PD Model vs S&P / Moody's / Fitch Long-Term Ratings

Numbers in this document are from the Q1 2026 vintage scoring run. Refreshed each quarter alongside the model rescore.

Headline

Our PD model and agency long-term ratings are not measuring the same thing, and the difference is the whole point of looking at both. Our model estimates a bank's standalone probability of default from its own balance sheet, with no credit for government or parent support. Agency long-term ratings deliberately blend that standalone view with support uplift: the too-big-to-fail presumption, sovereign backstops, and parent-to-subsidiary lift. So the question is not "do we match the agencies," it is "where we differ, is the gap explained by the support assumption?" For a counterparty-risk decision, that gap is exactly what you want surfaced rather than hidden.

Read through that lens, the two line up closely. On the 78 banks with a directly comparable long-term agency rating, our model sits at median 0 notches versus S&P, lands within ±2 notches on 81% of S&P pairs (65% Moody's, 69% Fitch), and within ±3 on 92% versus S&P. The divergences fall into two predictable groups, both by design:

• More conservative on the largest banks: where the agencies add the most support uplift, we add none, so our grade lands several notches lower. If you are sizing credit limits, this is the deliberately cautious read: price the exposure as if support might not arrive.
• More constructive on healthy mid-caps: where the agencies apply through-the-cycle caution, our model reads current capital, Texas Ratio, and earnings, and lands these banks higher.

Detail below in "Where the agencies and our model diverge."

Method

Source data: a manually curated external-ratings file covering 90 entities (49 bank-level, 41 BHC-level). For each entity × agency we pick the long-term rating via a priority list: Long-Term Issuer → Long-Term Deposits → Senior Unsecured → Senior Debt → Long-Term. Short-term, subordinated, and preferred-stock ratings are excluded.

Mapping to bank-level scoring:

• Bank-level ratings (49 entries) join directly to the PD bundle by FDIC certificate number.
• BHC-level ratings (41 entries) are expanded to constituent bank subsidiaries via the Federal Reserve NIC top-tier holder mapping, then the agency rating is improved by one notch before comparison. This reflects the standard agency-rating convention that bank subsidiaries are usually rated one notch above their parent holding company - FDIC-insured deposits make the bank entity safer than the holding company that owns it.

Notch scale: S&P's 22-tier scale (AAA = 0, D = 21). Moody's ratings are mapped to S&P-equivalent notches (Aaa→AAA, Aa1→AA+, etc.) for the alignment math. Our model's pd_letter uses the same S&P scale.

Result: 78 bank rows in the comparison set after BHC expansion + de-duplication. 4 BHC-rated entries had no bank subsidiaries in our PD bundle (investment-only BHCs or otherwise out-of-scope).

Notch alignment

Read: PD model aligns most closely with S&P (median 0 notch diff). Versus Moody's and Fitch, our model is systematically ~1–2 notches more bearish. None of the three agencies disagree wildly - 92% of banks land within ±3 notches of S&P, 82% within ±3 of Fitch.

Why S&P is closest: S&P's bank ratings have historically focused on stand-alone credit profile (SACP) without the government / parent uplift Moody's and Fitch apply more aggressively. Our PD model, trained on FDIC failures, sees roughly the same bank-only picture as S&P.

Where the agencies and our model diverge

The divergences are directional and structural, not random. They cluster into two segments, and within each segment the gap runs the same way for the same reason. Counts, median notch gaps, and ranges below are from the Q1 2026 comparison set.

Read: the model is systematically more cautious exactly where agencies lean hardest on support assumptions, and more constructive exactly where agencies apply the heaviest through-the-cycle discount. On the conservative side the agencies are arguably right on the specific names, since those banks would not be allowed to fail; but for a counterparty-risk product the standalone read is the defensible one: if you are underwriting credit limits, size them as if government support might not materialize.

Calibration anchor - mean model PD per agency long-term rating

This is the alignment table that goes into the methodology / MRM file: for each agency letter bucket, what's the mean PD our model produces, and how does that compare to Moody's idealized 1-year PD for that letter? Ratios shown.

S&P long-term rating

Moody's long-term rating

Fitch long-term rating

Read across all three tables: the ratio of model PD to Moody's idealized PD is huge at the safe end (50×–250× at AA-tier) and converges to ~1× around Baa2 / BBB - exactly the inflection point predicted by the methodology doc §15.2.

Two reasons this gap exists at the safe end:

1. Bank failures don't happen at the rates implied by corporate-bond AAA/AA defaults. Moody's idealized AA-tier 1-year PD of 0.3 bps reflects long-dated investment-grade corporate-bond default frequency. The bank universe doesn't have AA-equivalent failure rates because no FDIC-insured bank fails at 0.003% per year - the floor for "well-capitalized regulated bank" failure is empirically around 5–30 bps annually. Our model is calibrated to that floor; the agencies' idealized curve assumes a corporate-bond floor that doesn't exist in our population.

2. Agency ratings factor in implicit government / parent support. Moody's specifically uses a "supported rating" methodology that adds notches based on assessed sovereign + parent support. For large banks the lift is often 2–4 notches. Our PD model captures bank-standalone risk without that adjustment. This is by design - for counterparty risk we want to know failure probability assuming support might not materialize.

Why this isn't a calibration problem:

The methodology doc explicitly anchors v1.2's calibration to realized FDIC failure rates, not to agency idealized PDs. The risk_tier field (Severe / High / Elevated / Moderate / Low) was added in v1.2 precisely because agency PD anchoring isn't the right calibration target for bank failure prediction. The ratio table above documents the gap honestly; it doesn't argue that the gap is "wrong."

Coverage limitations

1. 78 banks out of ~9,300 in the scoring universe (~0.8%). This analysis applies only to the rated subset. Bank ratings are concentrated in the megabank + mid-cap public BHC universe; most community banks never get rated. One of our product's value propositions is scoring the 99%+ of banks the agencies don't.
2. BHC-level → bank-level mapping is approximate. We apply +1 notch as the standard convention, but it's a simplification. Some BHC structures have ratings that don't translate cleanly, the holding company and its lead bank subsidiary can sit a notch or two apart and the precise subsidiary rating is not always published. Where bank-level S&P/Moody's/Fitch ratings are available, we use them directly without adjustment.
3. The "long-term rating" type per agency is normalized. Different agencies use different naming conventions (Long-Term Issuer / Long-Term Deposits / Senior Unsecured / Senior Debt). We pick the first matching type per agency via a priority list; for some entities the priority resolves to slightly different rating concepts. Not a methodology problem but worth flagging when you're explaining specific banks' rating diffs.
4. Rating action staleness. Agency ratings have an as-of date; some are 6 - 18 months old at any given time. Our PD model refreshes quarterly. Some divergences will narrow at the next round of rating action.

Further reading

• PD Model Methodology - the foundational doc; the calibration choice discussed above (anchoring to realized FDIC failure rates rather than agency idealized PDs) is covered in the calibration section.
• Letter Distribution Explainer - why ~76% of banks land in A+/A/A- on our letter scale and why corporate-bond rating universes look different.

For a deeper validation pack (the full 78-row comparison table, per-quarter trend, calibration plot files, and a model-risk-management pack covering SR 11-7 alignment), contact info@counterpartyinsight.com.

Not a credit rating. Counterparty Insight is not a credit rating agency and is not registered as a Nationally Recognized Statistical Rating Organization (NRSRO) under Section 15E of the Securities Exchange Act. The letter grades and probability-of-default estimates referenced here are independent, quantitative analytical assessments for informational use only. They are not "credit ratings" as defined under the Credit Rating Agency Reform Act, are not a recommendation regarding any security or credit decision, and must not be used as the sole basis for any investment, lending, or counterparty decision. The comparison to agency ratings is for benchmarking and is not a representation of equivalence; the letter scale is used for interpretability and is not affiliated with, endorsed by, or derived from any rating agency.