Arena Analytics

The prediction is the product. Every module on the company's site is a model output sold to an employer: Retention Prediction scores the likelihood that a specific applicant will show up and stay in a specific role at a specific location; Talent Rerouting sends applicants to other openings where they are predicted to succeed; Talent Discovery routes candidates who are statistically likely to succeed into an employer's pipeline; Flight Risk flags emerging attrition by role, location and team; Internal Mobility recommends internal moves; and the Workforce Dashboard and Workforce Development products sell the same outcome intelligence to employers and workforce boards.

The company describes machine learning models trained on its own career outcome dataset (4.2 million people, growing by about 70,000 a month), external data and applicant data, with a separate bias mitigation model. There is no candidate network, job board or service that works without the models.

The models score, route and prioritize candidates, and no oversight structure around them is published. The home page says the platform routes high-fit candidates directly into an employer's pipeline, prioritizes candidates statistically more likely to stay before offers are made and reroutes others to better-fit roles, and verifies credentials for licensed roles before candidates enter the pipeline. Talent Rerouting, which offers a lower-fit applicant other roles rather than discarding them, lowers the cost of a wrong prediction to the person and is a real design choice. Recruiters are described as spending more time with applicants; the employer makes the hire.

What is not published is whether a low retention prediction can remove a candidate from consideration without a recruiter seeing the application, at what score that happens, whether recruiters see the reasons for a score, and what an applicant is told. A wrong positive costs the employer a hire that does not last; a wrong negative costs a person a job, and nobody sees it.

The approach and its inputs are named without version or update practice. The company describes machine learning models fitted to each role and organization using scores of data points, combining its own career outcome dataset (4.2 million people, growing by about 70,000 a month), external data and applicant data, and a separate bias mitigation model built on adversarial fairness, a method described in the fairness literature, in which two models are trained against each other. The predicted outcomes are named: showing up on day one, retention in the role, and attrition risk.

No model card, feature list, methodology paper, version or retraining schedule is published. Because models are fitted per role and organization on top of pooled data, a buyer cannot tell how much of a prediction comes from its own history and how much from other customers', which decides whether early performance holds.

The model layer is the company's own, which keeps the chain short, and the rest of it is unnamed. The Retention Prediction and Why Arena pages describe the company's own machine learning and bias mitigation models trained on its proprietary dataset, so no outside model provider receives applicant data for scoring. The privacy policy says applicant information goes to the client employers the person applied to and to one or more third parties the company uses to create and maintain the application, and names Google Analytics for the website; the hosting provider and other processors are not named, and no subprocessor list is published.

Pooled training is the other half of this question: the privacy policy says each person's information is used with everyone else's, de-identified and aggregated, to improve the algorithms that serve every customer.

Outcome percentages are published without method, denominator or named customers. The home page reports a 21 percent average improvement in first-year retention, a 4x minimum return on investment in year one, a 20,000 employee hospital system that reduced turnover by 35 percent and reported more than $19 million in savings, and a quote from an unnamed academic medical system of 40,000+ employees reporting 90-day turnover down 78 percent and 360-day turnover down 26 percent after two years. The case studies page carries one study, an unnamed 800 employee organization that reduced short-term turnover by 7 percent with Talent Discovery. Talent Rerouting is said to produce 2.8 to 3 applications per applicant.

None of these names the customer, the cohort, the comparison or the period, and they measure different things, so a buyer cannot reconcile a 7 percent result with a 78 percent one. No independent or peer reviewed evaluation is published. The post-hire reports the company offers each customer are the mechanism that could produce this evidence in aggregate.

Two specific commitments are published, alongside a grant to train on everyone's data. The privacy policy says the company does not collect Social Security, bank or card numbers, that demographic information such as gender, ethnicity or sexual orientation, if collected, is used only to validate that predictions reduce bias, and that each person's information is used with everyone else's to improve the company's algorithms and predictions in de-identified, aggregate form. Data is retained for as long as reasonably necessary for the purposes collected and for contractual and legal obligations, with no period stated, and is stored on the company's servers in the United States with encryption.

Training on pooled data from every customer's applicants is a practice disclosed against the people scored, whatever its de-identification, and it means one employer's applicant pool informs predictions sold to competing employers in the same labor market. The people most affected are applicants who were never hired, who have no account and may not know a score exists. Safety engineering, meaning how a harmful or wrong prediction is caught and handled, is not published. No patient information is processed.

A privacy policy reaches the service, and no business associate status or scope position is stated. The privacy policy covers applicants whose information the company processes on behalf of its clients, describes collection, use, sharing, storage and rights requests, and says nothing about protected health information. The product scores job applicants and employees rather than patients, and nothing on the site describes patient data, but the company does not say so; a plain statement that the service handles no protected health information, with its scope explained, is what a buyer would rely on.

The home page says credentials are verified for licensed roles before candidates enter the pipeline, and the relationship with an employer is governed by a customer agreement that is not published.

A recognized attestation is named without the artifact. The Why Arena page says the platform is SOC 2 Type II compliant, meaning its security has been independently examined against the AICPA's criteria; no report date, period, scope or auditor is published, and the site has no trust center. The privacy policy describes encryption and industry standard methods and places the risk of breaches caused by stolen user credentials on users.

The scope question matters because the company connects into employers' Workday, ADP, UKG, Oracle and other HR systems and holds identified employment histories and scores for applicants who were never hired.

The product predicts employment outcomes and makes no clinical claim, so no device claim is made and none applies. What governs it is employment discrimination law and the growing body of rules on automated employment decision tools, such as New York City's requirement that an employer using such a tool for roles in the city obtain and publish an independent bias audit. The company states that its recommended candidate groups exceed the EEOC's 80 percent threshold and describes its adversarial fairness method, but it publishes no bias audit, candidate notice or statement of how it supports employers' compliance duties.

The governing law is employment and AI hiring law rather than a device pathway, and in covered jurisdictions the audit duty sits with the employer.

A mitigation method and a parity claim are published, with no evaluation behind them. The Retention Prediction and Why Arena pages say a proprietary bias mitigation model uses adversarial fairness, pitting two machine learning models against each other to remove biased data points, and the home page says protected characteristics and related signals are removed from the models and that the recommended candidate group consistently exceeds the EEOC's 80 percent threshold, often surpassing 95 percent. The privacy policy adds that demographic information, where collected, is used only to validate that predictions reduce bias.

The 80 percent figure is a selection rate ratio across groups, but no groups, counts, roles, customers or period are published, and no independent audit or adverse impact analysis is available to read. A parity claim of this kind is named here and does not raise the grade. Removing proxies for protected characteristics is the hard part of fairness in hiring models, because turnover history, commute distance and employment gaps correlate with protected characteristics, and a model trained on who stayed can reproduce who has historically been allowed to stay.

A guarantee is advertised without its terms, and nothing published stands behind the output. The Retention Prediction page says the company guarantees greater retention for candidates it rates Likely and can prove it with post-hire reports, and the Why Arena page says its results are backed by a guarantee; neither states what is measured against what, over what period, or what the customer receives if retention falls short. The terms of service provide the service as is and disclaim any representation as to the accuracy of the service or the results to be achieved, so the guarantee lives, if anywhere, in customer contracts that are not published.

The post-hire reports are a real measurement mechanism, because they compare what happened to the people the models scored. They reach the employer, not the applicant: the person scored Unlikely or rerouted has no published route to see or contest the score, and the privacy policy's access and correction rights reach personal data, not the prediction. A wrong low score costs a person a job and is never observed, because a rejected applicant's retention cannot be measured.

No clinical record integration exists: the product scores job applicants and employees, and nothing connects to an electronic health record. The systems that matter are the employer's applicant tracking and human resources platforms, and the Retention Prediction page names nine of them: Workday, ADP, UKG, Paycom, iCIMS, TalentReef, Jobvite, Oracle and symplr.

What is not published is the depth of any of them: whether scores are written back into the applicant record, whether exchange is two way and real time or batch, which customers run each connection in production, and how much implementation work each takes.

A single hosted service, with the United States stated as where it is operated and where data may be stored. The privacy policy says the services are controlled and operated from the United States, that information may be stored there, and that data is kept on the company's servers and other systems with encryption; it names no hosting provider, region or tenancy model.

The data path runs through employers' applicant tracking and HR systems: candidate records move from an employer's recruiting system into the company's environment for scoring and back. The Workforce Development product also serves public workforce boards, which bring procurement and public records obligations a private employer does not.

The selling posture is discoverable and no figure is published. The pricing page presents two packages, one for medium sized businesses that starts with a small set of products and grows into the suite, and one for enterprise, both leading to a conversation. The home page says Talent Discovery is paid only when the employer hires, which states a unit of charge without an amount. An ROI calculator frames the purchase against avoided turnover cost, and the retention guarantee advertised on the Retention Prediction page has no published terms.

The suite is modular, one product is priced per hire, and the rest is quoted.

Coverage is claimed broadly and not specified. The home page describes frontline and licensed roles, with quotes from an unnamed statewide health system, a senior living community and an academic medical system, and the Workforce Development page gives an example drawn from healthcare aides who become registered nurses; the product is sold to workforce boards across regional economies as well as to employers. No list of settings, roles or specialties the models are built and validated for is published, the outcome figures are anonymized, and the site does not say which industries outside healthcare the models serve.

Because models are fitted to each role and organization, coverage is really a question about where enough outcome history exists.