9 The Platform as a Research Substrate
Author: Todd B. Adams Reinforces: Proposal §1 — Executive Summary and §3 · Reading order: chapter 01 of this part; follows the overview, precedes the end-to-end pipeline.
What this chapter establishes. Most factor-network research is validated on static, survivor-biased, cost-free academic panels — the very conditions the proposal criticises traditional risk models for assuming. We argue that a live instrument built to be hard on itself removes those conveniences, and that surviving those harder conditions is what makes it a credible testbed for a claim as strong as “we can detect market phase transitions.”
9.1 Why a production trading platform, not a notebook
A network model of the kind the proposal advances must survive four failure modes that quietly kill most quant-network research. The value of the substrate is that it treats each as a standing condition on evidence, not a one-off cleanup — so a result that clears them is making a far stronger claim than one fit on a clean academic panel.
| Failure mode | How research fails when unguarded | The evidence-grade condition imposed here |
|---|---|---|
| Look-ahead leakage | Features quietly use future data; the “signal” is time travel | The point-in-time analytical layer is constructed so a historical date sees only what was knowable then |
| Survivorship illusion | Dead companies dropped, manufacturing predictability | 4,858 delisted names are spliced back with realistic terminal losses; the evidence gauntlet subtracts a measured survivorship haircut rather than assuming none |
| Cost fantasy | Gross edges that die after the spread | A per-name spread and capacity estimate let every result be reported net of cost |
| Overfitting | Best-of-many-tries kept and reported as if tested once | The deflated Sharpe ratio, the probability-of-backtest-overfitting estimate, purged cross-validation, and honest family-wise trial counts |
A network model that clears these conditions makes a stronger claim than one fit on a clean panel. That is the substrate’s entire value to the research.
9.2 What the substrate offers the network model
- Point-in-time tensors without new engineering. The proposal’s supra-adjacency tensors — factor loadings and dynamic edge weights — would be assembled from the point-in-time analytical layer, inheriting its no-look-ahead guarantee (chapter 02).
- A real cross-section, not a synthetic panel. The proposal’s simulator (§6) uses 100 synthetic stocks to prove the physics; the platform offers the empirical graduation target — a liquid, survivorship-corrected equity universe spanning roughly two decades of daily history.
- An executability verdict. A phase-transition signal is useful only if it can inform a tradeable book. The pipeline terminates in cost-aware, risk-capped, broker-submitted orders (chapter 02), so a network signal can be judged on executed profit-and-loss rather than in-sample fit.
- A validation gauntlet already calibrated to the literature. The anti-overfitting machinery (chapter 08) applies to any new producer — a network producer included — so graph-model claims face the same bar as every factor.
- Reproducible experiments. Every run is immutable and audited (chapter 09).
9.3 The honest boundary
We are careful to distinguish scope from result. The platform does not today implement multiplex networks, graph neural networks, supply-chain ingestion, or institutional-ownership (13F) data — those are proposed, not yet built, and constitute the research. What it provides is the substrate and the discipline: the reason the research can be done credibly here rather than started from zero. Where each proposed piece attaches is chapter 06; what is missing, and in what order it would land, is chapter 07.
Cross-links: platform whitepaper · platform capability map · proposal §1