Parts Demand Forecast Accuracy (PDFA)

GMI-RE-031Renewable EnergyPilot-ReadyPattern O-1: Confidence Calibration

Definition

PDFA measures the mean absolute percentage error between AI-predicted spare parts demand and actual parts consumption, segmented by part criticality and lead time. This metric reveals whether condition-based demand signals are more or less accurate than the historical consumption models they replaced, and whether procurement decisions based on AI forecasts are grounded in reliable predictions.

The Measurement Gap

The Measurement Gap

When organizations deploy AI-driven predictive maintenance, the condition monitoring system generates alerts that predict component failures. These predictions drive not only maintenance scheduling but also parts procurement. If the AI predicts a gearbox bearing will fail in 90 days, the procurement team orders the bearing in advance. If the AI predicts elevated blade erosion across a fleet, the organization pre-orders repair materials. The accuracy of these demand forecasts determines whether the organization stocks the right parts at the right time or whether it over-stocks (tying up working capital) or under-stocks (causing downtime waiting for emergency parts).

The problem is that nobody validates whether the AI demand forecasts are actually accurate. Traditional inventory metrics (stockout rate, fill rate, inventory turns) measure whether parts were available when needed, not whether the forecast was correct. A stockout rate of 2% could reflect excellent forecasting or excellent emergency procurement that compensates for poor forecasting. Without PDFA, the organization cannot distinguish between these scenarios.

The consequence of unmeasured forecast accuracy is misallocated inventory investment. If the AI consistently over-predicts demand for critical spares, the organization carries excess inventory that ties up capital. If it under-predicts demand for long-lead-time components, stockouts occur weeks after the AI should have triggered an order.

The Decision This Metric Supports

PDFA answers whether procurement and inventory decisions based on AI demand forecasts are grounded in accurate predictions. When PDFA is above 0.75 for consumables and above 0.60 for critical spares, the forecast is adding value over historical averages. When PDFA drops below 0.50 for any category, the forecast is not adding value over simple historical consumption models, and the organization is bearing AI complexity costs without AI accuracy benefits.

Formula

Primary Formula:

PDFA(c) = 1 minus (1/N_c) x SUM(|Predicted_Demand_i minus Actual_Demand_i| / max(Actual_Demand_i, 1))

Where:

Variable Definition
c part category (critical long-lead, critical short-lead, consumable, commodity)
Predicted_Demand_i AI-forecasted demand for part i in the measurement period, in units
Actual_Demand_i actual consumption of part i in the measurement period, in units
N_c number of parts in category c
Bias(c) (1/N_c) x SUM((Predicted_Demand_i minus Actual_Demand_i) / max(Actual_Demand_i, 1))

Measurement frequency: Quarterly, aligned with typical inventory review cycles.

Segmentation requirements: Track PDFA by part criticality (critical versus consumable), by lead time (long-lead versus short-lead), and by forecast horizon (30-day versus 90-day versus 180-day). Longer forecast horizons naturally produce lower accuracy, and the acceptable accuracy threshold should adjust accordingly.

Edge cases requiring special handling:

  • Parts with zero actual consumption: the formula uses max(actual, 1) in the denominator to avoid division by zero. Parts with non-zero predictions and zero consumption should be flagged separately as "false demand signals" because they represent forecast-driven procurement for parts that were never needed.
  • Parts ordered but not consumed (received into inventory but not issued to work orders): track procurement decisions triggered versus actual installation. PDFA measures consumption accuracy, not procurement execution.
  • Seasonal parts consumed only during specific periods (heating elements in winter, cooling components in summer): compute PDFA by season rather than annually to avoid penalizing seasonally correct forecasts.

Proposed formula, pending validation pilots.

At a Glance

Unit
Percentage
Trend Direction
Higher; more accurate
Reporting Frequency
Quarterly by part category
Category
Parts and Inventory Decision Quality
← RUL Estimation Accuracy Index (REAI)Rebalancing Availability Impact Score (RBAIS) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment