PULSE
A working prototype for KRI intelligence and early warning: an indicator registry, statistically backtested thresholds, and a signal lab that grades which indicators actually lead losses.
The problem
Ask a bank how its key risk indicators are doing and you will get a dashboard. Ask which of those indicators has ever actually led a loss, by how many months, with what precision, and the room goes quiet. In most institutions the KRI library is folklore: indicators inherited from a framework project years ago, labelled "leading" because someone asserted it in a workshop, with thresholds set by anecdote and never revisited. The board sees green, amber, and red. Nobody can say what evidence stands behind the colors.
The supervisory texts assume better. Basel's Principles for the Sound Management of Operational Risk expect monitoring and reporting built on indicators that actually inform. Appetite statements cascade to metrics whose thresholds are supposed to mean something. An early-warning system whose warnings have never been tested against loss history is not an early-warning system. It is decoration with an audit risk attached.
What it does
PULSE (Platform for Unified Leading-Signal Evaluation) treats a KRI as what it actually is: a hypothesis that deterioration in some metric precedes losses. Then it tests the hypothesis. Seven modules cover the lifecycle:
Command Center. The enterprise view: active indicators, quality flags, open breaches with SLA countdowns, movers, and appetite posture, with a Basel category by business line heat map that drills through to the underlying indicators.
Indicator Registry. The library as a governed inventory: ownership, lineage, data quality, and review cadence, not a spreadsheet tab.
Threshold Studio. Statistical calibration with backtesting. Quantile, SPC, EWMA, and volatility-adjusted methods run side by side against the indicator's own history, showing expected amber and red episodes per year, backtest performance, and warnings where a method's assumptions fail the series. Thresholds become instruments, not folklore.
Signal Lab. The heart of the build. Every indicator that claims to be leading is tested against the loss series across multiple lags and graded on evidence: precision, recall, F1, and best lead time, with an explicit materiality floor so that multiple testing does not manufacture significance. The result is a ranked library and, usually, a scandal quadrant: indicators that claim "leading" on the board pack but show no lead evidence at all.
Monitoring & Breaches. Breach episodes with entry and exit rules, SLA tracking, and second-line workflow rather than a traffic-light snapshot.
Appetite & Governance. Board appetite metrics cascade to the indicator library, so the connection between the appetite statement and the operational metrics is explicit and inspectable.
Audit Trail. Every calibration, proposal, approval, and override logged as it happens.
The prototype ships with the same demo tenant as DELPHI: Meridian Bancorp, a synthetic mid-tier bank with 60 months of history, realistic roles, and a banner making clear that nothing in it is real data. It also ships with an "honest limitations" panel on its own results: multiple testing, short history, and the difference between correlation and causation, stated plainly inside the tool.
Why it matters
KRIs sit at the join between risk appetite and daily operations. If the indicators do not lead, the appetite cascade is theatre and the board is watching a lagging speedometer. Grading the library on evidence changes the conversation: which indicators to keep, which to recalibrate, which to retire, and which claims should never have reached a board pack. That is the difference between monitoring and early warning, and it is what the supervisory expectation has always actually meant.
Honest framing
PULSE is a working prototype and a blueprint, not a product. The statistics are deliberately conservative and the tool says so on screen: grades are decision support for validation and retirement, not proof of mechanism. It exists to demonstrate that an evidence-graded KRI library is buildable now, and to document the design decisions that make one defensible.
The live prototype runs in the browser with synthetic data. Open it, take the guided tour, and try to break the reasoning. That is what it is for.