26  Session 17: The GE-LAV® Platform — Architecture, Data, Calibration

Unit 3 — Decision and Application
Book Chapter 9 (sections 9.1–9.5)
Track Common core (both tracks)
Assessment milestone PS3 drops at end of class

26.1 Learning Objectives

By the end of this session, students will be able to:

  1. Describe the GE-LAV® platform’s architecture: data ingestion, calibration engine, valuation pipeline, output formats.
  2. Identify the data inputs required: secondary market discounts, fund cash flows, public benchmarks, asset-class parameters.
  3. Run the calibration workflow for the OU process and the quadratic premium function on a small sample dataset.
  4. Interpret platform outputs: LAV mark, GE-LAV mark, exit boundary recommendation, stress scenario values, Pigouvian tax recommendation.
  5. Evaluate the platform against alternative implementation strategies (closed-form formulas, custom code, spreadsheet-only).

26.2 Pre-Class Assignment

  • Read: Book Chapter 9, sections 9.1–9.5 (~12 pages)
  • Register: Course site provides platform access credentials. Log in before class.
  • Optional: Skim Table 9.1 (full parameter listing) — you’ll reference it in PS3

26.3 In-Class Outline (75 minutes)

Time Segment Format
0:00–0:05 Recap Unit 3 so far · Today: how to actually do it Lecture
0:05–0:25 Platform architecture overview Lecture
0:25–0:45 Data inputs and calibration workflow Lecture + demo
0:45–1:00 Output interpretation Lecture + walk-through
1:00–1:15 Alternative implementations · PS3 introduction Lecture

26.4 Discussion Questions

  1. The platform requires quarterly secondary market data. Some asset classes (private credit, smaller infrastructure deals) have thinner secondary market data than PE buyout. How should the platform handle data-sparse asset classes?
  2. The default calibration uses 22 years of secondary data (2003–2024). Should crisis periods (2008–2010, 2020) be weighted differently than normal periods in the calibration? Why or why not?
  3. If a sophisticated LP wanted to build their own implementation in Python, what’s the minimum viable pipeline? Could it be done in <500 lines of code?

26.5 Worked Example: Calibration on Sample Data

Setup: Suppose you have 20 quarterly observations of secondary discounts:

Q1: 8%, Q2: 6%, Q3: 9%, Q4: 12%, Q5: 18%, Q6: 25%, Q7: 32%, Q8: 28%,
Q9: 22%, Q10: 15%, Q11: 10%, Q12: 8%, Q13: 7%, Q14: 9%, Q15: 11%,
Q16: 14%, Q17: 12%, Q18: 9%, Q19: 7%, Q20: 8%

Step 1: Convert to \(L_t\) series

Using linear discount-to-L mapping (book Eq. 2.5): \(L_t = 1.0 - 5 \times (\text{discount}/0.50)\)

Yields: \(L_t\) ranges from −1.2 (Q7) to +0.5 (Q19) over 20 quarters.

Step 2: MLE for OU parameters

OU log-likelihood (Euler discretization): \(\ell(\kappa, \bar{L}, \sigma) = -\dfrac{1}{2} \sum_{t=2}^{T} \left[ \log(2\pi\sigma^2 \Delta t) + \dfrac{(L_t - L_{t-1} - \kappa(\bar{L} - L_{t-1})\Delta t)^2}{\sigma^2 \Delta t} \right]\)

Maximize numerically. Output: - \(\hat{\kappa} = 0.52\)/yr - \(\hat{\bar{L}} = -0.05\) (close to 0 — sample is just noisy around 0) - \(\hat{\sigma} = 0.41\)

Step 3: Compare to canonical calibration

Parameter Sample MLE Canonical
\(\kappa\) 0.52 0.45
\(\bar{L}\) -0.05 1.0 (different normalization)
\(\sigma\) 0.41 0.32

Differences: Higher \(\sigma\) likely because sample includes 2008–2010 stress. Reasonable agreement otherwise.

26.6 What to Expect Next Session

Session 18 is a deep platform demo + validation session. We’ll:

  • Run a complete portfolio valuation end-to-end
  • Validate platform outputs against the GFC, COVID, 2022 historical episodes
  • See where the platform’s predictions matched and where they diverged
  • Discuss limitations and known gaps

PS3 drops today. Both tracks. Due Session 24. The Track 1 PS3 uses the platform extensively. Get familiar with the UI this week.

Reading: Book Chapter 9, sections 9.6–9.10 (~10 pages).

← Session 16 | Schedule | Next: Session 18 →

Compute it live: The π(L, T) market-clearing surface in this session is interactive at liquidityillusion.com.