Skip to main content

ADR-0048: Financial Analysis Agent

Status: Accepted Date: 2026-04-22 Supersedes: none Superseded by: none

Context​

Compliance officers evaluating corporate customers need to form a view on the target's financial health as part of EDD. Until 2026-04-22 the pipeline produced a financial_health_report (line items + ratios from NBB CBSO, BODACC, or sbírka listin) and a single categorical label ("stable" / "filings_only"). Officers were left to interpret the raw numbers themselves — which is exactly the kind of high-effort, inconsistent judgement regulatory supervisors expect automated systems to reduce.

AMLR Art. 28(2)(d) (enhanced due diligence) and the EBA risk-factor guidelines require that financial risk signals be weighed during onboarding and ongoing monitoring. Without a structured analytic layer the platform was exporting "here's what the registry said" rather than "here's what it means" — the latter is what a supervisor expects.

Decision​

Introduce a dedicated Financial Analysis Agent that runs after the registry/NorthData step and before the investigation synthesis step. It ingests a financial_health_report and the target's NACE codes and produces a FinancialAnalysisReport plus a list of typed Finding objects that flow into the standard findings array.

The agent composes five orthogonal analytic stages:

  1. Ratio extraction (ratios.py) — liquidity, leverage, efficiency, profitability, coverage. Pure arithmetic on FHR line items.
  2. Distress prediction (distress_models.py) — three orthogonal models drawn from the peer-reviewed literature:
    • Altman (1968) Z, Altman (1983) Z′, Altman (1995) Z″ — three variants for public manufacturers, private manufacturers, and service/non-manufacturers respectively.
    • Ohlson (1980) O-score — logistic regression on nine accounting ratios.
    • Zmijewski (1984) — simpler three-ratio logistic. Each model emits a DistressVerdict or None when inputs are insufficient. Disagreement across models is itself surfaced as a finding.
  3. Trend analysis (trend_analysis.py) — year-over-year signals on revenue, equity, cash, working capital. Flags inversions and step-changes.
  4. Anomaly detection (anomaly_detection.py) — statistical outliers vs. the entity's own history.
  5. Peer benchmarking (peer_benchmarks.py) — ratio comparison against a NACE-code cohort.

Findings are emitted with severity (critical / high / medium / low / verified) and mapped to canonical regulatory citations (AMLR, EBA guidelines, CNB vyhlášky for CZ entities) so an officer can defend the decision during an inspection.

Architectural constraints​

Deterministic-by-design. None of the five stages calls an LLM. Every output carries its source FHR fields and the formula that produced it; the audit trail can replay the computation byte-identically from stored inputs. This is non-negotiable for EU AI Act Art. 12 (automatic logging) and Annex IV technical documentation.

Decoupled from the synthesis agent. The Financial Analysis Agent does not narrate — it produces typed structure. A later synthesis step (the PydanticAI agent fleet) is free to quote the structured output back in natural language, but only the raw report is load-bearing for the decision record.

Degrades gracefully. When the FHR has only metadata (no line items), the analyzer returns a report with going_concern_risk="insufficient_data" and an empty findings list rather than raising. The pipeline treats absence of data as a known state, not as an error — consistent with ADR-0046 (NBB CBSO honest data-gap surfacing).

Why these specific distress models​

  • Five models, not one. No single bankruptcy-prediction model is robust across industries, sizes, and accounting regimes. Running three orthogonal families (Altman, Ohlson, Zmijewski) and surfacing agreement/disagreement gives officers a calibration signal rather than a single false-precision number.
  • All peer-reviewed, all deterministic. Every formula has a citation trail (Altman 1968, 1983, 1995; Ohlson 1980; Zmijewski 1984). Regulators ask "why did you decide the target was distressed?" — pointing at a 1968 Journal of Finance paper is a stronger answer than pointing at an LLM.
  • No commercial ML models. Paid distress-prediction APIs (Moody's RiskCalc, CreditModel) are out of scope for the PoC — they add licence cost and audit complexity without adding explanatory power.

Consequences​

  • Pipeline position. New stage inserted between registry/NorthData and synthesis. Downstream consumers (risk matrix, chatbot, PDF conclusion) already consume typed Finding objects, so the integration cost is zero beyond wiring the new agent in.
  • UI surface. Financial Analysis findings are rendered in the officer dashboard tile (Overview tab) and as a dedicated panel on the case detail page.
  • Chatbot integration. The chatbot's read tools expose the FinancialAnalysisReport structure so officers can ask "what was the Altman Z score on the 2024 filing?" and get a factual answer with the exact formula quoted.
  • Test coverage. 34 unit tests across tests/test_financial_analyzer.py, tests/test_distress_models.py, tests/test_peer_benchmarks.py.

Alternatives considered​

Single-model approach (Altman Z only). Simpler, but brittle — private-company FHRs don't always have the inputs Altman Z-1968 needs, and a single number offers no calibration signal.

LLM-narrated verdicts instead of typed structure. Rejected. An LLM cannot be replayed byte-identically, and an officer cannot cite "the language model said so" when defending a decision to a supervisor.

Defer to OSINT synthesis agent. Rejected. That agent's job is natural-language narrative construction, not numeric analysis. Mixing the two obscures provenance and violates the EU AI Act Art. 12 logging requirement (every numeric output must have a reproducible source).

References​

  • AMLR Art. 28(2)(d) — enhanced due diligence
  • EBA/GL/2023/03 — risk-factor guidelines §4.9 (financial indicators)
  • Altman, E. I. (1968). Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy, Journal of Finance
  • Ohlson, J. A. (1980). Financial Ratios and the Probabilistic Prediction of Bankruptcy, Journal of Accounting Research
  • Zmijewski, M. E. (1984). Methodological Issues Related to the Estimation of Financial Distress Prediction Models
  • Related: ADR-0020 (EBA risk matrix), ADR-0029 (cost-optimized model tiers), ADR-0046 (NBB CBSO honest data-gap surfacing)