12 The Factor Layer as Multiplex Layer 1
Author: Todd B. Adams Reinforces: Proposal §3 (Layer 1) and §6.1 · Reading order: chapter 04 of this part; follows the multiplex data substrate, precedes systemic-risk & crowding groundwork.
What this chapter establishes. The proposal’s first network layer — the bipartite stock ↔︎ factor graph with weighted edges (factor loadings) — is not something the research must build. It exists in production, point-in-time and governed by a full lifecycle. The proposal simulates this layer with structural stochastic differential equations (§6.1); here it is empirical.
12.1 The bipartite layer is already a factor store
The proposal’s Layer 1 is a bipartite graph: stock nodes on one side, factor nodes on the other, edge weight equal to the factor loading. The platform’s factor store is that edge list, and it is empirical rather than synthetic:
- One row per (date, instrument, factor) — literally a timestamped weighted edge between a stock node and a factor node.
- Factor loadings proper — market beta and the broader factor cross-section — are computed point-in-time, so the edge weights are the real, knowable-on-the-date values, not a synthetic draw.
- The active node set is governed, and factors carry economic style tags, so the factor-side nodes are typed (value, momentum, quality, positioning, and so on).
The factor mathematics and the r = α + β·market + ε decomposition are set out in the factor-models-and-benchmarks chapter; the factor-engineering pipeline is in the factor library appendix. These are apparatus notes for the edge weights, not the point — the point is that the edges are measured and pre-validated before any research consumes them.
12.2 The edges are trustworthy, not merely present
A network is only as good as its edge weights. Before an edge is trusted, the platform subjects each factor to a validation funnel (the “mathematics of trust”, chapter 08): an information-coefficient hurdle, permutation significance, the survivorship haircut, a net-of-cost check, the overfitting and lockbox audits, and an orthogonality test. A factor that fails stays experimental; only a validated factor’s edges back a real strategy. The Layer-1 graph the research would consume is therefore pre-cleaned — its edges have already survived a literature-grade gauntlet.
12.3 What the proposal adds on top of Layer 1
| Proposal element | Relationship to the built factor layer | Build status |
|---|---|---|
| Bipartite stock–factor layer (§3, Layer 1) | Is the point-in-time factor store today | implemented |
| Dynamic (time-varying) edge weights | Factor values are recomputed every run, point-in-time | implemented |
| Node features (rolling price, volatility, factor arrays) | Already columns in the end-of-day and factor stores | implemented |
| Embedding Layer 1 into a supra-adjacency tensor with inter-layer coupling | The tensor assembly across Layers 1–3 | proposed — not yet built |
| Heterogeneous message-passing over the tensor | The geometric deep-learning engine | proposed — not yet built |
12.4 Why this matters for the research
The proposal’s simulator exists precisely because a synthetic Layer 1 is needed to prove the physics before touching real data. On this platform the empirical Layer 1 is already available, point-in-time and validated — so the research can move from synthetic proofing to real-data evaluation without first building a factor-loading pipeline. The stock–factor layer is the bridgehead; Layers 2 and 3 (chapter 03) and the tensor and engine (chapter 06) build outward from it.
Cross-links: factor-models primer · factor library · factor & strategy lifecycle · proposal §3/§6