Financial Analysis Agent

A dedicated pipeline stage that turns a raw financial_health_report (line items from NBB CBSO, BODACC, sbírka listin, etc.) into a structured FinancialAnalysisReport plus a list of typed Finding objects. Runs after the registry/NorthData fallback and before the synthesis agent. See ADR-0048 for the full architectural rationale.

Why this exists

Compliance officers need more than raw balance-sheet numbers — they need a view. Until this layer existed, the platform exported "here is what the registry said" and left interpretation to the officer. AMLR Art. 28(2)(d) and the EBA risk-factor guidelines expect automated EDD to actually weigh financial signals, not just relay them.

The five analytic stages

Every stage is pure Python, deterministic, and byte-identically replayable from stored inputs — an EU AI Act Art. 12 requirement.

1. Ratio extraction (ratios.py)

Liquidity, leverage, efficiency, profitability, coverage. Pure arithmetic over FHR line items. No thresholds applied here — the downstream stages consume the ratios.

2. Distress prediction (distress_models.py)

Five models across three orthogonal families, drawn from the peer-reviewed literature:

Model	Reference	Use case	Distress threshold
Altman Z	Altman (1968), J. Finance	Public manufacturers	Z < 1.81
Altman Z′	Altman (1983)	Private manufacturers	Z′ < 1.23
Altman Z″	Altman (1995)	Service / non-manufacturer / emerging markets	Z″ < 1.1
Ohlson O	Ohlson (1980), J. Acct. Research	Logistic regression on 9 ratios	O > 0.5
Zmijewski	Zmijewski (1984)	Simpler 3-ratio logistic	Z > 0.5

Why five models, not one. No single bankruptcy-prediction model is robust across industries, sizes, and accounting regimes. Running three orthogonal families and surfacing agreement/disagreement gives officers a calibration signal rather than a single false-precision number. When models disagree, the disagreement itself becomes a finding.

3. Trend analysis (trend_analysis.py)

Year-over-year signals on revenue, equity, cash, working capital. Flags inversions (equity turning negative) and step-changes (>50% shifts year over year).

4. Anomaly detection (anomaly_detection.py)

Statistical outliers vs. the entity's own history. Distinct from peer benchmarks (which compare against industry); anomaly detection compares against the entity's own baseline.

5. Peer benchmarking (peer_benchmarks.py)

Ratio comparison against a NACE-code cohort. Surfaces entities that deviate meaningfully from industry norms — a common indicator of mis-reported financials or business-model divergence.

Output shape

FinancialAnalysisReport(
    going_concern_risk: Literal["low", "medium", "high", "critical", "insufficient_data"],
    ratios: FinancialRatios,
    distress: DistressScores,          # All 5 models, each with verdict or None
    trends: TrendSignals,
    anomalies: list[Anomaly],
    peer_comparison: PeerComparison,
    findings: list[FinancialFinding],   # Typed, severity-tagged, citation-linked
)

Each finding carries a canonical regulatory citation (AMLR article, EBA guideline, CNB vyhláška for CZ entities) so an officer can defend the decision during an inspection.

Graceful degradation

When the FHR has only metadata (no line items) — e.g., the NBB CBSO CSV endpoint is degraded (see ADR-0046) — the analyzer returns a report with going_concern_risk="insufficient_data" and an empty findings list. The pipeline treats absence-of-data as a known state, not an error.

Chatbot integration

The chatbot's read tools expose the FinancialAnalysisReport structure, so officers can ask questions like "what was the Altman Z-score on the 2024 filing?" and get a factual answer with the exact formula quoted — not an LLM-hallucinated number.

Why deterministic, not LLM

An LLM cannot be replayed byte-identically. An officer cannot cite "the language model said so" when defending a decision to a supervisor. All numeric verdicts stay in code where they're auditable; a future LLM-narrated layer on top is possible (e.g. industry-context reasoning about why a ratio deviation matters for a specific NACE code) but will never touch the numbers.

References

• ADR-0048 — architectural decision
• AMLR Art. 28(2)(d) — enhanced due diligence
• EBA/GL/2023/03 §4.9 — financial indicator guidelines
• Altman, Ohlson, Zmijewski — see ADR-0048 for full citations