Enhanced HMDA Report - Fair Lending Attestation with STRATA + CREST

Ingests a standard HMDA LAR file (2024 FFIEC format, 85 fields), applies STRATA to proxy race/ethnicity on non-disclosed records, runs CREST disparity estimation with 95% confidence intervals, and produces two outputs: a machine-readable JSON attestation and a human-readable HTML compliance report. Designed for HMDA/ECOA/CRA fair lending monitoring. STRATA and CREST both run inside this product - no other subscriptions to set up.

BISG, the de facto proxy standard, misclassifies more than one in four minority borrowers as White (28.0% White false-positive rate on a 981,288-record voter-file holdout). STRATA reduces this to 17.8% (88.8% overall accuracy on the same holdout). More critically, BISG produces a point estimate with no confidence bounds - examiners cannot distinguish a reliable disparity finding from a statistical artifact. CREST provides 95% confidence intervals using a de-biased, Fisher-consistent weighted method-of-moments estimator, so each disparity estimate is accompanied by a clear actionability flag.

• JSON attestation: disparity point estimates and 95% CIs for White/Black, White/Hispanic, White/Asian approval rate gaps; STRATA coverage rates; lookup level distribution; CREST seed identifier for cross-institution comparability
• HTML compliance report: cover page, executive summary with actionability flags, coverage table, full disparity table with color-coded CI results, methodology note, aggregate-use attestation statement
• Actionability flag per comparison: CI excludes zero -> examiner should investigate; CI includes zero -> not statistically reliable at this institution's portfolio size

CREST uses a published, versioned simulation seed (CREST_SEED_2024Q4). Every institution running the Enhanced HMDA Report uses identical perturbation paths for shared census tracts. A state regulator receiving attestations from multiple institutions can aggregate results directly - disparity intervals are structurally comparable across institutions without any loan-level data sharing. See the interactive demo at hmda.aequum.ai.

The default output is the marginal (unconditional) disparity - the correct screening statistic for identifying where to focus examination resources, analogous to a raw HMDA disparity ratio. Under race-correlated geography confounding (common in segregated markets), this estimate may overstate the true race effect by approximately 1-3.5 percentage points. For confirmatory analysis controlling for legitimate underwriting factors, use the covariate-adjusted mode (adjust_cols parameter).

Runs as a SageMaker Model Package, deployed as a real-time endpoint entirely within your AWS account: POST a LAR file, get back a zip response. Input PII never leaves your environment. STRATA and CREST both execute in-process inside this product's own container - no separate STRATA or CREST subscription, no outbound call to any other endpoint, no data egress to aequumAI at any point.

• No other aequumAI subscriptions needed - STRATA and CREST are both bundled
• Input: HMDA LAR file (pipe-delimited or CSV, 2024 FFIEC format)
• Instance: ml.m5.xlarge recommended (ml.m5.2xlarge for larger filers)
• Endpoint config: set ContainerStartupHealthCheckTimeoutInSeconds to 900 - the bundled STRATA model takes up to ~3 minutes to cold-load, and the endpoint will fail to come up without this
• IAM: sagemaker:CreateModel/CreateEndpointConfig/CreateEndpoint/InvokeEndpoint
• Output: a zip archive (JSON attestation + HTML report) returned directly in the invocation response

Subscribe, create a SageMaker model from the subscribed package ARN, deploy a real-time endpoint (remember the 900-second startup health-check timeout), then POST your LAR file to the endpoint and unzip the response. Full usage instructions and an interactive demo with four synthetic NY institutions are available at hmda.aequum.ai. Contact terry@aequum.ai  to request a methodology walkthrough or discuss a pilot engagement.

aequumAI builds privacy-preserving demographic estimation tools for regulated industries. The STRATA methodology is documented at arXiv:2504.21259 (v2), submitted to Data-centric Machine Learning Research (DMLR). The Enhanced HMDA Report methodology is described in full at hmda.aequum.ai/#method.

STRATA and CREST are licensed only for aggregate, population-level analysis. Output must not be used to make, support, validate, or influence a decision about an identified or identifiable individual.