What is the difference between NMCS and PMCS in military readiness reporting?

NMCS stands for Not Mission Capable Supply — a vehicle or system that is grounded specifically because a required spare part is unavailable, not because of a labor or technical constraint. PMCS stands for Partially Mission Capable Supply, indicating the system can perform some but not all of its assigned missions due to a supply shortage. The distinction matters for forecasting because NMCS events directly quantify the supply gap: every NMCS day represents a part that the forecasting model failed to pre-position. Tracking the ratio of NMCS to total downtime reveals whether the inventory shortfall or the maintenance labor pool is the binding readiness constraint. A well-calibrated spare parts forecasting system should drive NMCS rates toward zero for Class IX high-criticality items.

How is a National Stock Number (NSN) structured and what fields are relevant for demand forecasting?

An NSN is a 13-digit identifier structured as a 4-digit Federal Supply Class (FSC) followed by a 9-digit National Item Identification Number (NIIN). The FSC encodes the broad commodity group — for example, 2530 covers vehicular brake, steering, axle, wheel, and track components. The NIIN uniquely identifies the item within that class. For demand forecasting, the FSC provides the grouping variable that pools demand history across similar items when individual NIINs have too few transactions for reliable estimation. Additional item-master attributes from FEDLOG that drive forecasting include the Unit of Issue (UI), Unit Price, Controlled Inventory Item Code (CIIC), Shelf Life Code, and the Essentiality Code — a rating from 1 to 9 that encodes the operational impact of a stockout, directly informing service-level targets.

Why does Croston's method outperform exponential smoothing for military spare parts?

Standard exponential smoothing implicitly assumes that demand occurs in every review period, which produces systematically biased forecasts when demand is intermittent — that is, when most periods see zero demand and rare periods see positive demand. Croston's method separates the forecasting problem into two independent exponential smoothing models: one for the inter-demand interval (how many periods between positive demand occurrences) and one for the demand size conditional on a positive demand event occurring. The combined estimate of mean demand per period is the demand size estimate divided by the interval estimate. For military spare parts, where a component may be demanded only once per quarter across a fleet of fifty vehicles, this separation prevents the zero-inflation problem that causes exponential smoothing to chronically underforecast. The Syntetos-Boylan Approximation (SBA) variant applies a bias correction of (1 − α/2) to the Croston estimate, further reducing mean squared error on real military item histories.

What is the ABC-XYZ matrix and how is it adapted for military inventory management?

The ABC-XYZ matrix classifies items on two independent dimensions. The ABC dimension ranks items by total annual demand value: A items represent roughly 20% of SKUs but 80% of spend; C items represent 50% of SKUs but only 5% of spend. The XYZ dimension ranks items by demand variability: X items have stable, predictable demand; Z items have highly intermittent or lumpy demand patterns, often driven by large infrequent maintenance events. In commercial inventory management, the matrix drives investment allocation — AX items justify sophisticated forecasting; CZ items are often managed with simple reorder points. Military adaptation adds a fourth dimension not present in commercial frameworks: the Essentiality Code from the NSN item master, which can override the cost-driven ABC ranking entirely. A low-cost, highly intermittent (CZ) item with Essentiality Code 1 (mission-essential, no substitute) may warrant higher safety stock than the commercial model would suggest, because the cost of the stockout — NMCS downtime on a critical platform — far exceeds the cost of carrying excess inventory.

How do you calculate readiness-based safety stock and how does it differ from a service-level approach?

A standard service-level safety stock calculation sets the buffer to cover demand uncertainty over the replenishment lead time at a specified fill-rate or cycle-service-level target, typically derived from an economic trade-off between holding cost and stockout cost. Readiness-based safety stock substitutes a military readiness requirement for the economic cost signal. The input is a required fleet readiness rate R (say, 90% of vehicles mission-capable) and a reliability model that maps component failure rates to expected concurrent demands during a lead-time window. Safety stock is then set so that the probability of a stockout during any single lead-time window is at most (1 − R), where R is computed at the fleet level rather than the per-item level. This is a tighter constraint than a commercial 95% cycle service level because fleet readiness is a joint probability across all critical item classes simultaneously. The practical implication is that safety stock levels for NSN Essentiality Code 1 items often exceed what a pure cost-minimization model would recommend.

What is FEDLOG and how is it used as a data source for spare parts forecasting?

FEDLOG (Federal Logistics Data) is the U.S. DoD's authoritative item identification and logistics data product, published by the Defense Logistics Agency (DLA) and available in both CD/DVD and online formats through WebFLIS (Web Federal Logistics Information Service). For a spare parts forecasting system, FEDLOG is the item master: it maps each NSN to its Federal Supply Class, unit of issue, unit price, shelf life, Controlled Inventory Item Code, and the cross-reference tables that link NSNs to manufacturer part numbers, Commercial and Government Entity (CAGE) codes, and interchangeable/substitutable (I&S) relationships. The I&S relationships are particularly important for forecasting: when a specific NSN is on back-order, a substitute NSN can fulfill the demand, so demand histories for I&S groups should be pooled before fitting a forecasting model. FEDLOG data is refreshed bimonthly; an operational forecasting system must track version changes because NSNs are periodically superseded, cancelled, or split.

How should lumpy demand be handled differently from intermittent demand in military spare parts forecasting?

Intermittent demand is characterized by many zero-demand periods interspersed with small positive demands — a single O-ring ordered once every two months, for example. Lumpy demand adds a second layer of variability: when demand does occur, the demand size itself is highly variable — perhaps zero for five months, then a fleet-level overhaul event demands 150 units in a single transaction. Croston's method handles intermittency by separating demand occurrence from demand size, but it still assumes that demand sizes are relatively stable when they do occur. Lumpy demand violates this assumption. The recommended treatment for lumpy items in a military context is a compound Poisson model: model demand occurrence as a Poisson process and demand size conditional on occurrence as a negative binomial or log-normal distribution. This two-component structure captures both the rarity of demand events and the high variance in their magnitude when they do arrive. Items with Squared Coefficient of Variation (SCV) above 0.5 and Average Demand Interval above 1.32 periods fall in the lumpy zone of the Syntetos-Boylan-Teunter demand classification diagram and warrant this treatment.

What GCSS-Army and LMP data feeds are required for an automated spare parts forecasting pipeline?

A full automated pipeline requires three categories of data from GCSS-Army (Global Combat Support System — Army) and LMP (Logistics Modernization Program, which covers depot-level supply and maintenance). First, demand history: the transaction ledger of supply requests by NSN, unit, date, quantity requested, and quantity issued, going back at least 36 months to capture low-frequency seasonal patterns and overhaul cycles. Second, on-hand and on-order positions: current stock on hand by storage location, open requisition quantities, and requisition status including estimated delivery date — these define the starting inventory position for replenishment calculations. Third, equipment density by unit: the authoritative equipment list (AEL) showing how many of each platform type each unit operates, because a demand forecast at the NSN level must be calibrated against the at-risk population. Without equipment density data, a draw-down in unit size will be misread by the model as a genuine demand reduction, producing under-stocked safety stock for the remaining fleet.

How is Mean Absolute Percentage Error (MAPE) interpreted and why is it problematic for intermittent demand items?

MAPE measures forecast error as a percentage of the actual demand: |Actual − Forecast| / Actual, averaged over all periods. For items with non-zero demand in every period, MAPE is interpretable and comparable across items with different demand scales. For intermittent demand items, MAPE is undefined in any period where actual demand is zero (division by zero) and is misleadingly large in periods where actual demand is one or two units and the forecast is slightly off. The standard alternative metrics for sparse and intermittent demand are Mean Absolute Scaled Error (MASE), which scales the error by the in-sample naive-forecast error and remains well-defined at zero, and the Periodic Fill Rate, which measures the proportion of demand transactions satisfied without a stockout — directly mapping to the readiness-based service-level target. For model comparison during development, the scaled interval score from the M5 competition framework (which penalizes both interval width and interval violations) is increasingly used in defense inventory research.

What is the Syntetos-Boylan Approximation (SBA) and when should it be used instead of standard Croston's method?

The Syntetos-Boylan Approximation is a bias-corrected variant of Croston's method derived from the observation that Croston's original formula produces a systematically upward-biased estimate of mean demand per period. The source of the bias is that Croston's estimator of the mean demand interval is itself a random variable, and dividing by a random variable introduces a positive bias in the ratio. The SBA correction applies a multiplicative factor of (1 − α/2) to the Croston demand-rate estimate, where α is the smoothing parameter used in the demand-size exponential smoother. Empirical studies across commercial and defense spare parts datasets consistently show that SBA produces lower mean squared error than Croston on intermittent items and performs no worse on non-intermittent items. The operational guidance is to use SBA as the default for all items with Average Demand Interval greater than 1.25 periods, which covers the majority of Class IX military spare parts. Standard Croston remains useful only as a pedagogical reference; production systems should implement SBA or the more recent Teunter-Syntetos-Babai (TSB) method, which adds a probability-of-occurrence term that handles obsolescence more gracefully.

Spare parts forecasting for military fleets

Spare parts demand forecasting for military fleets is the quantitative discipline of predicting how many units of each National Stock Number (NSN) will be consumed by a vehicle or equipment fleet over a future planning horizon — and pre-positioning those parts before demand materializes rather than reacting after a platform goes down. The operational stakes are direct: a vehicle that is Not Mission Capable Supply (NMCS) because a $40 seal is unavailable contributes nothing to readiness, regardless of how well the rest of the fleet is maintained. This guide covers the full technical stack — item-master data architecture, demand classification, stochastic forecasting models, readiness-driven safety stock, ERP integration, and closed-loop accuracy management — in the depth required to design or evaluate a production system for a defense logistics organization.

Why spare parts forecasting matters in military readiness

Military vehicle and equipment fleets are rated against a readiness standard — the percentage of authorized platforms that are fully mission capable (FMC) at any given time. A typical brigade-level armored formation may carry a 90% FMC requirement for its organic ground vehicles, meaning that in a formation of 100 vehicles, no more than 10 may be non-mission-capable at any point during a readiness reporting cycle. Every vehicle that is down due to a missing part directly erodes this figure.

The regulatory taxonomy distinguishes two types of supply-driven non-availability. NMCS (Not Mission Capable Supply) describes a platform that has been deadlined because a required part is absent from the supply chain — the maintenance action is understood, the maintainer is available, but the part has not arrived. PMCS (Partially Mission Capable Supply) describes a platform that can perform some but not all assigned missions because a non-critical component is on back-order; it remains available for reduced-scope tasking. The distinction is operationally important: NMCS events represent an absolute supply failure, while PMCS events represent a degraded but not zero contribution to readiness. A well-designed forecasting system aims to drive NMCS rates to zero on high-Essentiality-Code items while tolerating brief PMCS conditions on lower-criticality parts where carrying excess stock would not be economically justified.

The connection between spare parts availability and readiness also has a time dimension that distinguishes military from commercial inventory management. A commercial retailer that runs out of a product loses a sale; a military unit that cannot field a required capability during a deployment window may fail its mission. The demand signal in military logistics is therefore not purely customer-driven — it is also readiness-driven, meaning the required service level is defined by the operational requirement rather than by a cost-minimization calculation. This asymmetry is the foundational fact that shapes every architectural decision in a military spare parts forecasting system. For background on how predictive maintenance for military fleets integrates with supply forecasting to reduce the overall NMCS burden, see the companion guide on CBM+ platforms.

Data inputs: NSN item master, FEDLOG, and historical demand signals

Every reliable spare parts forecast begins with a clean item master. In the U.S. defense logistics system, the authoritative item master is the National Stock Number catalog published by the Defense Logistics Agency (DLA) and distributed through FEDLOG (Federal Logistics Data) and its web interface, WebFLIS (Web Federal Logistics Information Service). An NSN is a 13-digit identifier structured as a 4-digit Federal Supply Class (FSC) followed by a 9-digit National Item Identification Number (NIIN).

The FSC provides the first grouping variable for forecasting: items within the same FSC are broadly interchangeable from a maintenance-function standpoint, so FSC-level demand aggregates can be used to construct priors for individual NIINs that have sparse transaction histories. Relevant item-master attributes for forecasting include:

FEDLOG Field	Forecasting Use
Federal Supply Class (FSC)	Demand pooling; FSC-level prior for Bayesian shrinkage
Unit of Issue (UI)	Demand unit normalization across transactions
Unit Price	ABC demand-value classification; holding cost calculation
Essentiality Code (EC)	Service-level target override; NMCS risk classification
Shelf Life Code	Maximum stock position; obsolescence risk flag
I&S Group (Interchangeable & Substitutable)	Demand pooling across substitute NSNs; fill-rate calculation
CIIC (Controlled Inventory Item Code)	Storage constraint; security-driven replenishment routing

The demand history schema required for forecasting should be assembled at transaction granularity — one row per supply request — with at minimum: NSN, requesting unit identifier, transaction date, quantity requested, quantity issued (which may differ from requested if stock was partially available), and a fulfillment flag indicating whether the request resulted in a back-order. A minimum of 36 months of history is needed to capture seasonal patterns and the low-frequency overhaul cycles that dominate many Class IX item demand profiles. FEDLOG is refreshed bimonthly; an operational system must reconcile NSN supersessions and cancellations at each refresh cycle so that demand history for a cancelled NSN is correctly attributed to its successor before fitting any model.

Demand classification: slow-mover vs intermittent vs lumpy

Before selecting a forecasting model, each item must be classified by its demand pattern. The Syntetos-Boylan-Teunter (SBT) classification framework — the standard in spare parts inventory management research — uses two statistics computed from the demand history:

Average Demand Interval (ADI): the mean number of time periods between consecutive positive demand observations. An ADI above 1.32 periods signals intermittent demand.

Squared Coefficient of Variation of demand sizes (SCV): the variance of positive demand quantities divided by the square of the mean positive demand quantity. An SCV above 0.49 signals high size variability.

The four-quadrant demand classification diagram using these thresholds:

          SCV ≤ 0.49          SCV > 0.49
          ─────────────────────────────────────
ADI ≤     │  SMOOTH            │  ERRATIC        │
1.32      │  (Holt-Winters)    │  (SES + extra)  │
          ├────────────────────┼─────────────────┤
ADI >     │  INTERMITTENT      │  LUMPY          │
1.32      │  (SBA / Croston)   │  (Compound      │
          │                    │   Poisson)      │
          ─────────────────────────────────────

In practice, military Class IX spare parts inventories are dominated by intermittent and lumpy items. A survey of U.S. Army authorized stockage lists consistently shows that more than 60% of NSNs by line count fall in the intermittent or lumpy quadrant, with the smooth quadrant populated largely by high-turnover consumables (batteries, filters, lubricants) and the erratic quadrant capturing items with occasional large-lot demands driven by fleet exercises.

The ABC-XYZ matrix adapts the SBT classification to include value. A items (top 20% of NSNs by annual demand value) that are also in the Z demand-variability tier (highly intermittent or lumpy) are the items that require the most sophisticated forecasting treatment and the most careful safety stock configuration. They are typically also the items most likely to carry high Essentiality Codes, because the high unit price correlates with component complexity and the long procurement lead times that make stockouts consequential.

Stochastic forecasting models

The central insight behind stochastic spare parts forecasting is that intermittent demand cannot be adequately described by a single point estimate — the forecast must characterize the full distribution of demand over the replenishment lead time, because safety stock is calculated from the tail of that distribution, not from the mean.

Croston's method and the Syntetos-Boylan Approximation (SBA). Croston's method, introduced in 1972, separates the forecasting problem into two independent exponential smoothing models: one for the inter-demand interval q̂ and one for the demand size ẑ conditional on a positive demand event. The estimate of mean demand per period is ẑ/q̂. The SBA variant applies a bias correction — multiplying the Croston estimate by (1 − α/2) where α is the smoothing parameter — that has been shown analytically and empirically to reduce mean squared error across intermittent demand datasets. For a military system, the SBA estimate should be initialized from FSC-level priors for items with fewer than 12 non-zero historical observations:

# SBA demand rate estimate (Python pseudocode)
alpha = 0.15          # smoothing parameter (tuned per item)
z_hat = demand_size_ema     # EMA of positive demand quantities
q_hat = interval_ema        # EMA of inter-demand intervals

# Croston rate
croston_rate = z_hat / q_hat

# SBA bias-corrected rate
sba_rate = (1 - alpha / 2) * croston_rate

# Lead-time demand distribution: Poisson with mean = sba_rate * lead_time_periods
import scipy.stats as stats
lt_demand_dist = stats.poisson(mu=sba_rate * lead_time)

Compound Poisson models for lumpy demand. When both ADI exceeds 1.32 and SCV exceeds 0.49, the demand process is lumpy — the rare events that do occur involve highly variable quantities. The standard model is a compound Poisson process: demand occurrence follows a Poisson process with rate λ, and demand quantity conditional on occurrence follows a secondary distribution, typically negative binomial for non-negative integer quantities. The lead-time demand distribution is then the sum of a Poisson-distributed number of negative binomial draws, which has a closed-form probability generating function but requires numerical integration for quantile calculation. In a production system, the quantile function is typically pre-computed and stored as a lookup table indexed by (λ, r, p, lead_time) to avoid runtime integration.

Poisson models for low-volume Essentiality Code 1 items. For items that are demanded at most a few times per year and carry an Essentiality Code that makes any stockout unacceptable, a simple Poisson model — where the rate parameter is estimated as the total observed demands divided by the total observed periods, shrunk toward the FSC prior — is often preferred over SBA. The simplicity of the Poisson model makes the safety stock calculation fully transparent and auditable, which matters when a supply officer must justify a requisition that appears to overstock a rarely demanded item. Transparency is an underappreciated design requirement in defense inventory systems.

Readiness-based safety stock calculation

Commercial safety stock formulas derive the buffer quantity from a target service level that is itself derived from an economic trade-off between holding cost and stockout cost. In military inventory management, the service-level target is set by the operational requirement — a readiness rate mandated by the operational concept — and the holding cost constraint is secondary. This inversion has significant practical consequences for safety stock levels.

The readiness-based safety stock calculation for a single NSN over a replenishment lead time L (in periods) proceeds as follows:

1. Specify the required fleet readiness rate R (e.g., 0.90 for a 90% FMC requirement).

2. Identify the set of NSNs classified as mission-essential for the platform type. Let this set have cardinality n.

3. Distribute the allowable stockout probability budget across the n items. The joint probability that no item is in stockout is the product of per-item non-stockout probabilities, so the per-item fill-rate target P_i satisfies the constraint:

Product(P_i for i in 1..n) >= R

# Equal allocation (conservative baseline):
P_i = R^(1/n)

# Example: R=0.90, n=20 mission-essential NSNs
# P_i = 0.90^(1/20) = 0.9947
# Each NSN must achieve 99.47% fill rate

4. Compute safety stock SS_i for each NSN as the quantile of the lead-time demand distribution at level P_i, minus the mean lead-time demand:

mu_LT = sba_rate * L          # mean lead-time demand
SS_i  = quantile(dist, P_i) - mu_LT

# Reorder point:
ROP_i = mu_LT + SS_i

The equal-allocation approach above is conservative: it treats each NSN as independently responsible for the full readiness shortfall. A more sophisticated approach uses importance weights derived from the Essentiality Code and the expected downtime cost per stockout event to allocate the budget asymmetrically — giving more of the stockout allowance to cheaper, more easily sourced items and less to long-lead, high-cost items. The asymmetric allocation typically reduces total inventory investment by 10–20% compared to equal allocation while maintaining the same aggregate readiness rate.

It is also worth noting the interaction between spare parts forecasting and military inventory management software: the safety stock levels computed here must be written back to the supply management system as authorized stockage list quantities, and the replenishment logic in that system must then enforce the reorder point. A forecasting model that produces correct safety stock targets but cannot propagate them to the execution system provides no operational benefit.

Integration with ERP and maintenance management systems

A spare parts forecasting model that runs in isolation from the supply execution system is an academic exercise. Production value requires bidirectional integration: the forecasting system consumes data from ERP and maintenance systems, and its outputs drive replenishment actions in those same systems.

GCSS-Army (Global Combat Support System — Army) is the primary unit-level supply and maintenance system for U.S. Army formations. It exposes transaction history through SFTP-based batch file exports in fixed-format flat files (the legacy SARSS/ULLS-G format) and, in more recent versions, through web service APIs. A forecasting pipeline consuming GCSS-Army data must handle: NSN supersession (when DLA replaces an NSN, open requisitions and on-hand records may temporarily carry the old NSN); unit reorganization events (when a unit restructures, its equipment density changes and historical demand rates become non-representative); and the distinction between demands (supply requests) and actual issues (fulfilled requests), since the demand quantity is the true forecast target but only issues appear in standard transaction logs unless specifically requested.

LMP (Logistics Modernization Program) covers depot-level supply, maintenance, and procurement for the U.S. Army. Integrating with LMP is necessary to incorporate depot repair cycle times into lead-time estimates — a component with a 90-day depot overhaul pipeline produces a very different lead-time demand distribution than one with a 5-day commercial procurement. LMP exposes data through SFTP-delivered flat files and, for authorized partners, through SAP-standard BAPIs. The AWRDS (Army War Reserve Deployment System) interface adds theatre-level pre-positioned war reserve stock positions that must be excluded from unit-level stockage calculations to avoid double-counting available inventory.

The data synchronization pattern that works reliably in practice is an extract-transform-load pipeline running on a daily cycle, with an event-driven supplement that triggers an immediate position update when a large demand transaction (above a threshold quantity) arrives. The daily cycle maintains inventory position accuracy for normal demand patterns; the event-driven supplement prevents the system from operating on stale positions when a battalion-level maintenance event generates a demand spike that would otherwise not be reflected until the next day's batch.

The integration with maintenance management also enables demand sensing — the use of open work orders as a leading indicator of imminent parts demand. When a maintenance management system records that a technician has opened a work order for a specific repair action, the spare parts system can look up the bill-of-materials for that repair type and pre-stage a pull request for the required NSNs before the parts are formally requested. This pattern, well established in commercial aerospace MRO, reduces the effective lead time for high-priority maintenance actions and is directly transferable to the defense context using GCSS-Army's work order data. The broader context for this type of AI-driven demand anticipation is covered in detail in the article on AI optimization in military logistics.

Continuous improvement: closed-loop forecast accuracy

A spare parts forecasting system that is not continuously measured and updated will degrade as fleet composition changes, operational tempo shifts, and the demand process drifts away from the statistical patterns observed in the training data. Closed-loop accuracy management is not an optional enhancement — it is the mechanism that keeps the model's outputs operationally valid over time.

The primary accuracy metrics for intermittent demand items are Mean Absolute Scaled Error (MASE) and Periodic Fill Rate. MASE scales the absolute forecast error by the in-sample naive-forecast error (using the last observed positive demand as the naive forecast), making it comparable across items with very different demand volumes and avoiding the undefined behavior of MAPE at zero-demand periods. Periodic Fill Rate measures the fraction of demand transactions satisfied without a back-order during the measurement window, directly mapping to the readiness-based service-level targets established in the safety stock calculation step.

A practical monitoring framework for a military spare parts system:

Monitoring cadence:
  Monthly:  MASE and Fill Rate by NSN and by FSC aggregate
  Monthly:  Bias tracking (signed error, to detect systematic over/underforecast)
  Quarterly: Full model re-selection and parameter re-estimation
  Event-driven: Retraining trigger when:
    - Fill rate drops >5 pp below target for 2 consecutive months
    - Equipment density table changes by >10%
    - New platform variant added to the fleet
    - Overhaul cycle event recorded in maintenance system

Escalation routing:
  FSC fill rate < target - 5 pp → Supply analyst review
  NSN MASE > 2.0 for 3 months → Model class reassignment
  Systematic positive bias (>20% of mean) → Check for demand shift
  Systematic negative bias → Check for population growth or tempo increase

The closed-loop feedback also requires tracking manual overrides. Supply officers frequently override model-generated replenishment recommendations based on local knowledge — an anticipated exercise, a known supplier disruption, a recently observed failure pattern not yet visible in historical data. These overrides should be captured in the system with a reason code, and the outcome (whether the override was justified by subsequent demand) should be fed back into the model's training data. Override outcomes are a rich source of domain knowledge that is otherwise lost; a system that ignores them cannot learn from the supply officer's operational judgment.

Model drift detection is the final element of the closed-loop system. Military demand processes are not stationary: fleet sizes change with unit activations and deactivations, operational tempo varies with the training calendar and deployment cycles, and maintenance doctrine changes when new technical manuals are issued. A Cumulative Sum (CUSUM) control chart applied to the signed forecast error at the NSN level will flag a demand-level shift within a few periods of its occurrence, triggering a parameter re-estimation before the safety stock falls below the minimum required level. The CUSUM threshold should be calibrated so that the expected time to detection of a 50% demand increase is less than one replenishment lead time, ensuring that the model update can drive a replenishment action before a stockout occurs.

The discipline of closed-loop forecast accuracy management is what separates a spare parts forecasting system that maintains its initial accuracy over a multi-year deployment from one that degrades into an expensive shelf-ware installation. Defense logistics organizations that invest in the monitoring and retraining infrastructure consistently report sustained readiness rate improvements, while those that treat the initial model fit as the endpoint of the program typically see accuracy erode within 12 to 18 months as the underlying demand process evolves.

Spare parts demand forecasting for military fleets