FinObservatory / About
About FinObservatory
FinObservatory is an open financial-economics observatory: two centuries of financial crises, the balance sheet of every insured US bank, sovereign debt, credit cycles, cross-border banking and financial-crime reference data, unified on one platform where every figure traces to a primary source.
It is built by one person, working with AI agents, to one standard: a page may only show a number the build itself computed from the publisher’s data, and where the data cannot support a claim, the page says so.
The organizing view
The platform is organized the way the field is organized. Financial economics studies claims on future, uncertain consumption; its central objects are the price of risk, the institutions that intermediate it, and the recurring failure of the credit that funds it. The modules map onto that structure. Conditions tracks the time-varying price and availability of credit. Banks and cross-border banking cover the intermediaries, whose funding of illiquid assets with demandable claims is the fragility Diamond and Dybvig (1983, Journal of Political Economy) formalized. Systemic measures the comovement that turns one institution’s distress into everyone’s. And Crises treats two centuries of banking, currency and sovereign-debt crises as panel evidence that the failure mode is endemic, not accidental. The empirical spine follows the credit-cycle literature: what credit does before, during and after the break.
How it is built
Almost every page is prerendered: the numbers are computed in DuckDB over a parquet estate at build time, in the same TypeScript functions the pages render from. There is no CMS. The data catalog names each exportable dataset’s publisher, licence and coverage, with the coverage queried from the live data when the page is built, and lists the 18 research-licensed datasets that are analyzed but never redistributed, each with its reason stated.
The discipline is enforced, not promised. Computed results are audited by independent recomputation before they ship. Claims that data is absent are proved by query, because an absence claim is still a claim. A scripted quality gate blocks the failure modes this codebase has actually hit: silently empty renders, dead queries, licence regressions, and numbers typed into prose that should have been queried. The exact series selections and derivations behind each module are on its methodology page.
The analyst
The analyst is the strictest expression of that standard: the language model never generates a number. 54 fixed tools call the same functions the pages are built from; the model chooses tools, fills in their parameters, and writes prose around the figures the tools return. If no tool can supply a number, the analyst says so rather than inventing one.
The evaluation is published, not summarized: 73 questions scored against ground truth the harness recomputes from the estate at run time, including eleven fabrication attacks that must be answered as absence and six out-of-scope questions that must be refused with a reason. The latest full live pass (2026-07-17) scored 73 of 73. The question list, the scoring rules, and the honest history of harness fixes are on the analyst methodology page.
A tour, with the literature
The Schularick-Taylor (2012, American Economic Review) credit-boom result, re-derived: an event study across 84 advanced-economy banking-crisis onsets (Jorda-Schularick-Taylor Macrohistory) and 142 emerging-market onsets (Laeven-Valencia), every cell recomputed at build time, with the conditioning problem stated rather than hidden: most elevated credit readings are not followed by a crisis, and the page leads with that base rate.
CoVaR (Adrian and Brunnermeier 2016, American Economic Review), marginal expected shortfall (Acharya, Pedersen, Philippon and Richardson 2017, Review of Financial Studies) and its long-run form LRMES (Brownlees and Engle 2017, Review of Financial Studies), and an absorption-ratio comovement measure, computed for US banks with the estimation choices documented. LRMES is never presented as full SRISK: the dollar capital-shortfall level needs balance-sheet data the estate's licences cannot supply, and the methodology says so.
A composite health score for every active FDIC-insured bank in the spirit of supervisory CAMELS ratings, built only from public Call Report fields, on a published formula, explicitly disclaimed as not CAMELS. The out-of-sample record is stated up front: the score missed Silicon Valley Bank, whose unrealized losses are invisible in the public fields, and flagged First Republic.
A grounded AI analyst built on a strict division of labor: the model plans queries and writes prose; the estate supplies every number. It cites the source of each figure, refuses questions outside the estate, and answers questions about data that does not exist by saying the data does not exist.
Comparative financial structure in the World Bank GFDD tradition, joined with IMF Financial Soundness Indicators, the BIS credit aggregates and Global Findex, unified on one page per economy with each layer's coverage and staleness stated, so a missing series reads as the source's gap, not a zero.
Replications against the estate with the sample differences stated: Schularick and Taylor (2012), Jorda, Schularick and Taylor (2015, Journal of Monetary Economics), Mian, Sufi and Verner (2017, Quarterly Journal of Economics), Kaminsky and Reinhart (1999, American Economic Review), Frankel and Rose (1996, Journal of International Economics), Reinhart and Rogoff (2013, Journal of Banking and Finance). A failed replication is a headline, not a footnote.
Who builds this
FinObservatory is designed, engineered and verified by Md Deluair Hossen, a PhD economist who builds AI-augmented data platforms for the two fields he works in: financial economics here, and trade economics at TradeWeave. The method is the same on both: primary sources, computed rather than typed numbers, published methodology, and AI that is safe to trust by construction rather than by promise.