SECTION III

Assumptions, Cautions & Competitive Landscape

Thea Portfolio Analytics  ·  NAVcasting™

Simple Assumptions

Private funds have:

returns correlated with the public market, and estimable beta (β) and alpha (α).

idiosyncratic returns that are significantly — but only partly — explained by publicly traded comparable assets.

market-clearing prices that are distinct from GP-reported NAVs and other appraisals, which exhibit smoothed, arbitrageable characteristics.

realized cash-on-cash performance that is not affected by GP reporting.

secondary prices that are not stale, but are likely conditional on quotes being close to the GP-reported NAVs.

Private fund GPs:

have incentives to report smooth NAVs that are tied to the underlying economic value nonetheless.

face scrutiny from LPs on how reliable their fund's unrealized performance has been — limiting the capacity to provide an (unforced) price that two informed parties would trade at.

Implications for Real-Time Valuation Methods

Models should embed cash flows with NPV constraints that allow returns to vary through time.

Fund returns are substantially but not fully explained by public market and comparable assets.

Fund data (NAVs and cash flows) and third-party data (secondaries, deep-dive appraisals) should be modeled as noisy, possibly biased, but nonetheless informative signals of the fund asset value.

The fund asset value estimates must satisfy the no-arbitrage conditions.

Hard to argue with these assumptions… but even harder to measure!
Thea Portfolio Analytics  ·  NAVcasting™

“Caution: Do Not Try This At Home”(The method is half of it. The infrastructure around it is the other half.)

Why Sophisticated Models Need Infrastructure

Peer-reviewed research as the foundation — then years of scientific advancement, infrastructure development, and validation at scale with real funds and deeper simulation stress testing.

Sophisticated models buy real advantages — but they can break in ways simple deterministic models do not.

Those breaks can be loud or quiet. Anyone claiming they don’t occur has either not looked, or would fail the quant interview. Catching them takes monitoring infrastructure grounded in scientific understanding and domain expertise.

Some Words of Caution

However it is framed, statistically and scientifically, this problem carries almost every wrench in the book.

Most problems have one or two non-standard features — a chapter of an advanced econometrics or time-series text handles them.

This one has nearly all of those chapters, in a single modelling problem (albeit answering many questions simultaneously).

Extracting the signal — as opposed to unwittingly “fitting to the noise” with black-box models — requires principled scientific approaches with clear, auditable decompositions: both from a diagnostic perspective and from an explainability / liability perspective.

The signal is there… extracting it reliably is the infrastructure!
Thea Portfolio Analytics  ·  NAVcasting™

The Competitive Landscape

Other methods in the literature traffic in:

Look-ahead bias (using data not yet seen).

Inappropriate application of aggregate / industry models to a single fund.

Erratic outcomes (without monitoring or infrastructure around them).

Published statistical instabilities (unhealthy diagnostics — often collinearity issues).

Lack of identifiability (heavy collinearity).

Or worse, no auditable model decomposition (black box) — an even greater challenge for monitoring and solving.

Low frequency, or only estimating the past.

Inability to blend all fundamental sources of data

cash flows + market returns + holding-data-informed comparable assets

GP-reported NAVs + 3rd-party valuations from ground-up holding data

scientifically decipher signal from noise for all components coherently and reliably

Assumed a priori fund terminal date.

Static parameters that don't update as fund-specific data emerges and time passes.

Small validation sets (not ready for the wild).

In sum:

Many of these approaches may be worthwhile for attribution of the past, aggregate studies, and in highly curated / controlled lab environments.

None of these methods are truly fit for real-world purpose, and cannot:

intake all relevant bottom-up and top-down fundamentals data

coherently separate the signal from the noise

model fund-specific data

do it at high frequency

in real time

and do it reliably

Can’t we just throw a bunch of compute at this problem?

After all, we solved Navier–Stokes…
  1. Math & Physics: roughly more deterministic & clean problems.
    • Finance/Economics: not quite as clean a space
      1. Oxygen molecules always behave like oxygen molecules
      2. Cash flows for one fund almost never behave like cash flows for another
    • Our space is a significantly ‘dirtier’ environment
      1. not defined by the laws of physics
      2. contains tremendous heterogeneity and troublesome edge cases
  2. Naturally, this messier playing field requires:
    • Deep fundamental training in estimation techniques and literacy + domain knowledge
    • You could throw X billion tokens at this problem, but….
      1. Without expert forensic review and well-informed (and intuitive) guided experiments….
      2. You are likely going to burn those X billion tokens to nowhere…
  3. But they are useful research assistants:
    • “Managers of machines” – Citadel grows headcount
      1. We already have a framework for this (Professors and post-docs)
    • It’s an accelerating universe
      1. Those ahead get disproportionately further ahead!