Architecture

Architecture Overview

Private Credit uses a hierarchical deep generative framework with four levels:

1
2
3
4
5
6
7
8
9
10

Level 1: MACRO SCENARIO GENERATOR
    |
    v
Level 2: TRANSITION TRANSFORMER
    |
    v
Level 3: LOAN TRAJECTORY MODEL
    |
    v
Level 4: PORTFOLIO AGGREGATOR

Level 1: Macro VAE

Purpose: Generate correlated macroeconomic time series

  • Architecture: Conditional Variational Autoencoder with LSTM encoder/decoder
  • Input: Historical macro series, scenario specification
  • Output: Correlated macro path (GDP, unemployment, spreads)
  • Key Feature: Scenario conditioning (baseline/adverse/severe)

Learn more about Macro VAE

Level 2: Transition Transformer

Purpose: Predict cohort-level transition matrices

  • Architecture: Transformer Encoder with cross-attention
  • Input: Macro path, cohort features (vintage, asset class)
  • Output: Time-varying transition probabilities
  • Key Feature: Captures systematic risk and macro sensitivity

Learn more about Transition Transformer

Level 3: Loan Trajectory Model

Purpose: Generate individual loan paths

  • Architecture: Autoregressive Transformer Decoder + Diffusion Head
  • Input: Loan features, cohort transitions, macro path
  • Output: State sequence, payment sequence, default timing
  • Key Feature: Captures idiosyncratic risk within cohorts

Learn more about Loan Trajectory Model

Level 4: Portfolio Aggregator

Purpose: Aggregate to portfolio and tranche level

  • Architecture: Differentiable waterfall simulation
  • Input: Loan trajectories
  • Output: Portfolio cashflows, loss distribution, tranche returns
  • Key Feature: End-to-end differentiable for joint training

Learn more about Portfolio Aggregator

Correlation Structure

Correlation is captured at multiple levels:

Level Mechanism
Macro All loans affected by same macro path
Cohort Loans in same cohort share transition dynamics
Factor Latent factors for industry/geography clustering
Idiosyncratic Residual loan-specific variation

Training Strategy

  1. Stage 1: Pre-train components separately
  2. Stage 2: End-to-end fine-tuning with portfolio targets
  3. Stage 3: Calibration to historical data

Back to Home