Grade Crossing Risk Currency Index (GCRCI)

GMI-RAIL-044Rail TransportationIntegration-RequiredPattern O-2: Model Drift / Currency

Definition

Percentage of grade crossings where the ML risk model's training data inputs (traffic counts, road geometry, development patterns) are less than N months old, measuring whether the model's risk rankings reflect current or historical conditions.

The Measurement Gap

Why This Metric Didn't Exist Before

Grade crossing risk models have existed for decades: the FRA's own models use traffic counts, geometry, and accident history to prioritize crossings for improvement. What changed is that ML models now ingest far more data sources (satellite imagery, real-time traffic feeds, development permits) and produce continuously updated risk scores that drive resource allocation.

The problem nobody was measuring: how stale are the inputs feeding those models? A crossing that was low-risk when surveyed in 2019 may be high-risk today if a distribution center opened nearby, doubling truck traffic. The ML model doesn't know this because its traffic count input is five years old. It's still scoring that crossing as low-risk, and the capital improvement dollars are going elsewhere.

Here's the decision GCRCI supports: Which crossings need updated data collection before their risk scores can be trusted? Without GCRCI, you trust the model's output uniformly: you assume every crossing's risk score is equally valid. GCRCI segments crossings into those with current data (trustworthy scores) and those with stale data (scores that need revalidation before acting on them).

This is the O-2 Model Drift pattern applied to safety-critical infrastructure. The model hasn't drifted: the world has drifted away from the model's data. GCRCI measures the gap.

Formula

Formula:

GCRCI = Crossings_with_current_data / Total_scored_crossings

Variable Definitions:

Variable Definition Unit Source Classification
Crossings_with_current_data Grade crossings where all critical model inputs (traffic count, geometry, adjacent development) are within the currency window Count Risk model + data management DERIVED
Total_scored_crossings All grade crossings with ML risk scores Count Risk model database NATIVE
Currency window Maximum acceptable data age for each input type Months Configured N/A

Recommended Currency Windows: 36 months for traffic counts, 60 months for geometry, 24 months for adjacent development.

Sign Convention:

1.0 = all crossings have current data. Lower = more crossings scored with stale data.

Data Requirements:

Data Element Source System Classification
ML risk model scores by crossing FRA crossing database / railroad risk model NATIVE
Model input data vintage (date of last update per input type) Data management system NATIVE (if tracked) / NOT CAPTURED (if not)
Traffic counts by crossing State/county DOT data or railroad counts INTEGRATION
Adjacent development data GIS / land-use databases INTEGRATION
Geometric survey data Engineering records NATIVE

Edge Cases and Measurement Boundaries:

Crossings with no changes present a legitimate question: a rural crossing with no development changes and stable traffic may have genuinely unchanged risk. But the data needs to be confirmed as unchanged, not just unupdated. An "unchanged" data point from 2018 is stale; a confirmed-current data point from 2024 with "no change" is valid. New crossings with less than 3 years of data don't have enough history for the ML model's pattern analysis: flag as "insufficient history" rather than scoring.

Proposed formula, pending validation pilots.

At a Glance

Category
Passenger Experience and Safety Analytics Integrity
← Disruption Rebooking Equity Index (DREI)Near-Miss Severity Classification Accuracy (NMSCA) →

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