TEMPEST
A working prototype for operational loss projection under stress: frequency and severity modeling, a 50,000-iteration simulation engine, legal reserves, and results built for the CCAR submission cycle.
The problem
Every stress-testing cycle, banks project nine quarters of operational losses, and in most institutions the machinery behind that number is a patchwork: a regression fitted years ago, a spreadsheet that splices frequency and severity assumptions nobody can fully reconstruct, and a legal reserve overlay negotiated in meetings that leave no trail. The number reaches the capital plan; the method would not survive a determined examiner reading it end to end.
The supervisory bar is explicit. The Federal Reserve’s capital-planning letters (SR 15-18 for the largest firms, SR 15-19 for the next tier) expect estimation approaches whose assumptions are documented and supportable, with challenge and governance around them. Op-loss projection is routinely the weakest link in that chain: material to the result, thin on evidence.
What it does
TEMPEST (Tail Event Modeling and Projection Engine for Stress Testing) rebuilds the projection pipeline end to end, as connected modules:
Data Foundation. Governed loss data in, with generators and browsable history, so every downstream estimate traces to its inputs.
Frequency Lab. Candidate frequency models fitted and compared, champion selection surfaced with the diagnostics rather than buried in a model file.
Severity Lab. Spliced severity fits with explicit body and tail treatment, and a stabilization policy so tail estimates do not whipsaw between cycles.
Simulation Engine. A 50,000-iteration Monte Carlo engine that runs in the browser in seconds, producing the nine-quarter loss paths and the economic-capital lens on the same governed inputs.
Legal Module. Asserted and unasserted legal exposure handled as its own pipeline, because litigation is where op-loss projections most often lose credibility.
Fed Lens. A challenger view that replays the supervisory-style benchmark against the internal model, with the waterfall from one to the other made explicit.
Results and governance. Output shaped for the FR Y-14A reporting context, with backtests, approvals, and a submission bundle export, so the governance is the workflow rather than a memo written afterwards.
The prototype ships with a synthetic Category IV bank profile so the full cycle, data to submission bundle, can be walked in the browser with nothing real at stake.
Why it matters
The op-loss projection is one of the few operational risk numbers that directly moves capital. A process that produces it from spreadsheets and memory is a finding waiting to happen; a process that produces it from governed data, compared models, explicit challengers, and logged approvals is a capital-planning asset. The difference is not sophistication. It is reconstructability.
Honest framing
TEMPEST is a working prototype and a blueprint, not a product. The models are deliberately standard; the contribution is the pipeline around them: every estimate traceable, every choice challenged, every output reproducible. It exists to demonstrate that submission-grade governance of op-loss projection is buildable now, in a browser, by one practitioner who has sat through the cycle.