LeatherSagiKnowledgebase

Research Journal — CRE Asset Pricing Model Development (Feb–Mar 2026)

4 min read

Problem

The CRE project has ~10 years of accumulated context and a codebase that started life in Julia 0.6. Someone new to the project — or a future Claude Code session operating without the full CLAUDE.md — has no linear path through what was done, why, and in what order. This journal exists to provide that linear path for the Feb–Mar 2026 window in which the project transitioned from “partially working legacy codebase” to “incumbent on simulated-data validation arc with an MAP estimate”.

Key idea

A 26-phase chronological log of everything the project did between roughly 2026-02 and early 2026-04, written by Daniel Leather with Claude Code as the collaborating agent. Each phase records the objective, what was tried, what broke, and what was learned. It is the most authoritative source for “when did X happen and why”.

Method

The journal’s structure follows the actual development order, not a post-rationalized outline:

Results

The journal is itself the primary artifact. The substantive numerical results it reports are captured in experiment pages; the journal’s unique value is the ordering and the why.

Selected highlights from the narrative:

  • Every previous estimation approach over a decade failed (Sobol grid, MCMC, neural surrogates, TD learning, VFI). The common thread was the feasible volume fraction.
  • The RBPF where particles carry regime histories only was the key abstraction that made the problem tractable.
  • The TSM-MLE v2 global search was where the project first discovered that the MLE sits on a feasibility boundary — a finding that shaped every subsequent constrained-estimation and basin-finder experiment.
  • The Riccati convergence correction was a deep bug in the coupled recursion that had been miscomputing prices until it was found and fixed.
  • Particle anatomy (Phase Robustness) became the origin of the idea that required particle count can be predicted from cheap model features without running the particle filter.

Limitations

  • Chronological, not topical. Finding a specific topic requires scanning the table of contents. Use experiment pages or topic pages for topical entry.
  • Single-author. Written in Daniel Leather’s voice with Claude as collaborator; not externally reviewed.
  • Window is Feb–Mar 2026. The data-contamination recovery arc (2026-03-22 → 2026-04-12) is not fully covered here; see design-brief-global-optimization-pipeline and synthesis-final-day-sprint-decisions for that window.

Open questions

  • What is the right way to keep this kind of research journal going? It is high-value but extremely high-labor; most projects do not sustain them.
  • Should the experiment pages auto-link back to the journal phases that produced them, or is the one-way link (journal → experiments) enough?

My take

This journal is the single most useful onboarding document in the repository. Any future session that needs to understand why a particular architectural choice was made (why is m_g constrained? why is T_bar = 1400 safe? why is the crisis scale 10?) will save hours by reading the relevant phase of the journal first. Its role in the wiki is not to propose new claims but to supply motivational context for the experiment and claim pages that do propose claims.

Graph View

Backlinks