Apogee — Credit Risk Intelligence

Module 01 // Foundations

What Is Credit Risk

The probability a counterparty fails to meet a financial obligation — and the financial consequence of that failure.

Credit risk is the risk of loss from a borrower's failure to repay. It has three dimensions: the probability of failure, the amount at risk, and the fraction that cannot be recovered. All credit models quantify these three dimensions.

Credit risk is distinct from market risk (price volatility) and liquidity risk (inability to transact). They require separate analytical frameworks and capital treatments.

Corporate Lending

Banks assess whether a borrower's cash flows can service principal and interest. PD is estimated from financial ratios, industry data, and macro factors.

Bond Markets

Investors price credit spreads above the risk-free rate. Spreads embed both actual default probability and a risk premium above the actuarially fair rate.

Infrastructure Finance

Project lenders model cash flow shortfalls against debt service over 15–25 year tenors. DSCR covenants define the early warning threshold before technical default.

Orbital Asset Finance

Satellites generate revenue streams. Those streams can default if the asset fails, demand falls, or the operator is financially distressed. SFIS applies credit risk methodology at asset level.

Risk Type	Driver	Measurement	Capital Treatment
Credit Risk	Counterparty failure	PD × LGD × EAD	Basel III IRB / SA
Market Risk	Price volatility	VaR, ES	IMA / SA-MR
Liquidity Risk	Funding mismatch	LCR, NSFR	Liquidity buffer
Operational Risk	Process / system failure	AMA / BIA	Standardised approach

SFIS context: SIGMA quantifies the operational dimension of orbital asset risk. ALPHA translates that into a financial credit risk output: PD at 1, 3, and 5-year horizons. Together they bridge physical risk and financial risk for institutional audiences.

Module 02 // Foundations

The Core Formula

Every credit model traces to one identity. Master this and the rest is parameterisation.

Expected Loss Identity

EL = PD × LGD × EAD

PD: Probability of Default | LGD: Loss Given Default | EAD: Exposure at Default

Expected Loss is a statistical mean. A lender with a 1,000-loan book earns its spread on 975 loans and loses capital on roughly 25. EL is the average loss per unit of exposure. Pricing, IFRS 9 provisioning, and regulatory capital are all calibrated from this identity.

EL does not require predicting which loan defaults. It requires knowing the distribution of outcomes. Credit risk modelling is a statistical discipline, not a case-by-case prediction problem.

Pricing implication: the minimum spread equals EL / EAD in basis points. A spread below this destroys value on a risk-adjusted basis before accounting for funding costs or capital charges.

Module 03 // Foundations

Probability of Default

The most analytically intensive input. Always estimated, never directly observed.

PD is estimated from three broad methodological families. The choice of methodology determines whether the output is suitable for pricing, provisioning, stress testing, or none of the above.

Approach	Data	Best Use	Key Limitation
Historical Frequency	Cohort default rates	Large rated portfolios	Lags cycle. Needs comparable samples.
Market-Implied	CDS spreads, bond yields	Liquid corporate bonds	Risk-neutral. Embeds premium above true PD.
Structural (Merton)	Asset value, volatility	Private firms, infra	Asset value is unobservable.
Bayesian / ML	Mixed signals + telemetry	Sparse data, novel assets	Requires principled prior specification.

Market-Implied PD

PD ≈ Credit Spread (bps) / (LGD × 100)

Example: 450 bps spread, LGD = 60% → PD ≈ 7.5% (risk-neutral, not physical probability)

ALPHA engine: no liquid CDS market exists for private satellite operators. ALPHA uses Bayesian inference — orbital telemetry and financial signals update a prior built from satellite failure rate distributions. Output: posterior PD distribution at 1, 3, and 5-year horizons.

Module 04 // Components

Loss Given Default

How much is lost when default occurs. The complement of the recovery rate.

LGD Identity

LGD = 1 − Recovery Rate

Recovery Rate = Recovered Amount / EAD | LGD ranges from 0 (full recovery) to 1 (total loss)

Asset Class	Seniority	Recovery Rate	LGD
Corporate Loans	Senior Secured	65–80%	20–35%
Corporate Bonds	Senior Unsecured	35–50%	50–65%
Infrastructure Debt	Senior Secured	70–90%	10–30%
Sovereign Debt	Unsecured	25–60%	40–75%
Satellite Assets	Operator-dependent	5–40% (est.)	60–95%

Orbital LGD risk: a failed satellite cannot be repossessed. Its orbital slot may not transfer to creditors. Secondary markets for used spectrum capacity are thin or absent. Without insurance, effective LGD approaches 100%. This is the single largest structural risk in satellite lending.

Module 05 // Components

Exposure at Default

The total outstanding claim at the moment of default. Static for term loans. Dynamic for revolvers.

EAD for Revolving Facilities

EAD = Drawn Balance + CCF × (Committed Limit − Drawn Balance)

CCF: Credit Conversion Factor. Borrowers draw down revolvers under stress. Typically 0.5–0.9 under Basel III.

For amortising infrastructure loans, EAD declines as principal is repaid. Early-year EAD is high but PD may be lower as construction risk resolves. Late-year EAD is lower but cash flow risk can rise as maintenance costs increase and residual asset value declines.

Satellite loans: most are structured as term loans with scheduled amortisation. EAD is deterministic once the schedule is set. The key risk: the satellite's operational life fails to cover the loan tenor, leaving outstanding EAD against a non-functioning asset.

Module 06 // Components

Expected vs Unexpected Loss

EL is priced into spreads. UL is held as capital. Confusing the two destroys banks.

Expected Loss is the mean of the loss distribution. Lenders price it into spreads and provision for it under IFRS 9. It is a cost of doing business. Unexpected Loss is the variance around the mean. Capital is held to absorb UL at 99.9% confidence under Basel III.

Economic Capital Requirement

Economic Capital = VaR(99.9%) − EL

Capital absorbs losses exceeding EL in 1 of every 1,000 years of operation.

Credit loss distributions are right-skewed with fat tails. Most years produce losses near EL. Rare events produce losses far exceeding EL. The tail is driven by correlated defaults — borrowers failing simultaneously due to shared macro shocks.

RAROC

Risk-Adjusted Return on Capital = (Revenue − EL − OpEx) / Economic Capital. Hurdle rate typically 10–15%. Enables comparison across credit exposures.

IFRS 9 Staging

Stage 1: 12-month EL. Stage 2: lifetime EL on significant credit deterioration. Stage 3: lifetime EL on defaulted assets. Triggers tied to PD migration.

Module 07 // Models

Structural Models

Equity is a call option on assets. Default occurs when asset value crosses the debt boundary.

Merton Distance to Default

DD = [ln(V/D) + (μ − σ²/2)T] / (σ√T)
PD = N(−DD)

V: Asset Value | D: Debt Face Value | σ: Asset Volatility | N(·): Standard Normal CDF

Distance to Default measures standard deviations between current asset value and the default boundary. A firm with DD = 4 requires a 4-sigma decline in asset value to default over the horizon. KMV extended this framework using equity market data to infer V and σ empirically.

V Asset Value ($M)$200M

D Debt Face Value ($M)$120M

σ Asset Volatility25%

4.21

DD

1-Year PD

0.013%

Leverage D/V

60%

Implied Rating

AA

Equity Buffer

$80M

SIGMA application: for private satellite operators, V is unobservable. SIGMA proxies asset value through a composite of orbital health indicators (inclination drift, fuel reserves, solar panel degradation), revenue backlog, and operator financials. This structural prior feeds ALPHA's Bayesian inference layer.

Module 08 // Models

Reduced-Form Models

Default as a random arrival. No mechanism assumed. Calibrated directly to market prices.

Survival Probability (Constant Hazard Rate λ)

P(no default by T) = e^(−λT)
PD(T) = 1 − e^(−λT)

λ: hazard rate (instantaneous default intensity) | T: time horizon in years

Reduced-form models treat default as a Poisson process. The hazard rate λ(t) can be time-varying and driven by state variables: macro factors, credit spreads, industry indicators. Jarrow-Turnbull and Duffie-Singleton are the dominant frameworks, both requiring liquid market prices for calibration.

	Structural	Reduced-Form
Default trigger	Asset value crossing boundary	Exogenous Poisson process
Data requirement	Balance sheet, equity vol	CDS / bond prices
Default predictability	Predictable in theory	Unpredictable by design

SFIS position: no CDS market exists for private satellite operators. ALPHA's Bayesian inference layer functions as an analogue: each telemetry observation updates the posterior hazard rate, equivalent to continuous recalibration without market prices.

Module 09 // Models

Machine Learning Approaches

Relax parametric assumptions. Let data determine functional form. Accept the interpretability trade-off deliberately.

Logistic Regression

Baseline. Interpretable coefficients. Linear log-odds assumption. Strong on tabular financial data. Preferred by regulators for transparency.

Gradient Boosted Trees

Handles non-linear interactions and missing data. XGBoost and LightGBM are industry standards. Less interpretable without SHAP decomposition.

Random Forest

Ensemble of decision trees. Reduces variance. Provides feature importance scores. Useful for variable selection before final model specification.

Bayesian Inference

Produces a PD distribution, not a point estimate. Updates as new data arrives. Handles sparse data via principled priors. Output: P(PD | data).

Bayes' Theorem for PD Estimation

P(PD | data) ∝ P(data | PD) × P(PD)
Posterior ∝ Likelihood × Prior

Prior: satellite failure rate distributions | Likelihood: telemetry + financial signals | Posterior: updated PD distribution

ALPHA architecture: ensemble ML with Bayesian inference and Shapley value explainability. The ensemble handles non-linear interactions between orbital and financial inputs. The Bayesian layer quantifies estimation uncertainty. SHAP outputs satisfy investment committee explainability requirements.

Module 10 // Application

Credit Ratings

Ordinal classifications that map borrowers to empirical default probability ranges.

Rating	Category	1-Year PD	5-Year PD	Typical Spread
AAA	Investment Grade	0.00%	0.09%	20–50 bps
AA	Investment Grade	0.02%	0.27%	40–80 bps
A	Investment Grade	0.06%	0.58%	70–130 bps
BBB	Investment Grade	0.18%	1.90%	120–250 bps
BB	High Yield	0.85%	7.81%	250–500 bps
B	High Yield	3.24%	19.12%	450–900 bps
CCC	Distressed	22.57%	49.33%	>900 bps

Investment grade threshold: BBB (PD < 1%) is the institutional investor threshold. Most pension funds, insurance mandates, and infrastructure debt facilities require investment grade. SFIS outputs are calibrated to this scale to enable direct comparison with rated peers.

Module 11 // Application

Portfolio & Correlation

Individual asset credit risk is insufficient. Portfolio risk depends on how defaults cluster.

Portfolio Expected Loss (Additive)

EL_portfolio = Σᵢ (PDᵢ × LGDᵢ × EADᵢ)

EL is additive. Unexpected Loss is not. Portfolio UL < Σ individual ULs unless correlation = 1.

Two borrowers with identical PDs produce very different portfolio risk depending on default correlation. At zero correlation, large-number effects smooth losses. At high correlation, losses cluster: near-zero defaults in good years, catastrophic losses in bad years.

Orbital portfolio risk: correlation is driven by operator concentration, orbital regime (LEO/MEO/GEO), launch vehicle dependence, and space weather exposure. Ten LEO satellites from two operators on one vehicle have far higher correlation than ten satellites from ten operators across multiple orbital regimes.

Module 12 // Application

Full Risk Calculator

All components integrated. Configure an exposure and read the complete institutional credit risk output.

EXPOSURE PARAMETERS

EAD Exposure ($M)$100M

PD 1-Year PD5.0%

LGD Loss Given Default65%

PORTFOLIO PARAMETERS

ρ Default Correlation0.20

h Target RAROC12%

T Tenor (yrs)7 yrs

Expected Loss

—

EL as % of EAD

—

Min break-even spread

Economic Capital

—

99.9% VaR buffer

Required Spread

—

EL + capital charge

Implied Rating

—

PD-implied

RAROC Feasibility

—

Reading the output: Expected Loss sets the minimum spread for break-even on credit costs alone. Economic Capital is the equity buffer required at 99.9% confidence. Required Spread includes both the EL component and the capital charge at the target RAROC. RAROC Feasibility flags whether this spread is commercially achievable in the current market for the asset type.

Module 13 // Bayesian

Why Bayesian for Credit Risk

Frequentist methods require large, labelled default samples. Satellite credit has neither. Bayesian inference is not a preference — it is the only statistically coherent option given the data environment.

Classical maximum likelihood estimation (MLE) treats model parameters as fixed unknowns. You estimate them from data alone. With 1,000 mortgage defaults, that works. With 12 recorded satellite operator defaults over 40 years of commercial spaceflight, it does not. MLE produces point estimates with artificially narrow confidence intervals that misrepresent the true state of uncertainty.

Bayesian inference treats parameters as distributions. You start with a prior belief, update it with whatever data you have, and produce a posterior distribution over the parameter. The width of that posterior is an honest representation of how much you do and do not know.

The Core Identity

P(θ | data) = P(data | θ) × P(θ) / P(data)
// Posterior = Likelihood × Prior / Evidence

θ: model parameters (e.g. PD, LGD drivers) | data: observed signals (telemetry, financials, defaults)
P(data) is a normalising constant. In practice we compute: Posterior ∝ Likelihood × Prior

The Frequentist vs Bayesian contrast

	Frequentist (MLE)	Bayesian
Output	Point estimate + standard error	Full posterior distribution over θ
Prior knowledge	Ignored	Incorporated formally as P(θ)
Small data behaviour	Unstable, overfit	Prior regularises the estimate
Uncertainty reporting	95% confidence interval (frequentist)	95% credible interval (direct probability statement)
Sequential updating	Requires full re-estimation	Prior posterior becomes next prior naturally
New asset types	Fails without labelled default data	Prior from analogous assets, updated as data arrives

ALPHA's context: as of 2025 there are approximately 12–15 documented commercial satellite operator credit events (defaults, restructurings, insolvencies) with sufficient financial data to build a labelled dataset. MLE on this sample produces useless estimates. Bayesian inference uses this sparse data as a likelihood update on a well-constructed prior, producing a defensible PD distribution.

Module 14 // Bayesian

Prior Specification

The prior encodes what you know before seeing the data. It is not a guess. It is a formal statement of information from analogous assets, expert judgment, and physical constraints.

The Beta distribution is the natural prior for a probability parameter. It is defined on [0,1], has two shape parameters α and β, and is conjugate to the Binomial likelihood — meaning the posterior is also Beta. This makes updating analytically tractable.

Beta Prior for PD

PD ~ Beta(α, β)
E[PD] = α / (α + β)
Var[PD] = αβ / [(α+β)²(α+β+1)]

Effective sample size of prior = α + β | Larger α + β → stronger prior → more data needed to move the estimate
Uninformative limit: α = β = 1 (uniform prior over [0,1])

Choosing the prior for orbital assets

You have three information sources for the prior. Each maps to a different prior parameterisation.

Historical Satellite Failure Rates

Geostationary failure rate ~1.5%/yr. MEO ~0.8%/yr. LEO ~2.3%/yr. Maps to Beta(2,130) for GEO. Informative but conflates operational failure with financial default.

Analogous Infrastructure Default Rates

Satellite operators resemble project finance issuers. Moody's PF 10-yr cumulative default ~5–8%. Maps to Beta(3,50). Accounts for financial structure without space-specific data.

Expert-Elicited Prior

Structured elicitation from underwriters (ESA, AXA XL). Median PD estimate ~4–6% over 5 years. Maps to Beta(3,60). Most defensible for novel asset class.

Prior sensitivity: a strong prior (high α + β) resists updating even with strong data. Always test how much the posterior moves when you double the prior ESS. If the posterior does not move, you are not learning from data — you are imposing your prior. Report this as a governance risk in the model documentation.

Module 15 // Bayesian

The Likelihood Function

The likelihood is the probability of observing your data given a particular value of PD. It is the mechanism by which data updates the prior.

For a binary default outcome (defaulted = 1, survived = 0), the natural likelihood is Binomial. Given n exposures and d defaults, the probability of observing exactly d defaults if the true PD is θ is:

Binomial Likelihood

L(θ | d, n) = C(n,d) × θᵈ × (1−θ)^(n−d)
log L(θ | d, n) = const + d·log(θ) + (n−d)·log(1−θ)

The likelihood is maximised at θ_MLE = d/n (the MLE). The Bayesian uses the full likelihood function, not just its maximum.

The shape of the likelihood function carries information. A wide, flat likelihood means the data is consistent with many PD values — the data has low information content. A sharp, narrow likelihood means the data strongly localises the PD. The posterior combines the prior's shape with the likelihood's shape.

Multi-signal likelihood for orbital assets

Financial default is not the only signal. Telemetry observations also carry information about default probability. ALPHA constructs a composite likelihood across three signal types.

Signal	Likelihood Form	Update Direction
Binary default event	Binomial L(θ \| d, n)	Direct PD observation
Orbital health index (SIGMA score)	Probit link: P(default) = Φ(β₀ + β₁·SIGMA)	High SIGMA → lower PD
DSCR deterioration	Logistic: log-odds shift per DSCR unit	DSCR < 1.1× → PD spike
Insurance coverage change	LGD adjustment, not PD directly	Coverage lapse → LGD up

Joint likelihood: assuming conditional independence across signals (a strong but tractable assumption), the joint likelihood is the product of individual likelihoods. Log-likelihood is additive. Each new telemetry cycle contributes a likelihood update — functionally equivalent to recalibrating the hazard rate in a reduced-form model.

Module 16 // Bayesian

Posterior Update

When the prior is Beta and the likelihood is Binomial, the posterior is Beta. This conjugacy makes updating exact and computationally free.

Conjugate Update Rule (Beta-Binomial)

Prior: PD ~ Beta(α, β)
Data: d defaults out of n exposures
Posterior: PD | data ~ Beta(α + d, β + n − d)

α increases by the number of defaults observed. β increases by the number of survivals observed.
This is the cleanest possible update rule: count the events and add them to the shape parameters.

The posterior mean is a weighted average of the prior mean and the MLE:

Posterior Mean as Shrinkage Estimator

E[PD | data] = [n/(n + α + β)] × MLE + [(α + β)/(n + α + β)] × Prior Mean

When n is small relative to (α + β): posterior is pulled toward the prior. When n is large: posterior converges to MLE. This is regularisation derived from probability theory, not imposed ad hoc.

Prior → Posterior Update — Interactive

Prior Parameters

α₀ Prior alpha2.0

β₀ Prior beta100

Observed Data

n Exposures40

d Defaults4

Prior Mean

—

MLE (data only)

—

Posterior Mean

—

Posterior Std Dev

—

95% Credible Interval for PD

———

Sequential updating

Bayesian inference is naturally sequential. When a new telemetry observation or financial report arrives, today's posterior becomes tomorrow's prior. No re-estimation from scratch is required. The update rule is the same conjugate formula applied again.

Module 17 // Bayesian

MCMC Sampling

When the posterior has no closed-form solution — because the model has multiple parameters or a non-conjugate structure — Markov Chain Monte Carlo generates samples from the posterior numerically.

The Beta-Binomial conjugate model is analytically tractable. But ALPHA's full model is not: it conditions PD on multiple correlated inputs (SIGMA score, DSCR, operator leverage, insurance coverage, orbital regime). The joint posterior across all parameters requires MCMC.

Metropolis-Hastings Algorithm

1. Start at θ_current
2. Propose θ* ~ q(θ* | θ_current) // proposal distribution
3. Compute acceptance ratio r = [P(data|θ*) × P(θ*)] / [P(data|θ) × P(θ)]
4. Accept θ* with probability min(1, r)
5. Repeat N times → samples approximate the posterior

After burn-in, the chain's stationary distribution equals the posterior. Acceptance rate target: 23–44% for continuous parameters. Your pipeline achieved 8.3% — proposal scale too large.

Diagnosis: is the chain working?

1

Trace Plot

2

Acceptance Rate

3

Autocorrelation

4

R-hat

A healthy trace plot shows a "fuzzy caterpillar" — rapid mixing, no trends, no stuck regions. A pathological trace shows slow drift or periods where the chain does not move (rejection streaks).

Scale Proposal std dev0.040

Acceptance Rate

—

Target: 23–44%

Effective Sample Size

—

Useful samples after thinning

Chain Status

—

Acceptance rate is the fraction of proposed moves accepted. Too low (below 10%): proposals jump too far, almost all rejected, chain does not explore. Too high (above 60%): proposals are tiny, chain moves very slowly. Tune the proposal scale to hit 23–44%.

Your earlier result (8.3%): the proposal scale (MLE_SE × 1.5) was too large for the posterior geometry. Reducing the multiplier to 0.5–0.8 × SE typically restores the chain to the target range. Alternatively, use an adaptive MCMC that tunes the scale during burn-in.

Autocorrelation at lag k measures correlation between chain values θ_t and θ_{t+k}. High autocorrelation means the chain moves slowly and consecutive samples carry redundant information. Effective sample size (ESS) = N / (1 + 2 Σ autocorr_k). With 1,000 raw samples and high autocorrelation, ESS might be only 50.

ρ Autocorrelation0.30

Raw Samples

1,000

Effective Sample Size

—

R-hat (Gelman-Rubin statistic) compares variance within chains to variance between chains. Run M independent chains with different starting points. If they have converged to the same distribution, within-chain and between-chain variance should be equal. R-hat close to 1.0 (below 1.05) indicates convergence. R-hat above 1.1 is a red flag.

Practical rule: run at least 4 chains. Discard the first 50% of each as burn-in. Compute R-hat for each parameter. If any R-hat exceeds 1.05, double the number of iterations and recheck. For ALPHA's multi-parameter model, this diagnostic is mandatory before any PD output is used in a credit decision.

Module 18 // Bayesian

PD Output & Credible Intervals

The posterior distribution over PD is the deliverable. A single point estimate discards the uncertainty information that makes Bayesian inference valuable.

From the posterior samples, you extract three numbers for each horizon: the posterior mean (point estimate for pricing), and the 5th and 95th percentiles (the 90% credible interval for stress testing and capital allocation). The width of the credible interval tells the credit committee how confident the model is.

Posterior Summary Statistics

The credible interval is a direct probability statement: "There is a 90% probability that the true PD lies in this range." This is what practitioners want but cannot get from a frequentist confidence interval.

Credit committee use: present the posterior mean as the base case PD for pricing. Present the 95th percentile as the stress PD for economic capital. The ratio of 95th percentile to mean is the model uncertainty multiplier. A multiplier above 3× indicates insufficient data — the credit committee should require more diligence before approving.

Module 19 // Bayesian

Sensitivity Analysis & SHAP

The posterior distribution answers "what is the PD?" Sensitivity analysis and Shapley values answer "why?" Both are required for institutional explainability.

Posterior sensitivity analysis moves one input at a time and measures the shift in posterior mean PD. It identifies which variables are driving the estimate. Variables with high sensitivity require more rigorous data validation. Variables with low sensitivity can be monitored at lower frequency.

Shapley Values

SHAP (SHapley Additive exPlanations) decomposes the model output into additive contributions from each input variable, averaged over all possible orderings of variable inclusion. It satisfies three axioms: efficiency (contributions sum to the total output), symmetry (equal contributions for equal variables), and linearity (additivity for independent models).

Shapley Value for Variable j

φⱼ = Σ_{S⊆F\{j}} |S|!(|F|−|S|−1)!/|F|! × [f(S∪{j}) − f(S)]

S: subset of features excluding j | F: full feature set | f(S): model output using only features in S
φⱼ measures the average marginal contribution of feature j across all subsets.

Module 20 // Bayesian

ALPHA Architecture

How everything in this module maps to ALPHA's production design. From raw data inputs to the PD output seen by the credit committee.

Full pipeline

1

Data Ingestion

2

SIGMA Score

3

Prior Construction

4

Likelihood Update

5

Posterior PD

6

Output Layer

Data ingestion layer. Space-Track TLE data (orbital elements, decay rates), CelesTrak (operational status, manoeuvre history), UCS Satellite Database (operator identity, launch date, orbital regime), operator financial statements (DSCR, leverage, revenue backlog), insurance registry (coverage type, limits, claims history). Data refreshed at 24-hour cadence for telemetry, quarterly for financials.

Source	Signal Type	Update Freq	Feeds
Space-Track	TLE orbital elements	Daily	SIGMA decay, manoeuvre components
CelesTrak	Operational status flags	Daily	SIGMA anomaly detection
Operator financials	DSCR, leverage, backlog	Quarterly	ALPHA likelihood function
Insurance registry	Coverage, claims	Event-driven	LGD module

SIGMA scoring layer. Seven-factor orbital risk framework. Each factor scored 0–100, weighted and aggregated into a composite SIGMA score. SIGMA is the primary structural proxy for unobservable asset value V in the Merton framework. High SIGMA (healthy orbit) → high implied asset value → lower PD prior.

Prior construction. ALPHA uses a regime-specific prior. Satellites are classified into three prior regimes based on orbital altitude and operator type:

Regime	Prior	Prior Mean	Rationale
GEO Commercial	Beta(2, 130)	1.5%	Low physical failure rate, high operator maturity
LEO Broadband	Beta(3, 80)	3.6%	New operators, constellation execution risk
MEO / Novel	Beta(2, 50)	3.8%	Limited analogues, higher uncertainty

Prior ESS = α + β. Higher ESS priors are harder to move with data. The LEO prior (ESS = 83) is harder to move than the MEO prior (ESS = 52) — reflecting greater confidence in the LEO prior from constellation launch history data.

Likelihood update. ALPHA constructs a composite log-likelihood from three signal types. The SIGMA score enters via a probit link function. DSCR enters via a logistic regression coefficient estimated from infrastructure default data. Insurance status enters as a binary modifier on LGD, not PD.

ALPHA Composite Log-Likelihood

log L(θ | signals) = log L_financial(θ | DSCR, leverage) + log L_orbital(θ | SIGMA) + log L_binary(θ | d, n)

Each new observation appends a term to the log-likelihood. This is the mechanism by which telemetry cycles continuously update the PD estimate without requiring full model re-estimation.

Posterior PD distribution. For the univariate Beta-Binomial model (analytical). For the multi-signal model: Hamiltonian Monte Carlo via NumPyro, 4 chains × 2,000 post-warmup samples, R-hat convergence criterion < 1.05, ESS per parameter > 400. The posterior is summarised as mean, 5th, and 95th percentile for each horizon.

Horizons

1 / 3 / 5 yr

PD distribution per horizon

MCMC Method

HMC

Hamiltonian Monte Carlo

Convergence

R̂ < 1.05

Mandatory gate before output

Output layer. Three outputs per asset, per horizon. Formatted for institutional credit committee consumption.

Output	Value	Use Case
Posterior Mean PD	Point estimate for pricing	Spread floor = EL = PD × LGD × EAD
95th Pct PD (stress)	Worst-case credible PD	Economic capital sizing at 99.9%
CI Width (95th − 5th)	Model uncertainty flag	Triggers enhanced diligence if > 3× mean
SHAP Attribution	Variable contributions	Explainability for credit committee
Implied Rating	PD mapped to S&P scale	Peer comparison to rated instruments

Institutional positioning: this output structure matches what a credit committee at a sovereign wealth fund, infrastructure debt facility, or space-focused insurer needs. PD mean for pricing. Stress PD for capital. CI width for diligence gates. SHAP for governance. Implied rating for peer benchmarking.

Module 21 // Advanced

Your Knowledge Map

What you have covered, what is in this module, and what sits at the frontier of institutional credit risk fluency.

6

COVERED

8

THIS MODULE

4

FRONTIER

78%

INSTITUTIONAL FLUENCY

DONE

EL = PD × LGD × EAD

Core identity, interactive calculator, spread pricing logic.

DONE

Probability of Default

Market-implied, structural (Merton DD), Bayesian. ALPHA architecture.

DONE

LGD & EAD

Recovery waterfall, orbital LGD drivers, amortising EAD profile.

DONE

Expected vs Unexpected Loss

Loss distribution, economic capital, RAROC, IFRS 9 staging.

DONE

Structural & Reduced-Form models, Merton, hazard rates, survival curves.

Model Families

DONE

Bayesian Inference

Beta-Binomial, MCMC, diagnostics, posterior PD, SHAP, ALPHA pipeline.

THIS

Monte Carlo Simulation

Portfolio loss simulation, correlated default paths, tail risk quantification.

THIS

Copula Models

Gaussian copula, t-copula, tail dependence, Li (2000) and its failure modes.

THIS

Stress Testing

Scenario analysis, reverse stress testing, DFAST, EBA methodology.

THIS

CVA / XVA

Credit Valuation Adjustment, DVA, FVA. Pricing counterparty credit risk.

THIS

IFRS 9 Deep Dive

Stage migration, ECL mechanics, significant increase in credit risk triggers.

THIS

Model Validation

Gini, KS, AUC, Brier score, backtesting, SR 11-7, EBA guidelines.

THIS

Credit Derivatives

CDS mechanics, CDS spreads as PD signals, CDO tranching, CLO structure.

THIS

Sovereign & Infra Credit

Moody's PF default study, DSCR floors, sovereign rating frameworks.

FRONTIER

Climate Credit Risk

Physical and transition risk, NGFS scenarios, TCFD credit integration.

FRONTIER

Network / Contagion Models

Interbank contagion, systemic risk, DebtRank, Eisenberg-Noe.

FRONTIER

Deep Learning for PD

LSTM for time-series default, attention mechanisms, variational autoencoders.

FRONTIER

Space-Specific: Kessler

Cascade probability modelling, regime-level correlation shocks, debris density as credit factor.

What "disgustingly knowledgeable" means in practice: you can run any credit committee conversation on methodology, defend model choices under scrutiny, spot flaws in competitor frameworks, and connect every technical component back to a capital allocation or pricing decision. By the end of this module you are there.

Module 22 // Advanced

Monte Carlo Simulation

Simulate thousands of correlated default scenarios. Aggregate into a loss distribution. Read off VaR and CVaR at any confidence level. This is how banks actually compute economic capital.

The analytical loss distribution (e.g. the normal approximation used in Basel IRB) makes strong parametric assumptions. Monte Carlo makes no distributional assumption about the portfolio loss. It simulates default outcomes directly for each asset in each scenario, then counts the losses. The result is an empirical loss distribution built from the data you gave it.

Portfolio Loss Monte Carlo — Algorithm

For each simulation s = 1 to S:
  Draw systematic factor Z ~ N(0,1) // shared macro shock
  For each asset i:
    Draw idiosyncratic ε_i ~ N(0,1)
    X_i = √ρ · Z + √(1−ρ) · ε_i // asset return
    Default_i = 1 if X_i < Φ⁻¹(PD_i)
  Loss_s = Σᵢ Default_i × LGD_i × EAD_i
Sort {Loss_s}. VaR(99.9%) = 999th percentile of 1,000 simulations.

ρ: asset correlation to systematic factor | Φ⁻¹: inverse normal CDF
This is the one-factor Vasicek model underlying Basel II/III IRB capital formula.

The systematic factor Z represents the macro state of the world. When Z is very negative, the economy is in severe stress and many assets breach their default threshold simultaneously. This is how correlated defaults emerge from individual asset-level simulations without explicitly modelling pairwise correlations.

Live Portfolio Monte Carlo — 2,000 Simulations

n Assets in portfolio20

PD Average asset PD5.0%

LGD Loss given default65%

ρ Asset correlation0.20

EAD Avg exposure ($M)$30M

S Simulations2,000

Expected Loss (EL)

—

Mean of distribution

99% VaR

—

1-in-100 loss

99.9% VaR

—

Economic capital basis

CVaR (ES) 99%

—

Avg loss beyond VaR

Worst Simulation

—

UL / EL Ratio

—

Tail risk multiplier

Max Simultaneous Defaults

—

CVaR vs VaR — why CVaR is superior for capital

VaR(99%) answers: "what is the loss at the 99th percentile?" It says nothing about what happens in the 1% tail beyond that point. CVaR (Conditional Value at Risk, also called Expected Shortfall) answers: "given that we are in the tail beyond VaR, what is the average loss?" CVaR is subadditive — portfolio CVaR is less than or equal to the sum of individual CVaRs. VaR is not. Basel III moved from VaR to ES (Expected Shortfall) for market risk capital exactly for this reason.

CVaR Definition

CVaR_α = E[Loss | Loss > VaR_α]
= (1/(1−α)) × ∫_{VaR_α}^{∞} L · f(L) dL

In Monte Carlo: CVaR_99% = average of the worst 1% of simulated losses (top 20 of 2,000 simulations)

Orbital application: Kessler cascade scenario

SFIS Gaussian copula Monte Carlo: SFIS runs this simulation at the portfolio level across 1,763 assets. The systematic factor Z is decomposed into regime-specific sub-factors: macro-financial (shared with terrestrial credit), orbital-specific (space weather, debris density), and operator-idiosyncratic. Correlation structure is parameterised from orbital regime co-location data, not assumed uniform.

Module 23 // Advanced

Copula Models

A copula separates the marginal distribution of each asset's default probability from their joint dependence structure. This is what allows portfolio credit risk to be modelled rigorously.

Sklar's theorem: any joint distribution F(x₁,…,xₙ) can be written as C(F₁(x₁),…,Fₙ(xₙ)) where C is a copula and F₁…Fₙ are the marginal CDFs. The copula captures correlation structure independently of the marginals. You can combine any set of individual PD models with any dependence structure.

Sklar's Theorem

F(x₁,…,xₙ) = C(F₁(x₁),…,Fₙ(xₙ))
// C is the copula: the joint distribution of uniform marginals
Gaussian copula: C(u₁,…,uₙ) = Φₙ(Φ⁻¹(u₁),…,Φ⁻¹(uₙ); Σ)
t-copula: C(u₁,…,uₙ) = tₙ,ν(t_ν⁻¹(u₁),…,t_ν⁻¹(uₙ); Σ)

Φₙ: multivariate normal CDF with correlation matrix Σ | tₙ,ν: multivariate t CDF with ν degrees of freedom

The Li (2000) Gaussian copula — and why it failed

David Li's 2000 paper introduced the Gaussian copula to credit correlation modelling. It became the standard for CDO pricing. The Gaussian copula has zero tail dependence: in the limit of extreme stress, defaults remain approximately independent. This is mathematically elegant but empirically wrong. In the 2008 crisis, defaults became highly correlated in the tail — exactly the region the Gaussian copula said was uncorrelated.

Copula	Tail Dependence	Parameters	Use Case	Limitation
Gaussian	Zero (upper and lower)	Correlation matrix Σ	Normal times, large portfolios	Underestimates crisis correlation
Student-t	Symmetric, non-zero	Σ, ν (degrees of freedom)	Fat-tail environments	Symmetric: same upper/lower tail dep.
Clayton	Lower tail only	θ (one parameter)	Downside clustering	No upper tail dependence
Gumbel	Upper tail only	θ (one parameter)	Boom correlation	No lower tail dependence
Frank	Zero both tails	θ	Symmetric mid-range	Similar failure to Gaussian in tails

SFIS copula choice: SFIS uses Gaussian copula for computational tractability across 1,763 assets. The governance documentation must explicitly acknowledge the zero tail dependence limitation and pair the Gaussian copula output with a t-copula stress overlay for Kessler/debris cascade scenarios. This is the same approach used by Moody's Analytics in their infrastructure portfolio models.

Module 24 // Advanced

Stress Testing

Stress testing asks what happens to the portfolio under severe but plausible scenarios. It is not the same as VaR. VaR is a statistical percentile. Stress testing is a conditional scenario.

A stress test specifies a macro or sector scenario, translates it into shifts in PD, LGD, and EAD across the portfolio, and computes the resulting loss. The scenario is defined by humans, not by a model. This is its strength: it can capture scenarios that have never occurred in historical data.

Three types of stress test

Historical Scenario

2008 GFC

Apply observed macro shifts from a real crisis to current portfolio. Defensible but backward-looking.

Hypothetical Scenario

Solar Storm

Define a plausible but unobserved event. Translate to asset-level shocks. Most relevant for orbital credit.

Reverse Stress Test

Find Break Point

Work backwards from a capital threshold breach. Ask: what scenario causes this loss? More informative than forward stress tests.

Reverse stress testing

Reverse stress testing is conceptually different from scenario stress testing. Instead of asking "what does scenario X do to my losses?", it asks "what scenario causes my losses to exceed my capital buffer?" This forces explicit identification of the conditions that would breach solvency. Regulators increasingly require reverse stress tests because they reveal concentration risks that forward stress tests miss.

Reverse Stress Test Logic

Given: Capital buffer C, Portfolio EAD, LGD
Find: minimum PD multiplier m* such that VaR(99.9%, m*×PD, ρ) = C

Solve numerically by binary search over m. The result tells you how far PDs need to rise before capital is exhausted.

Module 25 // Advanced

CVA / XVA

Credit Valuation Adjustment prices the credit risk embedded in a derivative or bilateral contract. It converts counterparty credit risk into a dollar amount that must be deducted from the mark-to-market value of the position.

When you enter a derivative with a counterparty, you are exposed to two risks: market risk (the value of the derivative moves) and credit risk (the counterparty defaults while the derivative is in-the-money to you). CVA is the expected loss from counterparty default, priced into the derivative at inception.

CVA Formula

CVA = (1 − R) × ∫₀ᵀ EE(t) × dPD(t)
≈ (1 − R) × Σₜ EE(tₖ) × [PD(tₖ) − PD(tₖ₋₁)]

R: recovery rate | EE(t): expected exposure at time t (the expected positive value of the derivative)
dPD(t): marginal default probability in time interval [t, t+dt]
CVA is paid upfront as a reduction in the derivative's fair value.

The XVA family

Adjustment	Full Name	Sign	What It Prices
CVA	Credit Valuation Adj.	Debit	Counterparty defaults while contract is asset to us
DVA	Debit Valuation Adj.	Credit	We default while contract is liability to counterparty
FVA	Funding Valuation Adj.	Debit	Cost of funding uncollateralised derivative positions
MVA	Margin Valuation Adj.	Debit	Cost of posting initial margin under SIMM
KVA	Capital Valuation Adj.	Debit	Cost of holding regulatory capital against the trade
ColVA	Collateral Valuation Adj.	Both	Value of optionality in collateral agreement terms

Space finance relevance: satellite capacity agreements, spectrum leases, and launch contracts are bilateral long-dated contracts. CVA applies. The counterparty (a satellite operator with no public credit rating) has no CDS market for spread calibration. CVA must be computed using SFIS PD outputs — making ALPHA the direct input to CVA pricing for space capacity contracts.

Module 26 // Advanced

IFRS 9 Deep Dive

IFRS 9 replaced IAS 39 in 2018. It changed loan loss provisioning from incurred loss to expected credit loss (ECL). The mechanics are more complex than most practitioners realise.

Under IAS 39 (the old standard), a bank only provisioned for a loan when a loss event had actually occurred. The 2008 crisis revealed this was procyclical: banks built provisions only after the crisis hit, amplifying the shock. IFRS 9 requires provisioning for expected future losses from day one. This is the ECL model.

The three-stage model

Stage 1 — Performing

12-Month ECL

No significant increase in credit risk since origination. Provision = 12-month EL. Interest accrues on gross carrying amount.

Stage 2 — Deteriorated

Lifetime ECL

Significant increase in credit risk (SICR) but no default. Provision = lifetime EL. Interest still on gross amount.

Stage 3 — Defaulted

Lifetime ECL

Credit-impaired. Provision = lifetime EL. Interest accrues on net carrying amount (gross minus provision).

IFRS 9 Expected Credit Loss

ECL = PD × LGD × EAD × D
D = discount factor = 1/(1 + EIR)^t
Lifetime ECL = Σₜ PDₜ × LGDₜ × EADₜ × Dₜ // sum over loan life

EIR: effective interest rate (used as the discount rate)
PDₜ: conditional PD in period t (given no default in prior periods)
The formula looks like EL but discounted over the remaining life.

Significant Increase in Credit Risk (SICR)

The trigger for Stage 1 to Stage 2 migration is SICR. Banks must define SICR in their IFRS 9 methodology. The most common approaches:

Approach	Trigger	Threshold (typical)
Absolute PD threshold	PD exceeds a fixed level	PD > 1% for prime portfolios
Relative PD change	PD has doubled since origination	PD_current / PD_origination > 2×
Watchlist flags	Internal risk management flags	30 days past due (backstop)
Qualitative triggers	Covenant breach, rating downgrade	Below investment grade from IG

The cliff effect: the jump from 12-month ECL to lifetime ECL at Stage migration can be large. A loan migrating from Stage 1 to Stage 2 might see provisions increase 5-10× overnight. This creates P&L volatility that is disproportionate to the underlying credit change, particularly for long-dated infrastructure and satellite loans. IFRS 9 governance must address how SICR triggers are calibrated to avoid mechanical Stage migrations on minor PD movements.

Module 27 // Advanced

Model Validation

A PD model that cannot be validated is not a model — it is an opinion. Validation proves the model discriminates, calibrates, and remains stable over time.

Model validation has three components: discrimination (does the model rank-order risk correctly?), calibration (are the predicted PDs accurate on average?), and stability (does the model produce consistent outputs as inputs change?). Each requires different tests.

Discrimination metrics

Key Discrimination Statistics

AUC (Area Under ROC Curve): probability that a randomly chosen defaulter scores higher than a randomly chosen non-defaulter. AUC = 0.5 is random. AUC = 1.0 is perfect.

Gini Coefficient = 2 × AUC − 1 // ranges 0 (random) to 1 (perfect)

KS Statistic = max|F_defaulter(s) − F_non-defaulter(s)| // max separation of score distributions

SR 11-7 (Federal Reserve) and EBA Guidelines require minimum Gini of 0.30 for retail models, 0.25 for corporate. Your earlier pipeline achieved AUC 0.663 → Gini 0.326. Above minimum threshold but modest.

Calibration: Brier Score and reliability diagrams

Discrimination says nothing about calibration. A model can rank-order perfectly but systematically overstate or understate PDs. Calibration checks whether predicted PD of 5% corresponds to an observed default rate of approximately 5%.

Brier Score

BS = (1/N) × Σ (PD_predicted − Outcome)²

Outcome ∈ {0,1}. Perfect model: BS = 0. Random model predicting base rate p: BS = p(1−p).
Your result: BS = 0.186 vs naive benchmark BS = 0.207. Improvement exists but is modest.

Test	What It Checks	Pass Threshold	Your Pipeline
AUC / Gini	Rank ordering	AUC > 0.65, Gini > 0.30	AUC 0.663, Gini 0.326 ✓
KS Statistic	Score distribution separation	KS > 0.20	~0.29 (estimated)
Brier Score	Calibration accuracy	Below naive benchmark	0.186 vs 0.207 ✓
Hosmer-Lemeshow	Calibration across deciles	p-value > 0.05	Not reported — required
R-hat (Bayesian)	MCMC convergence	R̂ < 1.05	Not reported — required
Acceptance Rate	MCMC efficiency	23–44%	8.3% — failed
Population Stability	Score distribution over time	PSI < 0.10	Not yet implemented

SR 11-7 requirement: all models used in credit decisions must have formal documentation covering conceptual soundness, data integrity, ongoing monitoring, and outcomes analysis. This is not optional. For ALPHA outputs to be used by institutional clients in credit decisions, an independent validation report must exist. This is a pre-condition for regulatory use, not a nice-to-have.

Module 28 // Advanced

Credit Derivatives

CDS, CDOs, CLOs. Credit derivatives transfer credit risk without transferring the underlying asset. They created the market infrastructure that makes spread-implied PD possible — and nearly destroyed the financial system.

Credit Default Swap (CDS)

A CDS is a bilateral contract. The protection buyer pays a periodic premium (the CDS spread) to the protection seller. If the reference entity defaults, the protection seller pays the notional minus recovery. The CDS spread is the market price of credit risk for the reference entity.

CDS Spread — No-Arbitrage Relationship

PV(premium leg) = PV(protection leg)
s × Σₜ DF(t) × [1 − PD_cum(t)] × Δt ≈ LGD × Σₜ DF(t) × ΔPD(t)
// At market: s ≈ PD_hazard × LGD (for flat term structure)

s: CDS spread (bps/year) | DF(t): discount factor | ΔPD(t): marginal default probability
Inverting: hazard rate λ ≈ s / LGD. This is the reduced-form PD extraction you saw in Module 3.

CDO / CLO tranching

A CDO (Collateralised Debt Obligation) pools credit assets and tranches the cash flows. Senior tranches absorb losses last and receive lower yields. Junior/equity tranches absorb losses first and receive higher yields. The tranching redistributes credit risk — it does not eliminate it.

Tranche	Attachment Point	Loss Absorption	Rating Target	Yield Premium
Super Senior	30–100%	Last losses absorbed	AAA	10–40 bps
Senior	15–30%	After equity and mezz	AA–A	50–150 bps
Mezzanine	5–15%	After equity	BBB–BB	200–500 bps
Equity / First Loss	0–5%	First losses absorbed	Unrated	15–25% IRR target

CDO² and the 2008 crisis: CDOs of CDO tranches (CDO²) created synthetic exposures where the underlying correlation assumptions were amplified through multiple tranching layers. The Gaussian copula assigned near-zero tail dependence to diversified pools. In stress, correlation approached 1.0. Senior tranches rated AAA experienced losses exceeding the model's 99.99th percentile. The model was correct given its assumptions. The assumptions were wrong.

Sovereign & Infrastructure Credit

The two asset classes most analogous to orbital credit. Sovereign credit has the longest default history. Infrastructure project finance has the closest structural parallels to satellite finance.

Moody's Project Finance Default Study

Moody's publishes the definitive empirical study of infrastructure project finance defaults. Key findings from the 2023 update (1983–2022 cohort, 8,910 projects):

Sector	10-Year Cum. Default	Recovery Rate	LGD
Power (contracted)	5.2%	80%	20%
Transportation (toll)	8.1%	75%	25%
Oil & Gas (midstream)	6.8%	72%	28%
Telecom / Satellite	~7–12% (estimated)	40–60% (thin data)	40–60%
PF overall average	6.4%	78%	22%

The infrastructure premium: infrastructure project finance exhibits better recovery rates than equivalent-rated corporate bonds (78% vs 40–50%). This is driven by asset tangibility, long-term contracted cash flows, and senior secured structures. Satellite finance currently does not capture this premium because the asset class lacks the standardised credit framework required for institutional comparison. This is the gap SFIS addresses.

DSCR as the primary underwriting metric

Debt Service Coverage Ratio

DSCR = Operating Cash Flow / Debt Service (Principal + Interest)

LLCR (Loan Life Coverage Ratio) = NPV(Cash Flows over loan life) / Outstanding Debt
PLCR (Project Life Coverage Ratio) = NPV(Cash Flows over project life) / Outstanding Debt

Minimum DSCR covenant: typically 1.10–1.20× for investment grade infrastructure.
Below 1.0×: cash flow insufficient to service debt. Technical default trigger.
LLCR and PLCR give forward-looking coverage across the full term.

Sovereign credit rating methodology

Sovereign ratings integrate quantitative and qualitative factors. The four major pillars in Moody's sovereign methodology:

Economic Strength

GDP per capita, growth volatility, diversification

Weight: ~25% of overall rating

Institutional Strength

Rule of law, governance quality, policy track record

Weight: ~25% of overall rating

Fiscal Strength

Debt/GDP, fiscal balance, debt affordability

Weight: ~25% of overall rating

Susceptibility to Event Risk

Political risk, banking sector, external vulnerability

Weight: ~25% of overall rating

SWF relevance: sovereign wealth funds (your primary institutional target) are sophisticated about sovereign credit because they manage sovereign balance sheet risk daily. Framing SFIS outputs in the same analytical language as sovereign credit analysis — coverage ratios, structural protections, macroeconomic sensitivities — is the fastest path to credibility with SWF credit teams.

Module 29 // Regulatory & Capital

Basel III/IV Capital Rules

The regulatory framework that determines how much capital banks must hold against credit risk. IRB vs SA, output floors, and the road to Basel IV.

Basel III requires banks to hold capital against unexpected losses. The Internal Ratings-Based (IRB) approach uses the bank's own PD, LGD, and EAD estimates. The Standardised Approach (SA) uses fixed risk weights by asset class. Basel IV introduces an output floor: IRB capital cannot fall below 72.5% of the SA figure.

IRB Capital Requirement (K)

K = LGD × [N((1−R)^−0.5 × G(PD) + (R/(1−R))^0.5 × G(0.999)) − PD]
RWA = K × 12.5 × EAD
// Capital Ratio = Eligible Capital / RWA ≥ 8% (10.5% with buffers)

N(): cumulative normal | G(): inverse normal | R: asset correlation | PD: probability of default
The formula computes conditional expected loss at the 99.9th percentile minus EL, isolating unexpected loss.

Approach	PD Source	LGD Source	Risk Weights	Output Floor
SA	Not used	Fixed by regulation	20%–150% by rating	N/A (is the floor)
F-IRB	Bank estimate	Fixed by regulation	Formula-based	72.5% of SA
A-IRB	Bank estimate	Bank estimate	Formula-based	72.5% of SA

Basel IV timeline: The output floor phases in from 2023 to 2028 (50% rising to 72.5%). For banks with aggressive IRB models, the floor may bind — increasing capital requirements by 10–25%. This is the single largest regulatory capital impact of the decade.

Module 30 // Regulatory & Capital

Credit Migration Matrices

The probability of transitioning between credit rating grades over time. The foundation of multi-period credit risk modelling and IFRS 9 staging.

A transition matrix records the empirical probability of moving from one rating grade to another over a fixed time horizon (typically 1 year). S&P and Moody's publish annual transition matrices from cohort studies.

Multi-Year Transition Matrix

M(T) = M(1)^T // Matrix exponentiation for integer T
// Generator matrix approach: M(t) = exp(t × Q) where Q = log(M(1))

M(1): 1-year transition matrix | M(T): T-year matrix via matrix power
Cohort method: track rating at start and end of period. Duration method: continuous observation.

IFRS 9 connection: SICR (Significant Increase in Credit Risk) is triggered by multi-notch downgrade probability. The transition matrix feeds directly into lifetime PD calculation for Stage 2 classification. A BBB-rated exposure with a 5-year cumulative default probability exceeding the origination threshold moves to Stage 2.

Module 31 // Regulatory & Capital

Concentration Risk

When portfolio diversification breaks down. Single-name, sector, and geographic concentration amplify tail losses beyond what correlation models predict.

Concentration risk arises when exposure is not sufficiently diversified across obligors, sectors, or geographies. The Herfindahl-Hirschman Index (HHI) measures single-name concentration. The effective number of exposures is 1/HHI.

Herfindahl-Hirschman Index

HHI = Σ_i w_i² // where w_i = exposure_i / total_exposure
Effective Number = 1 / HHI
// Granularity adjustment: GA = (1/N_eff) × σ²(LGD) × correction factor

Perfectly diversified: HHI → 1/N (equal weights). Concentrated: HHI → 1 (single obligor).
Regulatory large exposure limit: single-name exposure ≤ 25% of Tier 1 capital.

Limit	Threshold	Source
Single-name large exposure	≤ 25% of Tier 1	Basel / CRR Art. 395
G-SIB connected limit	≤ 15% of Tier 1	BCBS 283
Sector concentration	No hard cap — Pillar 2	Supervisory review

SFIS orbital regime concentration: LEO mega-constellations create extreme concentration risk. Starlink alone operates >6,000 satellites in overlapping orbital shells. A single Kessler event could impair hundreds of assets simultaneously. Standard HHI measures fail to capture this spatially-correlated concentration.

Module 32 // Counterparty & Transfer

Wrong-Way Risk

When exposure increases precisely when the counterparty's credit quality deteriorates. The most dangerous form of model mis-specification in derivative portfolios.

Wrong-way risk (WWR) occurs when exposure and counterparty default probability are positively correlated. Standard CVA assumes independence: E[Exposure × 1_default] = E[Exposure] × PD. Under WWR, the actual expected loss exceeds this product.

Wrong-Way Risk Adjustment

CVA_WWR = (1+α) × CVA_standard // α captures correlation effect
E[Loss] = E[Exposure | Default] × PD × LGD // conditional exposure > unconditional
// General WWR: macro-driven (e.g., rates move with credit spreads)
// Specific WWR: structural link (e.g., put on own stock)

α: WWR multiplier (0 = no WWR, >0 = positive correlation between exposure and default)

Commodity Producer + Swap

Oil producer enters pay-fixed swap. Oil price drops → producer revenue falls (credit worsens) AND swap MTM rises (exposure increases). Classic specific WWR.

Satellite Capacity + Operator

Satellite operator sells capacity forward. If operator faces distress, capacity delivery risk rises AND the replacement cost of the contract increases. SFIS models this linkage explicitly.

Regulatory treatment: CRR Art. 291 requires banks to identify and manage WWR. The SA-CVA framework applies a 1.4× multiplier for general WWR. Specific WWR exposures must be excluded from netting sets and treated individually.

Module 33 // Counterparty & Transfer

Economic Capital Allocation

How to distribute portfolio-level economic capital to individual exposures. Euler allocation, marginal contributions, and RAROC.

Economic capital (EC) is the internal capital a bank holds to absorb unexpected losses at a chosen confidence level (typically 99.9%). Portfolio EC benefits from diversification — the sum of standalone ECs exceeds diversified EC. The question: how to allocate the diversification benefit fairly?

Euler Capital Allocation

EC_i = w_i × ∂EC/∂w_i // Euler's theorem for homogeneous functions
Σ_i EC_i = EC_portfolio // Euler allocations sum exactly to portfolio EC
RAROC_i = (Spread_i − EL_i) / EC_i

Marginal contribution: additional capital required if exposure i increases by a small amount.
RAROC: Risk-Adjusted Return on Capital. Hurdle rate typically 12–15%.

RAROC decision rule: if RAROC > hurdle rate, the exposure creates value. If RAROC < hurdle rate, the capital is better deployed elsewhere. Euler allocation ensures the decision is consistent: no exposure is subsidised by or penalised relative to its true risk contribution.

Module 34 // Counterparty & Transfer

Credit Risk Transfer

CDS, guarantees, securitisation, and sub-participation. The mechanisms by which credit risk moves from originators to investors — and the conditions under which regulators recognise the transfer.

Credit Risk Transfer (CRT) allows banks to reduce regulatory capital by transferring credit risk to third parties. The Significant Risk Transfer (SRT) test determines whether the bank receives capital relief. Basel requires that the originator transfers a meaningful portion of the risk — not just the first loss or just the senior tranche.

Tranche Expected Loss

EL_tranche = max(0, min(pool_EL, D) − A) / (D − A)
// A = attachment point, D = detachment point, pool_EL = pool-level expected loss
Tranche thickness = D − A

Equity tranche (0-A): absorbs first losses. Mezzanine (A-D): absorbs after equity. Senior (D-100%): last to absorb.

CRT Mechanism	Funded?	Capital Relief	Key Condition
CDS (single-name)	No	Substitution approach	Eligible protection seller
Guarantee	No	Substitution approach	Irrevocable, unconditional
Synthetic securitisation	No	SRT test required	Meaningful risk transfer
True sale securitisation	Yes	Full derecognition possible	True sale opinion + SRT
Insurance wrap	No	Partial (Pillar 2)	Insurer rating, claims history

Satellite insurance as CRT: launch and in-orbit insurance transfers the physical risk component of satellite credit. If insurance covers 100% of replacement cost, the LGD falls from ~90% to ~10%. SFIS models insurance as a funded credit protection mechanism, reducing both EL and EC allocations.

Module 35 // Frontier

Climate Credit Risk

Physical risk, transition risk, and NGFS scenarios. How climate change reshapes probability of default across every asset class.

Climate risk transmits to credit risk through two channels. Physical risk: extreme weather, sea-level rise, and chronic changes damage assets and reduce cash flows. Transition risk: carbon pricing, regulation, and technology shifts strand assets and alter competitive dynamics.

Climate-Adjusted PD

PD_climate = PD_base × (1 + β_phys × PhysRisk + β_trans × TransRisk)
// PhysRisk: physical risk score (0-1), TransRisk: transition risk score (0-1)

β coefficients calibrated to NGFS scenario outputs. Orderly transition: low physical, moderate transition.
Hot house: high physical, low transition. Disorderly: high physical, high transition.

Orderly Transition

Low physical risk, gradual carbon pricing, early policy action

PD uplift: +5–15%

Disorderly Transition

Delayed action, then sudden policy shift. High transition + physical risk

PD uplift: +20–50%

Hot House World

No policy action. Extreme physical risk. Chronic GDP impact.

PD uplift: +30–80%

Transmission Channel	Physical Risk	Transition Risk
Asset damage / impairment	Direct	Stranded assets
Revenue disruption	Supply chain breaks	Demand shift
Operating cost increase	Adaptation costs	Carbon pricing
Funding cost increase	Insurance repricing	ESG screening

Space sector climate nexus: satellites are infrastructure assets with 15-year horizons. Ground station flooding, launch facility exposure, and thermal management in orbit are physical risks. Transition risk is lower — satellites enable climate monitoring and are net contributors to ESG objectives.

Module 36 // Frontier

Network & Contagion Models

Systemic risk through interconnected defaults. When one node fails, the cascade propagates through the network, amplifying losses far beyond the initial shock.

Financial networks exhibit contagion: the default of one institution imposes losses on its counterparties, potentially triggering further defaults. The Eisenberg-Noe clearing model finds the unique clearing vector — the set of payments each node can make given incoming payments from others. DebtRank extends this to partial defaults.

DebtRank Algorithm

h_i(t+1) = min(1, h_i(t) + Σ_j∈N(i) W_ji × Δh_j(t))
// h_i: stress level of node i (0=healthy, 1=defaulted)
// W_ji: relative exposure of i to j (exposure / equity)

DebtRank captures the fraction of economic value lost in the cascade.
Systemic importance: the total DebtRank impact if a single node defaults.

Network topology matters: scale-free networks (few highly connected hubs) are robust to random failures but fragile to targeted attacks on hubs. The interbank market exhibits this topology. Orbital networks may exhibit similar properties — a few relay satellites or ground stations serve as critical hubs.

Module 37 // Frontier

Deep Learning for PD

LSTM, attention mechanisms, and transformers applied to default prediction. Higher accuracy, lower explainability — and the regulatory tension that creates.

Deep learning models capture non-linear interactions and temporal dependencies that logistic regression cannot. LSTM networks process time-series data (financial ratios, market signals, telemetry). Attention mechanisms identify which time steps and features matter most. But credit regulators require explainability (SR 11-7, SS1/23).

Model	Accuracy (AUC)	Explainability	Data Requirement	Regulatory Status
Logistic Regression	0.70–0.75	Full	Low (100s)	Approved
XGBoost / GBM	0.78–0.83	Partial (SHAP)	Medium (1000s)	Case-by-case
LSTM	0.80–0.86	Low	High (10000s+)	Challenger only
Transformer	0.82–0.88	Very low	Very high	Research only

Bias-Variance Tradeoff

MSE = Bias² + Variance + Irreducible Noise
// Simple models: high bias, low variance (underfitting)
// Complex models: low bias, high variance (overfitting)

The optimal model minimises total error. For credit (small data, high stakes), the bias-variance frontier often favours simpler models with interpretable outputs.

SR 11-7 requirements: US bank supervisory guidance requires that models used for credit decisions have “sound development, implementation, and use.” Black-box models fail the conceptual soundness test. The practical path: use deep learning as a challenger model alongside an interpretable champion. ALPHA uses this dual-model architecture.

Module 38 // Frontier

Kessler Cascade Modelling

Debris density as a credit factor. When collisions breed collisions, every asset in an orbital shell shares cascade risk. This is the space-specific frontier — SFIS is building first-principles models.

The Kessler syndrome describes a cascade where orbital debris from one collision increases the probability of subsequent collisions, creating a self-reinforcing chain reaction. For credit risk, this means the assets in a congested orbital shell are correlated not just through market factors but through shared physical destruction risk.

Cascade Probability Model

P_collision(t) = 1 − exp(−λ × D(t) × σ × v_rel)
D(t+1) = D(t) + ΔD(t) × P_collision(t) // debris grows after each collision
// λ: spatial density factor, D(t): debris count, σ: cross-section, v_rel: relative velocity

Each collision creates 10–1000+ trackable fragments. These fragments increase the collision rate for all remaining assets in the shell. The feedback loop is the cascade.

No published methodology exists. Kessler cascade modelling for credit risk is genuinely novel. ESA and NASA model debris propagation for mission safety, not financial risk. SFIS is building the first-principles framework that translates debris density, collision probability, and cascade dynamics into portfolio-level credit risk measures. This is the unique analytical contribution of the platform — the reason the credit output cannot be replicated by any existing model.