aequum STRATA: Race/Ethnicity Proxy Model (US Name + Address)

STRATA estimates the most likely race/ethnicity of individuals from name and US address for aggregate, population-level analysis such as HMDA/ECOA/CRA fair-lending monitoring, housing-equity research, and disparity studies. It is not licensed for decisions about identified or identifiable individuals - see the EULA for permitted-use and prohibited-use terms.

The model uses a published char-LSTM + geography architecture refined by gradient-boosted soft-voting, improving on traditional BISG/BIFSG approaches (see the held-out comparison under "How It Improves"). A geography waterfall selects the best-supported tier per record (block-group+tract, tract, ZCTA, or name-only) and reports the lookup level used.

Each scored record returns:

• Probabilities for five classes: White, Black, Hispanic, Asian, Other
• Argmax label (most likely class)
• Per-record uncertainty index (IU) - addresses a known BISG limitation where no per-record confidence is available. The uncertainty index was calibrated on Florida, Georgia and North Carolina voter records and PPP borrowers; its calibration in other populations is not established.
• Privacy-preserving cohort tokens (cohort_tract, cohort_zcta) for geographic grouping without exposing census identifiers
• Lookup level and quality codes indicating which geographic tier was resolved

Resolved geography is withheld from output to protect privacy.

• Character-level deep learning captures name-origin signals that surname-list approaches miss
• Geography waterfall uses the most granular available tier rather than a single fixed level
• Per-record uncertainty (IU) lets you flag low-confidence predictions for review - standard BISG provides no such signal
• Lower false-positive bias: On a held-out voter-registration sample, STRATA's block-group tier assigned 18.0% of non-white individuals to white, versus 28.5% for surname + ZIP-code BISG.

STRATA deploys as a SageMaker real-time endpoint or Batch Transform entirely within your AWS account. Your input PII never egresses to the vendor - the vendor never sees your data. The model inherits your existing AWS security posture including encryption in transit (TLS) and at rest (S3 server-side encryption for model artifacts). You control VPC configuration, IAM permissions, and network isolation.

• Model package size: approximately 1.45 GB
• Startup time (version 1.0.1): the container loads all models and geographic tables at startup and reports ready only when they are loaded; set ContainerStartupHealthCheckTimeoutInSeconds to 900 for real-time endpoints. On ml.m5.xlarge an endpoint reaches InService about 9 minutes after creation (measured: 524 s), when loading is complete, and the first request is then answered normally (measured: 7 s).
• Batch Transform: recommended for one-off or batch scoring to avoid standing hourly endpoint costs.

After subscribing, deploy the model package to a SageMaker endpoint or configure a Batch Transform job. Input records require name and US address fields. Contact the aequumAI team at the support email to request a methodology walkthrough or discuss a pilot scoring engagement for your use case.

aequumAI builds privacy-preserving demographic estimation tools for regulated industries. STRATA builds on the methodology described in arXiv:2504.21259. The deployed model is trained on a larger corpus than the paper's evaluation models. It is designed for fair-lending compliance workflows where accuracy and auditability are critical. Methodology reference: S. Chalavadi, A. Pastor, and T. Leitch, arXiv:2504.21259 (https://arxiv.org/abs/2504.21259 ).