Collateralized Loan Obligation Waterfall: Mathematical Foundations

This document provides rigorous mathematical derivations for CLO/SPV waterfall mechanics, including loss allocation, priority of payments, and differentiable approximations for end-to-end training.

1. Tranche Structure

1.1 Definition

Definition 1 (Tranche)

A tranche $k$ is defined by:

• Attachment point $A_k \in [0, 1]$: Loss level where tranche begins absorbing losses
• Detachment point $D_k \in [A_k, 1]$: Loss level where tranche is fully wiped out
• Notional $N_k = (D_k - A_k) \cdot N_{\text{total}}$

Example: Standard CLO structure

Tranche	Rating	Attachment	Detachment	Spread
A	AAA	30%	100%	L+120
B	AA	22%	30%	L+180
C	A	16%	22%	L+250
D	BBB	10%	16%	L+350
E	BB	5%	10%	L+600
Equity	NR	0%	5%	Residual

1.2 Loss Allocation

Definition 2 (Tranche Loss Function)

Given portfolio loss rate $L \in [0, 1]$, the loss allocated to tranche $k$:

\[\ell_k(L) = \min\left(\max\left(\frac{L - A_k}{D_k - A_k}, 0\right), 1\right) \tag{1}\]

The dollar loss:

\[\text{Loss}_k = \ell_k(L) \cdot N_k \tag{2}\]

Theorem 1 (Loss Allocation Properties)

The tranche loss function satisfies:

1. Non-negativity: $\ell_k(L) \geq 0$
2. Boundedness: $\ell_k(L) \leq 1$
3. Monotonicity: $\ell_k$ is non-decreasing in $L$
4. Piecewise linearity: $\ell_k$ has exactly two breakpoints

Proof

Properties 1-2 follow directly from the min-max formulation. For monotonicity, let $L_1 < L_2$: - If $L_2 \leq A_k$: $\ell_k(L_1) = \ell_k(L_2) = 0$ - If $L_1 \geq D_k$: $\ell_k(L_1) = \ell_k(L_2) = 1$ - If $A_k < L_1 < L_2 < D_k$: $\ell_k(L_1) = \frac{L_1 - A_k}{D_k - A_k} < \frac{L_2 - A_k}{D_k - A_k} = \ell_k(L_2)$ - Other cases: Similar analysis For piecewise linearity: $$ \ell_k(L) = \begin{cases} 0 & L \leq A_k \\ \frac{L - A_k}{D_k - A_k} & A_k < L < D_k \\ 1 & L \geq D_k \end{cases} $$ $\square$

1.3 Conservation of Loss

Theorem 2 (Total Loss Conservation)

For non-overlapping tranches covering $[0, 1]$:

\[\sum_{k=1}^{K} \ell_k(L) \cdot (D_k - A_k) = L \tag{3}\]

Proof

Consider $L \in [A_j, D_j]$ for some tranche $j$. Then: - Tranches below $j$: $\ell_k(L) = 1$, contribution = $D_k - A_k$ - Tranche $j$: $\ell_j(L) = \frac{L - A_j}{D_j - A_j}$, contribution = $L - A_j$ - Tranches above $j$: $\ell_k(L) = 0$, contribution = $0$ Sum: $$ \sum_{k < j}(D_k - A_k) + (L - A_j) = A_j + (L - A_j) = L $$ Since $\sum_{k < j}(D_k - A_k) = A_j$ by construction. $\square$

2. Interest Waterfall

2.1 Priority of Payments

Definition 3 (Interest Waterfall Priority)

Interest collections $I_{\text{total}}$ are distributed in order:

1. Senior fees and expenses: $F$
2. Tranche A interest: $I_A = \min(I_{\text{avail}}, r_A \cdot N_A)$
3. Tranche B interest: $I_B = \min(I_{\text{avail}}, r_B \cdot N_B)$
4. …
5. Equity residual: $I_{\text{equity}} = I_{\text{avail}}$

2.2 Formal Specification

Algorithm 1: Interest Waterfall

Input: Interest collections I_total, fees F, tranche notionals {N_k}, spreads {r_k}
Output: Interest to each tranche {I_k}

I_avail = I_total - F
For k = 1 to K-1 (senior to junior):
    I_k = min(I_avail, r_k * N_k)
    I_avail = I_avail - I_k
I_equity = I_avail  # Residual to equity

2.3 Interest Coverage Tests

Definition 4 (Interest Coverage Ratio)

\[\text{IC}_k = \frac{I_{\text{total}} - F}{\sum_{j=1}^{k} r_j \cdot N_j} \tag{4}\]

Trigger Condition: If $\text{IC}_k < \text{IC}_k^{\text{threshold}}$, divert cash to pay down senior tranches.

3. Principal Waterfall

3.1 Standard Principal Distribution

Definition 5 (Principal Waterfall)

Principal collections $P_{\text{total}}$ (scheduled + prepayments + recoveries):

1. Senior fees (if interest insufficient)
2. Interest shortfall coverage
3. Principal to Tranche A until paid down
4. Principal to Tranche B until paid down
5. …
6. Residual to equity

3.2 Overcollateralization Tests

Definition 6 (Overcollateralization Ratio)

\[\text{OC}_k = \frac{\text{Portfolio NAV}}{\sum_{j=1}^{k} N_j} \tag{5}\]

Theorem 3 (OC Trigger Effect)

When $\text{OC}_k < \text{OC}_k^{\text{trigger}}$:

\[P_{\text{redirect}} = \max\left(0, \sum_{j=1}^{k} N_j - \frac{\text{NAV}}{\text{OC}_k^{\text{target}}}\right) \tag{6}\]

is diverted from junior to senior tranches.

3.3 Reinvestment Period

Definition 7 (Reinvestment Waterfall)

During reinvestment period (typically years 1-4):

\[P_{\text{reinvest}} = P_{\text{total}} - P_{\text{triggers}} \tag{7}\]

Principal is reinvested in new loans rather than paying down tranches.

4. Full Numerical Example

4.1 Setup

Portfolio:

• Total notional: $N = $500\text{M}$
• Weighted average coupon: 7%
• Expected default rate: 3%
• Expected recovery rate: 60%

Capital Structure:

Tranche	Size	Rate	Attachment	Detachment
A	$350M	L+1.2%	30%	100%
B	$40M	L+1.8%	22%	30%
C	$30M	L+2.5%	16%	22%
D	$30M	L+3.5%	10%	16%
E	$25M	L+6.0%	5%	10%
Equity	$25M	Residual	0%	5%

4.2 Interest Waterfall (Normal Period)

Assumptions: LIBOR = 5%, Fees = $1M/period

Interest Collections: \(I_{\text{total}} = 500\text{M} \times 7\% / 4 = \$8.75\text{M}\)

Distribution:

Item	Calculation	Amount	Remaining
Fees		$1.00M	$7.75M
A Interest	$350M × (5% + 1.2%)/4	$5.43M	$2.32M
B Interest	$40M × (5% + 1.8%)/4	$0.68M	$1.64M
C Interest	$30M × (5% + 2.5%)/4	$0.56M	$1.08M
D Interest	$30M × (5% + 3.5%)/4	$0.64M	$0.44M
E Interest	$25M × (5% + 6.0%)/4	$0.69M	-$0.25M
E Shortfall			$0.25M
Equity		$0.00M

Result: Tranche E receives only $0.44M (64% of entitled amount).

4.3 Loss Allocation (Stress Scenario)

Assume portfolio loss = 8% ($40M):

Tranche	Attachment	Loss Rate	Loss Amount
Equity	0%	100%	$25.00M
E	5%	(8%-5%)/(10%-5%) = 60%	$15.00M
D	10%	0%	$0.00M
C	16%	0%	$0.00M
B	22%	0%	$0.00M
A	30%	0%	$0.00M
Total			$40.00M

Verification: $25 + 15 = 40$M = 8% × $500M

4.4 Coverage Ratio Computation

Interest Coverage (Class D):

\[\text{IC}_D = \frac{8.75 - 1.0}{5.43 + 0.68 + 0.56 + 0.64} = \frac{7.75}{7.31} = 1.06\]

Typical trigger: 1.20 → This would trigger diversion.

Overcollateralization (Class D):

After 8% loss, NAV = $460M:

\[\text{OC}_D = \frac{460}{350 + 40 + 30 + 30} = \frac{460}{450} = 1.02\]

Typical trigger: 1.10 → This would trigger principal diversion.

5. Differentiable Approximation

5.1 Problem Statement

The standard loss allocation uses non-differentiable operations (min, max). For end-to-end training, we need smooth approximations.

5.2 Soft Gating Functions

Definition 8 (Smooth Loss Allocation)

Replace hard thresholds with sigmoid gates:

\[\ell_k^{\text{soft}}(L; \tau) = \sigma\left(\frac{L - A_k}{\tau}\right) \cdot \left(1 - \sigma\left(\frac{L - D_k}{\tau}\right)\right) \cdot \frac{L - A_k}{D_k - A_k} \tag{8}\]

where $\sigma(x) = 1/(1 + e^{-x})$ and $\tau > 0$ controls sharpness.

Lemma 5.1 (Convergence)

\[\lim_{\tau \to 0^+} \ell_k^{\text{soft}}(L; \tau) = \ell_k(L) \tag{9}\]

5.3 Alternative: Softplus Clipping

Definition 9 (Softplus Approximation)

\[\text{clip}_{\text{soft}}(x, a, b) = a + \text{softplus}(x - a) - \text{softplus}(x - b) \tag{10}\]

where $\text{softplus}(x) = \log(1 + e^x)$.

Properties:

• Smooth everywhere
• Gradient always exists
• Asymptotically approaches hard clipping

5.4 Gradient Flow

Theorem 4 (Gradient Through Waterfall)

For differentiable loss allocation, the gradient with respect to portfolio loss $L$:

\[\frac{\partial \ell_k^{\text{soft}}}{\partial L} = \frac{1}{D_k - A_k} \cdot \mathbf{1}_{A_k < L < D_k} + O(\tau) \tag{11}\]

Proof Sketch

Taking derivatives of the soft formulation and evaluating at interior points where both sigmoids are near 0.5, the gradient simplifies to the constant slope $1/(D_k - A_k)$. The sigmoid derivatives contribute terms of order $\exp(-|L - A_k|/\tau)$ and $\exp(-|L - D_k|/\tau)$, which vanish as $\tau \to 0$ for interior $L$. $\square$

5.5 Training Considerations

Temperature Annealing:

During training, anneal $\tau$: \(\tau_t = \tau_{\max} \cdot \left(\frac{\tau_{\min}}{\tau_{\max}}\right)^{t/T} \tag{12}\)

Start with $\tau_{\max} = 0.1$ for smooth gradients, end with $\tau_{\min} = 0.001$ for accuracy.

6. Stochastic Extensions

6.1 Monte Carlo Waterfall

Algorithm 2: MC Waterfall Simulation

Input: N_sims, loss distribution parameters
Output: Tranche loss distributions

For s = 1 to N_sims:
    L_s ~ LossDistribution(params)
    For k = 1 to K:
        loss_k[s] = ell_k(L_s) * N_k

Return {mean(loss_k), std(loss_k), VaR(loss_k)}

6.2 Expected Tranche Loss

Theorem 5 (Expected Loss Integration)

For continuous loss distribution with PDF $f_L$:

\[\mathbb{E}[\ell_k(L)] = \int_0^1 \ell_k(L) f_L(L) \, dL \tag{13}\] \[= \int_{A_k}^{D_k} \frac{L - A_k}{D_k - A_k} f_L(L) \, dL + \int_{D_k}^1 f_L(L) \, dL \tag{14}\]

6.3 Beta Distribution Example

For $L \sim \text{Beta}(\alpha, \beta)$ with $\alpha = 2, \beta = 50$:

\[\mathbb{E}[L] = \frac{\alpha}{\alpha + \beta} = \frac{2}{52} \approx 3.8\%\]

Tranche E (5%-10%):

\[\mathbb{E}[\ell_E] = \int_{0.05}^{0.10} \frac{L - 0.05}{0.05} f_L(L) \, dL + \int_{0.10}^{1} f_L(L) \, dL\]

Using numerical integration:

• $\mathbb{P}(L > 0.05) \approx 15\%$
• $\mathbb{P}(L > 0.10) \approx 3\%$
• $\mathbb{E}[\ell_E] \approx 0.08$ (8% expected loss to tranche E)

7. Implementation

7.1 PyTorch Waterfall

def soft_tranche_loss(portfolio_loss, attachment, detachment, tau=0.01):
    """Differentiable tranche loss allocation."""
    # Soft lower bound
    above_attachment = torch.sigmoid((portfolio_loss - attachment) / tau)
    # Soft upper bound
    below_detachment = 1 - torch.sigmoid((portfolio_loss - detachment) / tau)
    # Linear interpolation
    linear_loss = (portfolio_loss - attachment) / (detachment - attachment)
    # Combine
    return above_attachment * below_detachment * linear_loss.clamp(0, 1)

7.2 Full Waterfall Class

class DifferentiableWaterfall:
    def __init__(self, tranches, tau=0.01):
        self.tranches = tranches  # List of (attachment, detachment, notional)
        self.tau = tau

    def allocate_loss(self, portfolio_loss_rate):
        losses = {}
        for name, (att, det, notional) in self.tranches.items():
            loss_rate = soft_tranche_loss(portfolio_loss_rate, att, det, self.tau)
            losses[name] = loss_rate * notional
        return losses

8. Summary

Concept	Formula	Purpose
Tranche Loss	$\ell_k = \text{clip}((L-A_k)/(D_k-A_k), 0, 1)$	Loss allocation
Interest Coverage	$\text{IC}_k = (I - F) / \sum_j r_j N_j$	Performance test
Overcollateralization	$\text{OC}_k = \text{NAV} / \sum_j N_j$	Credit enhancement
Soft Allocation	$\sigma((L-A_k)/\tau) \cdot (1-\sigma((L-D_k)/\tau))$	Differentiable

References

1. Fabozzi, F. J., & Kothari, V. (2008). Introduction to securitization. Wiley.
2. Duffie, D., & Garleanu, N. (2001). Risk and valuation of collateralized debt obligations. Financial Analysts Journal.
3. Longstaff, F. A., & Rajan, A. (2008). An empirical analysis of the pricing of collateralized debt obligations. Journal of Finance.
4. Bengio, Y., Leonard, N., & Courville, A. (2013). Estimating or propagating gradients through stochastic neurons. arXiv.