FDIC Portfolio Surveillance

Skill Type: RIGID

Follow every phase in order. Do not skip phases. Do not produce output before completing Phase 3.

When to Activate

Activate when the user asks to:

• Screen, triage, or surveil a group of institutions
• Build a watchlist for a state, asset tier, or peer cohort
• Rank banks by risk, health, or deterioration across a portfolio
• Identify which institutions in a universe warrant immediate attention
• Compare emerging risk signals across a population of banks

Do not activate for single-institution deep dives — use the fdic-bank-deep-dive skill instead.

Inputs

Dependency Table

Phase 1 — Universe Construction

Step 1: Build the institution roster

Call fdic_search_institutions with the user's universe definition.

• State-based: filters: "STALP:<XX> AND ACTIVE:1", fields: "CERT,NAME,ASSET,DEP,STALP,BKCLASS,ACTIVE", sort_by: "ASSET", sort_order: "DESC", limit: 500
• Asset-based: filters: "ACTIVE:1 AND ASSET:[<min> TO <max>]", same fields and sort
• CERT list: filters: "CERT:(<cert1> OR <cert2> OR ...)", same fields

Record the total count and the list of CERTs. This is the screening universe.

If the tool returns zero results or errors, hard-stop and report the failure to the user.

Step 2: Derive date parameters

If the user did not provide repdte, omit it from downstream calls (tools default to the most recent published quarter). If the user did not provide start_repdte, omit it (snapshot comparison defaults to one year prior).

Record the effective repdte and start_repdte from the first tool response that returns them. These are the date basis for the report and must be stated in the output header.

Phase 2 — Risk Screening and Ranking

Run the following three calls. They may be executed in parallel since they are independent.

Step 3: Detect risk signals

Call fdic_detect_risk_signals with the universe parameters.

• State-based: state: "<XX>"
• Asset-based: asset_min, asset_max
• CERT list: certs: [<cert1>, <cert2>, ...]

Set limit to at least the universe size (up to 100) so all flagged institutions are captured. Set min_severity: "warning".

Record for each flagged institution: CERT, name, critical count, warning count, signal codes, and the proxy summary band.

Step 4: Compare peer health

Call fdic_compare_peer_health with matching parameters.

• State-based: state: "<XX>"
• Asset-based: asset_min, asset_max
• CERT list: certs: [<cert1>, <cert2>, ...]

Set limit to at least the universe size (up to 100). Default sort_by: "composite".

Record for each institution: CERT, name, proxy score, proxy band, composite rating, component ratings (C/A/E/L/S), and flags.

Step 5: Compare snapshots for trend confirmation

Call fdic_compare_bank_snapshots with matching parameters.

• State-based: state: "<state name>" (full name, not abbreviation)
• Asset-based: The snapshot tool does not accept asset_min / asset_max directly. Use the CERT list collected from Step 1 instead: certs: [<cert1>, <cert2>, ...]. If the CERT list exceeds 100, truncate to the top 100 by asset size and note the truncation in the report.
• CERT list: certs: [<cert1>, <cert2>, ...]
• Provide start_repdte and end_repdte if the user specified them.

Use sort_by matching the user's ranking emphasis:

• overall or unspecified → asset_growth_pct (DESC)
• capital → roe_change (ASC, to surface capital erosion)
• earnings → roa_change (ASC)
• funding → deposits_to_assets_change (ASC)
• growth → asset_growth_pct (DESC)

Set limit to at least the universe size (up to 100).

Record for each institution: asset growth %, deposit growth %, ROA change, ROE change, and insight tags.

Phase 3 — Triage and Categorization

Step 6: Merge and classify

Merge the results from Steps 3–5 by CERT. Classify each institution into one of three tiers:

Escalate — Any institution meeting one or more of:

• Any critical-severity risk signal (capital undercapitalized, earnings loss, reserve coverage low)
• Proxy band of high_risk or weak
• Composite rating of 3 or worse
• Three or more warning-severity signals
• Persistent adverse trend: ROA declining AND asset growth negative or flat

Monitor — Any institution meeting one or more of:

• One or two warning-severity signals without critical signals
• Component rating of 3 or worse in any single CAMELS component
• Rapid asset growth (>15% over the comparison period) combined with any warning signal
• Proxy band of satisfactory with any component rated 3+

No Immediate Concern — Institutions with:

• No critical or warning signals
• Proxy band of strong
• All component ratings 1 or 2
• No adverse trend indicators

Step 7: Build the ranked watchlist

Within each tier, rank institutions by severity:

1. Escalate: Sort by critical signal count descending, then warning count descending, then proxy score ascending (lower proxy score = weaker institution; the proxy model maps 4.0 = strong, 1.0 = high risk).
2. Monitor: Sort by warning count descending, then worst component rating descending (higher component rating = worse; CAMELS components use 1 = Strong through 5 = Unsatisfactory).
3. No Immediate Concern: Sort by proxy score descending (higher proxy score = stronger institution).

Apply the user's limit to bound the total output. If the limit is smaller than the Escalate + Monitor population, prioritize Escalate institutions first.

For each institution in the Escalate and Monitor tiers, write a short driver text (1–2 sentences) summarizing the primary reason(s) for placement. Use explicit reason codes, not opaque scores:

• "Capital erosion: Tier 1 leverage at X%, classified as undercapitalized"
• "Earnings loss: ROA at -X% with efficiency ratio at Y%"
• "Funding stress: brokered deposits at X% of total deposits"
• "Credit deterioration: noncurrent loans at X%, trending adverse for N quarters"
• "Rapid growth with weakening profitability: asset growth of X% while ROA declined Y pp"
• "Repeated adverse trend signals: [list domains]"

Phase 4 — Targeted Follow-Through (Escalated Institutions Only)

Limit follow-through calls to the top 3 most material Escalate-tier institutions by default, or fewer if the Escalate tier is smaller. The user may override this limit.

Step 8: Detailed health assessment

For each targeted institution, call fdic_analyze_bank_health with cert: <cert> and repdte: <repdte> (using the effective report date established in Step 2). If the user did not specify a repdte, omit it to let the tool default — but all follow-through calls must use the same date basis as the screening workflow.

If the call fails, note "Detailed assessment data unavailable for [name] (CERT <cert>)" and preserve the institution's watchlist placement.

Step 9: Domain-specific follow-up (conditional)

Call domain-specific tools only when the institution's risk signals implicate that domain. Pass cert and repdte (the same effective report date) to each call:

• Funding signals present (funding_stress, wholesale_funding_elevated, funding_ltd_stretched) → call fdic_analyze_funding_profile with cert, repdte
• Credit signals present (credit_deterioration, credit_deterioration_trending, reserve_coverage_low) → call fdic_analyze_credit_concentration with cert, repdte
• Earnings signals present (earnings_loss, earnings_pressure, margin_compression) → call fdic_ubpr_analysis with cert, repdte

If a domain tool fails, note "[domain] follow-up could not be completed for [name]" and preserve the triage result.

Step 10: Regional context (optional)

Call fdic_regional_context once for the universe (using state or the first institution's CERT), passing repdte if established.

If unavailable, omit the Regional Context section or write: "Regional economic context unavailable; macro conditions were not factored into institution-level rankings."

Phase 5 — Report Assembly

Produce the report in the following fixed section order:

Section 1: Universe Definition

State the screening universe parameters, institution count, and effective date basis.

Universe: [state / asset range / CERT list description]
Institutions screened: [N]
Financial data as of: [REPDTE in YYYY-MM-DD format]
Comparison period: [start_repdte] to [end_repdte]

If the report mixes data from different date bases (e.g., quarterly financials and annual SOD data), state each basis explicitly.

Section 2: Screening Summary

Provide a brief quantitative overview:

• Total institutions screened
• Institutions with critical signals
• Institutions with warning signals
• Institutions with no signals
• Distribution across Escalate / Monitor / No Immediate Concern tiers

Section 3: Ranked Watchlist

Present the watchlist as a structured table or grouped list:

Escalate

Monitor

Same structure.

No Immediate Concern (include only if the user requests full population coverage or the limit allows)

Same structure, abbreviated.

Section 4: Escalated Institution Follow-Through

For each institution that received follow-through analysis in Phase 4, provide a concise summary:

• Health assessment highlights (composite rating, worst components, trend direction)
• Domain follow-up findings (if applicable)
• Recommended next step: "Full deep-dive analysis recommended" or "Continue monitoring with [specific focus]"

If follow-through was limited to a subset, state: "Follow-through analysis was performed for the top [N] escalated institutions. Remaining escalated institutions retain their watchlist classification pending further review."

Section 5: Caveats / Date Basis

Always include:

1. Date basis statement: "Financial data sourced from FDIC Call Reports as of [REPDTE]. Snapshot comparison covers [start] to [end]."
2. Proxy disclaimer: "All health scores and component ratings are derived from the public_camels_proxy_v1 analytical engine using public FDIC data. These are not official CAMELS ratings and do not reflect confidential supervisory information."
3. Peer limitation (if applicable): If the peer set was thin (<10 institutions) or partially mismatched, state: "Peer comparisons are limited due to [reason] — interpret relative rankings with caution."
4. Data staleness (if applicable): If the report date is more than 120 days old, note this explicitly.
5. Regional context gap (if applicable): Note if macro context was unavailable.

Output Rules

• Tone: Supervisory-safe, neutral, decision-oriented. Describe public-data signals and triage implications. Do not imply official CAMELS ratings or confidential supervisory conclusions.
• Reason codes over scores: Every institution in Escalate or Monitor must have explicit driver text. Do not rely on opaque numeric scores alone.
• Screening vs. deep-dive: This workflow identifies where attention should go next. It is not a full narrative report on every institution. Keep per-institution text concise.
• Bounded output: Default to a concise format suitable for portfolio review. Do not produce full-population detail unless explicitly requested.
• Three outcome states: Distinguish "No data," "Not applicable," and "Tool failure" per the FDIC Skill Builder rules. Never collapse these into a generic "n/a."
• Date basis transparency: If quarterly financial data and annual branch/geographic context are both referenced, state the date basis clearly for each and do not present them as if they are from the same reporting period.

Error Handling Summary