Skip to main content
Signal research examines whether new evidence changes a previously supported assessment. Monitoring must distinguish economic deterioration from normal variation, corrected records, and lost visibility. It identifies reviewable signals; institution staff determine intervention in their own systems.

​
Signal construction

​
Temporal design

Maintain a versioned baseline and a durable change sequence. Recompute only affected features while preserving sufficient historical context. A corrected old record becomes newly known evidence; it must not silently rewrite a historical backtest. Evaluate detection time using the actual time the system could have known the new information. Use trailing windows, seasonal comparisons, robust change-point methods, and explicit quality gates. Select thresholds on development data according to false-alert cost and review capacity. Apply documented grouping and cooldown rules so repeated evidence does not generate a misleading number of “detected events.” Review suppression logic for missed persistent problems.

​
Lead-time study

Predefine the eventual event, such as a specified delinquency, and the time at which an incumbent process first identifies it. Compare methods at matched sensitivity or matched false-alert burden. Measure paired lead time for jointly detected events and report detection coverage for the full eligible cohort. Include events never detected and right-censored histories instead of dropping them. Orders sometimes precede settlement by construction, but this does not prove a universal two-to-four-week risk advantage. Estimate lag distributions per source and product. Follow the integration refresh guidance: provider webhooks, AMLOK processing schedules, and business-event visibility are different clocks.

​
Metrics

Report event-level precision and recall, false alerts per 100 monitored borrower-months, alerts per analyst-hour, median and tail lead time, repeated-alert burden, stale-data coverage, and eventual outcomes. Distinguish signal generation, analyst acknowledgement, intervention, and loss. An alert associated with default does not establish that intervention would have prevented default. Use a prospective shadow phase before controlled intervention evaluation. Preserve existing institution safeguards and record interventions as time-varying factors. No experimental withholding of required protection is proposed.

​
Product correspondence

Signal includes monitoring and change-driven reassessment with existing and new evidence; a detected event does not launch a separately charged Insights underwriting. The account-month billing unit remains one borrower/business per workspace. Newly processed source versions count toward workload, while unchanged baseline data and retries do not count again. Billing rules are independent of statistical event counts. Cancellation stops processing immediately. A research collection or continuous-learning loop cannot override cancellation, consent, retention, or workspace isolation. Model updates follow a separate reviewed release process rather than learning automatically from every production event.