Glossary
This glossary fixes the vocabulary of this book, so that a word means the same thing in every chapter.
It is an overlay, not a replacement. The platform that produced our measurements maintains its own canonical domain thesaurus (docs/THESAURUS.md in the engineering documentation area), and that document remains authoritative for the platform’s own vocabulary. This glossary adopts those definitions by reference and adds three things on top:
- Research-layer terms — econophysics, network-science, and statistical vocabulary the platform thesaurus does not carry, because they belong to the science rather than to the instrument.
- Register differences — words that exist in both layers but mean something narrower, broader, or simply different here. These are the entries that earn the overlay its keep: an unmarked register difference is how two documents come to disagree while using identical words.
- Terms excluded from research prose — vocabulary that is correct inside the platform and wrong in this book, with what to write instead.
Where this glossary and the platform thesaurus genuinely conflict, this glossary governs prose in this book and the thesaurus governs the platform. Neither is wrong; they address different readers.
1. Research-layer terms
Terms belonging to the science. None of these are platform vocabulary, and none should be defined by reference to how the platform computes them.
Networks and multiplex structure
Multiplex network. A network on a fixed node set carrying several distinct edge types, each edge type forming a layer, with layers coupled through the shared nodes. In this programme the nodes are instruments and the candidate layers are factor exposure, supply-chain dependency, and institutional ownership. Distinguished from a multilayer network in the general sense, where node sets may differ across layers.
Layer. One edge type within a multiplex. Not to be confused with the platform’s data layers, which are an infrastructure concept and do not appear in this book.
Algebraic connectivity. The second-smallest eigenvalue of the graph Laplacian (the Fiedler value). A measure of how difficult the network is to disconnect; falling algebraic connectivity indicates a structure becoming easier to fragment.
Spectral gap. The separation between successive eigenvalues of a network operator, most often between the leading eigenvalue and the bulk. Widening indicates increasing dominance of a single collective mode — in market terms, a market moving as one thing.
Spectral density. The distribution of eigenvalues of a correlation or adjacency matrix. Compared against a random-matrix null (Marchenko–Pastur) to separate structure from sampling noise.
Network entropy. An entropy computed over a network’s spectral or degree distribution, measuring how evenly connectivity or dynamical weight is spread. Falling entropy indicates concentration.
Community fragmentation. Change in the meso-scale community structure of a network — communities merging, splitting, or dissolving — as distinct from a change in overall connectivity.
Centrality concentration. The degree to which centrality mass is carried by few nodes. A distributional statement about the network, reported with the concentration measure used.
Generative mechanism. A model of how a network is produced — Erdős–Rényi, Barabási–Albert, stochastic block model, or an empirically-derived topology. Used here to test whether an indicator’s detection performance generalises beyond the topology it was developed on.
Systemic risk and early warning
Systemic instability. A market state in which shocks propagate rather than dissipate. Always operationalised explicitly before use — the threshold, the horizon, and the outcome variable stated — never left as an intuitive appeal.
Early-warning signal. An observable whose change precedes an operationalised instability event by a measurable interval, more often than a matched null predicts. A signal with no stated null and no stated lead time is not an early-warning signal.
Lead time. The interval between an indicator’s trigger and the onset of the event it is claimed to anticipate, reported as a distribution with its dispersion, never as a single number.
Phase transition. A qualitative change in collective market behaviour at a critical point, with the associated critical phenomena — diverging correlation length, critical slowing down. Used in its statistical-mechanics sense; not a synonym for “regime change”.
Critical slowing down. Increasing autocorrelation and variance in a system’s fluctuations as it approaches a critical point, a candidate early-warning mechanism independent of network structure.
Estimation and evidence
Information coefficient. The cross-sectional rank correlation between a predictor’s value at one date and realised forward return over a stated horizon. Meaningless without horizon, population, period, and a gross-or-net label.
Point-in-time. A property of a historical measurement: at every date, only information knowable at that date was used. A point-in-time claim is a claim about the measurement, and is what makes a backtested result admissible evidence.
Survivorship bias. Upward bias from a sample that excludes instruments which ceased to exist. Reported here with its measured magnitude where we have estimated it, not as a generic caveat.
Look-ahead bias. Use of information at a date before it was knowable. Distinguished from survivorship bias, and from dissemination lag — the interval between a fact becoming true and becoming public, which is the mechanism by which look-ahead most often enters fundamental data.
Multiple-comparison correction. Adjustment of significance thresholds for the number of hypotheses tested. Always reported with the trial count it assumes; a correction applied at an understated trial count is worse than none, because it looks rigorous.
Deflated Sharpe ratio. A Sharpe ratio adjusted for the number of trials, non-normality, and sample length — an overfitting-aware significance statement rather than a performance statistic.
Probability of backtest overfitting. The estimated probability that a strategy selected as best in-sample underperforms the median out-of-sample. An estimate about a selection procedure, not about a strategy.
Baseline. The established method a new one must beat: a standard factor model, a volatility- or correlation-based risk measure, a published estimator. Reported with its parameterisation and its performance on the same sample.
Null result. A result in which the null was not rejected. A research output, reported with the same care as a rejection, and never described as a failure of the instrument.
2. Register differences
Words that exist in both layers. The platform’s sense is given for orientation; the research sense is what governs prose in this book.
| Term | Platform sense | Research sense in this book |
|---|---|---|
| Experiment | not a platform concept | A hypothesis test with its variables, sample, estimator, and decision rule fixed before it is run. A pipeline execution is not an experiment. |
| Run | one pipeline execution, carrying an identifier | Avoid. Write “the measurement of 26 May 2026”, or name the experiment. The identifier never appears in prose. |
| Factor | a declared entity with a lifecycle status | A cross-sectional explanatory variable — a candidate risk-premium or behavioural-bias proxy. Its lifecycle status is a property of our instrument, not of the science, and is not evidence. |
| Strategy | a factor paired with a mechanic and parameters | A trading rule under test, defined by its construction and rebalancing convention. |
| Universe | a versioned, point-in-time instrument set | The population under study, and the sample drawn from it. State inclusion criteria as sampling criteria. |
| Signal | a scored column produced by a model | A predictive quantity, defined by what it predicts, at what horizon, and with what estimator. |
| Gate | an automated promotion check | A pre-specified decision rule. Say what the rule is and what it decides; that it is automated is an infrastructure fact. |
| Validated | a lifecycle status | Avoid as a claim about truth. A factor is not “validated”; a null is rejected under stated conditions. Where the status itself is the subject, say “reached the validated stage of our lifecycle” and treat it as a description of the instrument. |
| Feature | a measured property of an instrument at a time | Avoid — it collides with the software sense. Write “variable”, “predictor”, or “observable”. |
| Model | often a trained artefact | The mathematical object: estimand, estimator, and assumptions, stated so it can be re-implemented in any language. |
| Coverage | the fraction of a population with non-missing data | Same quantity, but always reported as a data limitation with its effect on the estimate, never as a platform statistic. |
| Cost | modelled transaction cost | The economic friction a result must survive. Every performance number is labelled gross or net. |
3. Terms excluded from research prose
These are correct inside the platform and wrong in this book. The rule they serve is the two-layer separation: the research layer must remain intelligible to a reader who has never seen our codebase.
| Excluded | Why | Write instead |
|---|---|---|
| Internal identifiers of any kind — feature, optimisation, task, run, gate, bug, dashboard card | Encode our build order, not the argument | Name the mechanism or capability in plain English |
| Tracker status stamps | Bookkeeping vocabulary | “is implemented”, “is partially built”, “is not yet built” |
| Pipeline, ingestion, promotion, orchestration | Infrastructure construction | Describe the measurement and its point-in-time properties |
| Schema, table, column, endpoint, service, class | Implementation detail | Describe the data’s provenance, characteristics, and limitations |
| Dashboard, sidecar, single-page application, route | User-interface implementation | Omit; if a figure came from an operator screen, redraw it for the page |
| Deployment, environment, configuration switch | Operations | Omit; if a setting changes a result, it is an experiment parameter — record it as one |
| “The platform can now …” as a motivation | Capability is not a research question | State the question the capability unlocks, or drop the passage |
4. Maintenance
- Check before you coin. Before introducing a term in any research file, look here first, then at the platform thesaurus. If the concept has a settled term, use it exactly.
- Add in the same pass. A new term is added here in the commit that first uses it. A term used in the book and absent from this glossary is a defect, not a backlog item.
- Record register conflicts explicitly. When a word is already in the platform thesaurus with a different sense, add a row to §2 rather than silently redefining it. The two layers are allowed to differ; they are not allowed to differ invisibly.
- This is a seed. The entries above were drawn from the research-questions chapter, the capability map, and the register conflicts known at the time of writing. Extend it as the book grows rather than attempting a single exhaustive sweep — an unused definition is as much of a liability as a missing one.