3  Research Questions

This page states the research programme’s testable core. It is extracted from the research proposal §1 so that the current form of the research question is available without reading the full proposal; when the proposal’s hypothesis or questions change, both are updated together.

3.1 Central hypothesis

We hypothesise that changes in the spectral properties and meso-scale organisation of dynamic multiplex financial networks — constructed from factor exposures, supply-chain dependencies, and institutional ownership — provide statistically significant early-warning signals of systemic market instability, with measurable lead time relative to standard volatility- and correlation-based indicators. This is a proposed claim: the instrument to test it is described below, and the multiplex construction it rests on remains to be built and tested.

3.2 Overarching question

Can dynamic heterogeneous multiplex representations of financial markets provide robust early-warning signals of systemic instability that materially precede conventional volatility-based and return-correlation risk measures?

3.3 Hypotheses and sub-questions

We decompose the central hypothesis into four questions, each falsifiable on its own:

  1. Which multiplex network observables — algebraic connectivity, eigenvalue gap, spectral-density shifts, network entropy, community fragmentation, centrality concentration — reliably signal impending systemic instability?
  2. How much lead time do these indicators provide relative to standard risk metrics (implied/realised volatility, DCC-GARCH correlation spikes, realised drawdowns)?
  3. Under which network generative mechanisms (Erdős–Rényi, Barabási–Albert, stochastic block models, empirically-derived topologies) do the indicators and their detection performance generalise?
  4. Do graph neural network models improve predictive skill over spectral, hand-crafted indicators, and if so by how much, and at what cost to interpretability?

3.4 Open questions (not yet in the proposal’s formal scope)

  • Does a multi-layer (factor + supply-chain + ownership) network out-perform the single-layer factor and comovement view the platform already measures, or does it merely add noise? This is the boundary between what is established and what is proposed. The platform’s own single-layer experiment found that connectedness moved with stress rather than ahead of it — a measured, negative-for-early-warning result, reported in the platform’s asset-dependency-network review. Whether the multi-layer extension recovers genuine lead time is the open question that extension exists to answer, and its design is meant to fail cheaply if it does not.
  • How should a validated early-warning signal enter the trading system without becoming an optimiser that sizes the book — that is, remaining risk observability rather than a second lifecycle gate? The boundary is asserted in the whitepaper §7 and is not yet resolved in the research.

3.5 Where the answers will come from

  • Methodology — how each question is tested.
  • Experiments — the running record of what has been tried against each question.
  • Journal — the raw, dated account of what changed a question, closed one, or opened a new one.