervention Volume | Field engineer capacity and parts stocking | Depends on complete service history linked to the installed base |
| Spare Parts Demand | Depot and vehicle inventory | Improves with detailed installed base configuration data |
| Equipment Failure on the Line | Maintenance scheduling | Limited by sparse failure data from reliable equipment |
| Field Device Failure | Proactive service and reliability planning | Requires exposure-adjusted history and connected device signals |
| Capital Equipment Sales | Revenue planning and production | Few large, negotiated decisions make pipeline analysis more suitable than statistical models |
| New Product Uptake | Launch planning and supply | Limited comparable launches and strong external influences |
Forecasts should distinguish actual customer demand from shipments constrained by stock availability. A period where you could not supply looks like a period of low demand, the model learns the constraint rather than the need, and the resulting forecast perpetuates the shortage it was built to prevent.
A forecast is a composition of inputs, and its accuracy is bounded by the weakest one regardless of method. The third column is what each contributes to total uncertainty, which is the attribution most forecasting programmes never perform.
| Input | Where It Comes From | What It Contributes to Uncertainty |
|---|---|---|
| Installed Base | Shipment, service, telemetry and inference | Often the largest uncertainty contributor and the least examined |
| Utilization Rate | Telemetry from connected devices | Potential extrapolation bias because connected devices may not represent the full population |
| Consumption per Procedure | Product assumptions or sales history | Often relies on assumptions inherited from product definition |
| Historical Demand | ERP and order history | Generally reliable, but can be distorted by stocking behaviour |
| Service History | Field service systems | Useful when complete, but may omit unbilled interventions |
| Equipment and Process Data | Historians and MES | Strong for condition monitoring, but often limited for failure prediction |
You are probably forecasting orders and calling it demand.
Order history records what customers chose to purchase, which reflects their stocking policy, their budget timing, your promotions and their own uncertainty as much as it reflects consumption. Forecasting orders and planning production against them transmits customer behaviour straight into your supply chain and amplifies it. Where utilization data exists, forecasting consumption and deriving orders is a materially better model, and the difference shows up most in the periods where it matters.
A demand forecast informs a production plan and carries ordinary commercial governance. A reliability forecast informing a service strategy for a device in the field, or a prediction contributing to a quality decision, sits closer to the quality system. Those are different obligations and the distinction should be drawn before building, because applying quality controls to demand planning is wasteful and applying commercial controls to a reliability claim is exposure.
Scope
- Each forecast classified as commercial or quality-relevant before development, since the obligations diverge
- A named business owner accountable for whether the forecast is still performing, distinct from whoever built it
- Engineering controls capable of supporting the frameworks that apply, which may include FDA regulation, ISO 13485 and electronic record expectations
- Applicability, interpretation and validation strategy determined by your quality and regulatory functions rather than by us
Explainability
- Forecasts explainable in terms the planner can challenge, since an unchallengeable number is overridden quietly rather than debated
- Input contributions visible, so a user can see whether a change came from the installed base, utilization or the method
- Assumptions documented and versioned, particularly the extrapolations the forecast depends on
- Model and input versions attached to every published forecast, since both change behaviour
Security and commercial sensitivity
- Customer-level demand and utilization data treated as commercially sensitive under the relevant agreements
- Distributor and channel data used within the terms of the agreement, which frequently restrict analytical use
- Access governed by purpose, since supply planning and commercial targeting are different grants
- Patient-associated device data handled to the obligations your counsel defines, covered on our privacy page
Lifecycle
- Forecast performance scored continuously rather than validated once at implementation
- Retraining triggered by measured degradation rather than by calendar
- Input monitoring alongside model monitoring, since an ageing installed base degrades a forecast silently
- Retirement as a routine event, because an unscored forecast still being consumed is a liability
A demand forecast accurate at national level and wrong at every region produces exactly the inventory problem the forecast existed to prevent. Errors offset in aggregate and do not offset in the warehouse, and accuracy should be measured at the level the decision is taken: by product, by region, by depot and by period rather than at the top of the hierarchy.
Accuracy at the decision level
By product, region and period rather than at the aggregate where errors cancel.
Directional bias by segment
Persistent over or under-forecasting, which is correctable and usually present.
Interval calibration
Whether stated ranges contain outcomes at the stated rate, since planning depends on the range rather than the point.
Performance against a naive baseline
Last period, same period last year or a moving average, because many forecasts do not beat these.
Input drift as well as model drift
Since an installed base ageing or a utilization sample shifting degrades the forecast with no model change.
Forecast value, not just accuracy
Whether a decision changed as a result, since an accurate forecast nobody acted on delivered nothing.
Forecasting programmes are usually reported as models deployed and accuracy percentages. The outcomes that matter are whether the forecast beat the naive baseline, whether planners use it as issued, and whether anybody is scoring it.
| Category | What We Measure | Why It Matters |
|---|---|---|
| Decision Definition | Prediction, decision owner, required lead time, impact of errors, event frequency, data sufficiency, baseline, actionability and approval requirements | Defines the decision before model development begins |
| Performance Against Naive | Forecast error compared with the previous period or seasonal baseline | Provides an honest benchmark that many forecasts fail to beat |
| Accuracy at Decision Level | Error by product, region and period | Reflects where inventory and capacity decisions are made |
| Directional Bias | Persistent over- or under-forecasting by segment | Can often be corrected without new data or methods |
| Planner Override Rate | Frequency of forecast adjustments before use | Indicates how much planners trust the forecast |
| Input Contribution | Share of total uncertainty attributable to each input | Identifies where improvement efforts will have the greatest impact |
We will score your historical forecasts against outcomes and against a naive baseline, quantify the directional bias, decompose the forecast to find which input carries the uncertainty, assess whether utilization data can replace assumption, and establish how often planners override what is issued.
Share the forecast, the history behind it, and the decision you want to improve.
Security & Compliance
