Definition
CRA is the proportion of resource-period observations where actual measured utilization meets or exceeds the forecast minus a tolerance band τ. The directional framing matters: CRA is defined on the under-utilization failure mode (actual fell short of forecast) because that is the financially costly direction. When you commit Reserved Instances based on a forecast and actual utilization comes in significantly below that forecast, you have paid for capacity you did not use.
The Measurement Gap
Teams buy Reserved Instances and make Savings Plan commitments based on AI-generated utilization forecasts, then never check whether those forecasts were right. The finance team uses the forecast to justify the commitment. Engineering uses it to size the cluster. FinOps uses it to project quarterly spend. Thirty days later, when actual utilization comes in 25 percentage points below forecast, nobody connects the outcome back to the forecast quality. The over-commitment gets attributed to "workload variability" and the AI tool gets another year of unconditional trust.
The Flexera 2024 State of the Cloud Report found that 32% of cloud spend is wasted: provisioned capacity that goes unused. A significant portion of that waste originates upstream in the forecasting step, not in the allocation step. You cannot provision accurately from a forecast you have never validated.
The failure mode CRA is designed to catch is systematic directional bias in a forecasting model for a specific workload class. Your FinOps AI tool may forecast general-purpose compute utilization accurately while consistently over-forecasting GPU instance utilization for ML training jobs. The model has learned that ML jobs look like high-utilization workloads in the training data, but your actual ML pipeline runs in short burst jobs with long idle windows between them. The forecast says 80%; reality is 45%. The cluster runs half-empty, the team pays for committed capacity they are not using, and the quarterly FinOps review blames "unpredictable ML workloads" rather than a model that has not been validated against actual ML usage patterns.
The second failure mode is timing drift. Cloud forecasting models trained on historical workload patterns degrade when workload composition changes, a new product launch, a data platform migration, an ML pipeline moving from batch to streaming. The model continues generating confident-looking forecasts based on patterns that no longer describe the current workload. Without measuring CRA over time, you have no signal that the model's validity has expired.
Without CRA, you are flying formation with a compass you have never calibrated. The instruments look fine. The heading might be wrong.
Formula
CRA measures whether your AI-generated cloud resource utilization forecasts are accurate within a defined tolerance: specifically, the fraction of forecasted resource-periods where actual utilization came within τ percentage points below the forecast. It tells you how often the model is close enough to trust.
Definition:
CRA is the proportion of resource-period observations where actual measured utilization meets or exceeds the forecast minus a tolerance band τ. The directional framing matters: CRA is defined on the under-utilization failure mode (actual fell short of forecast) because that is the financially costly direction. When you commit Reserved Instances based on a forecast and actual utilization comes in significantly below that forecast, you have paid for capacity you did not use.
Formula: plain language:
For each resource type in each measurement period, check whether actual utilization was at least as high as the forecast minus tolerance τ. Count the fraction of all resource-periods where this condition holds.
Notation:
CRA = (1/N) × Σᵢ I(U_actual,i ≥ U_forecast,i − τ)
Where:
- N = total number of resource-period observations in the trailing 90-day window
- U_actual,i = actual utilization percentage for resource i in period i, sourced from cloud provider monitoring or billing APIs
- U_forecast,i = the AI-generated utilization forecast for resource i at the start of period i: the forecast that was actually used for capacity or commitment decisions, not a retroactive forecast computed today- τ = tolerance band in percentage points; default 10pp (the forecast is considered accurate if actual utilization was within 10 percentage points below the forecast)
- I(·) = indicator function, equal to 1 when the condition is true and 0 otherwiseWorked Example Your FinOps platform generates weekly EC2 utilization forecasts for 12 instance families over 20 weeks in Q3. That is 240 resource-period observations. You set τ = 10pp. Review each observation: for 198 of the 240 periods, actual utilization was at or above forecast minus 10pp. For 42 periods, actual utilization fell more than 10pp below forecast. CRA = 198/240 = 0.825 (82.5%). This clears the 80% threshold: accurate enough to make RI commitment decisions with reasonable confidence. Now examine the 42 failures. Thirty-eight of them come from two instance families: GPU instances used for ML training and M-family instances running variable batch workloads. Standard compute families for steady-state applications are well-forecast. The variable workloads are not. That narrows the remediation target considerably: you do not need to retrain the whole model, you need better handling for burst workload patterns.
Companion metric: Systematic Commit Deviation Index (SCDI):
SCDI = (1/N) × Σᵢ (U_forecast,i − U_actual,i)SCDI is the mean signed forecast error. Positive SCDI means the model systematically over-forecasts: you are consistently committing to more capacity than you actually consume. Negative SCDI means systematic under-forecast: you are leaving discount opportunities on the table because the model thinks you will use less than you do. CRA tells you how often the forecast is wrong. SCDI tells you which direction it is wrong in. A model with CRA of 0.75 and SCDI of +18pp has a very different problem than a model with CRA of 0.75 and SCDI of −4pp. The first is an over-committing model. The second is a slightly-under-forecasting model with occasional large misses. The fixes are different.
Sign convention:
Higher CRA is better. CRA = 1.0 means every forecast was accurate within tolerance. Minimum meaningful target before making binding financial commitments: 0.80.
Edge Cases
Changing workloads: CRA computed during a period of major workload change (a new product launch, a large data migration, a significant architecture change) will be artificially low. The model has not seen the new workload pattern and cannot be expected to forecast it accurately. Segment results by workload stability class. Do not use a CRA measurement anchored in a transition period to make commitment decisions. Tolerance band choice: τ = 10pp is a reasonable default for compute resources. Storage workloads (which tend to grow monotonically) and data transfer (which spikes with product releases) may warrant different tolerances. Document your choice and apply it consistently. Changing τ mid-measurement window invalidates the comparison. RI commitment horizon: For 1-year or 3-year Reserved Instance decisions, compute CRA over a 180-day window minimum and apply seasonal adjustment if your workload has quarterly or annual cycles. The 90-day default is appropriate for operational monitoring; commitment decisions need longer horizons.
Cross-references:
- Upstream: CPA (GMI-IT-009, Capacity Prediction Accuracy). CPA measures raw workload demand forecast accuracy; CRA uses those forecasts as inputs to resource allocation; systematic errors in CPA propagate directly into CRA failures
- Downstream: SCWR (GMI-IT-015, Stranded Commitment Waste Rate) (whether your team successfully captures available RI and Savings Plan discounts depends on having CRA high enough to commit with confidence; CPFI (GMI-IT-014, Commitment Portfolio Flexibility Index)) unit cost economics require stable denominator values, which depend on accurate resource forecasting
Proposed formula, pending validation pilots.