14  Gap Analysis & Research Roadmap

Author: Todd B. Adams Reinforces: Proposal §3 & §5 · Reading order: doc 07 of the research-platform package

The claim of this doc. Here is the honest ledger of what is built versus what the PhD proposes, expressed as an increment plan in the platform’s own idiom. Nothing below hides a gap; the point is that each increment is scoped against an existing seam, and the earliest, cheapest increment is the one that most de-risks the research.

14.1 Built vs. proposed — the ledger

Capability Proposal role Status Where
Point-in-time data + Gold model Feeds all tensors BUILT doc 02
Stock–factor bipartite layer (L1) Multiplex Layer 1 BUILT doc 04
Correlation / crowding / factor-ρ networks Layer-3 proxy, risk observability BUILT doc 05
FF5+UMD, AQR, FRED reference series Benchmarks / regime covariates BUILT doc 03
Validation gauntlet (MCPT/PBO/DSR/purged-CV) Anti-overfitting for network claims BUILT doc 08
EDGAR + FINRA ingestion machinery Sourcing path for L2/L3 PARTIAL doc 03
Supply-chain layer (L2) Directed unipartite stock–stock PROPOSED —
13F ownership layer (L3) Bipartite fund–stock PROPOSED —
Supra-Adjacency tensor builder §3 subsystem 2 PROPOSED doc 06
ST-GNN / HGT engine §3 subsystem 3 PROPOSED doc 06
Supra-Laplacian spectral analyzer §3 subsystem 4 (early warning) PROPOSED doc 06

14.2 The increment plan

Ordered so the earliest step is the cheapest and most decisive — the platform’s standard “spike before you build” discipline (the DY-connectedness spike is the precedent: it killed the weakest hypothesis for the price of a script).

  1. Increment A — Spectral observability producer (report-only). A sbnetwork/sbspectral leaf that assembles the built Layer-1 factor graph (plus the crowding correlation network) into a single-layer Laplacian and publishes its algebraic connectivity / spectral radius as a report-only sidecar. Reuses the sbcrowding template end-to-end. De-risks the core physics claim on real data before any GNN or new dataset.
  2. Increment B — Supply-chain layer (L2). Extend the EDGAR client from fundamentals verification to 10-K customer–supplier edge extraction → a directed stock–stock dataset. Largest data-engineering item.
  3. Increment C — 13F ownership layer (L3). Reuse the EDGAR client for 13F holdings → a bipartite fund–stock dataset, replacing the comomentum crowding proxy with holdings ground truth.
  4. Increment D — Multiplex tensor + ST-GNN engine. Assemble Layers 1–3 into the Supra-Adjacency tensor; introduce the isolated torch-geometric engine on research-day cadence.
  5. Increment E — Phase-transition evaluation & (eventually) a gate. Validate the multi-layer spectral early-warning signal through the full gauntlet (doc 08); only after survived evidence does it inform anything downstream (report-then-enforce).

14.3 Honest risks, stated up front

  • The core hypothesis may not hold. The platform’s own spike found single-layer connectedness coincident, not leading (findings §6). Increment A is designed to test whether the multi-layer spectral version does better — and to fail cheaply if it does not.
  • L2 supply-chain data is noisy. 10-K relationship extraction is an NLP problem with real precision limits; the roadmap treats it as a research task with its own evaluation, not a solved ingestion.
  • The GNN dependency is heavy. torch-geometric is quarantined to one research-day package (doc 06) so it never touches the nightly production path.

Cross-links: proposal §3/§5 · target architecture · DY-connectedness spike · the internal epics roadmap directory