This paper formally combines Narion's two previously independent research tracks — the Institutional Flow Anticipation Engine's (IFAE) microstructure state Xt, and the Regime Intelligence Engine's six-dimensional composite structural state Rt — into a single joint research question: does conditioning order-flow and liquidity dynamics on structural regime context materially change the probability of forward price displacement, including the probability of entering a volatility-expansion state.
Consistent with Narion's disclosure practice, the combined hypothesis, the integration architecture, and empirical findings drawn from Narion's internal validation program are disclosed in academic form. Closed-form feature mathematics, regime-scoring formulas, and proprietary calibration internals remain withheld — see Disclosure & Scope (Section 11).
Market microstructure research and financial regime-switching research have developed largely as separate literatures, despite an obvious complementarity. This paper's premise is that combining them is not a novel theoretical move so much as an overdue empirical one.
The microstructure literature establishes that order flow is informative about price formation and that price impact is a well-behaved, if nonlinear, function of flow. Sequential and strategic trade models — Glosten & Milgrom's specialist-quote model and Kyle's continuous-auction model — remain the founding formal treatments of how informed order flow moves prices [1,2]. Cont, Kukanov & Stoikov formalized order flow imbalance (OFI) as a short-horizon linear predictor of price change from order-book events [3], and a substantial empirical literature since has documented that price impact is better described as concave in order size than linear — the square-root impact law — and that the same signed flow can produce materially different impact depending on prevailing liquidity conditions [4,5]. Order-flow toxicity measures such as VPIN extend this by asking whether the composition of flow, not just its sign, carries forward information about liquidity-provider withdrawal [6].
Separately, the regime-switching literature — beginning with Hamilton's Markov-switching model of macroeconomic time series and extended extensively to financial returns, volatility, and asset allocation — establishes that financial variables often behave as though drawn from a small number of discrete, persistent states rather than a single stationary process, and that the parameters governing returns, volatility, and cross-asset relationships differ materially across those states [7,8].
What is comparatively underexplored is the direct combination of the two: whether a microstructure signal's own informational content — not merely the unconditional distribution of returns — is itself regime-dependent. A small applied literature has begun to test order-flow-conditioned impact and machine-learning approaches to forecasting order flow imbalance under varying conditions [9,10], but a systematic study conditioning a validated microstructure displacement signal on an explicit, multi-dimensional structural regime taxonomy — trend, trend strength, volatility expansion, positioning, and price structure jointly — is, to our knowledge, not established in the published literature. This paper reports Narion's empirical attempt to close that gap using its own production research infrastructure.
Narion's two research tracks were, until this paper, deliberately kept separate: IFAE studies the order-flow information content of trade and order-book activity in isolation (state Xt), and Regime Intelligence studies whether the market's broader structural condition is itself persistent and non-random (state Rt), independent of any specific downstream signal. This paper states, for the first time, the combined research bet that motivated keeping both programs running in parallel: that these two states are not independent inputs to be averaged, but jointly determine the probability of forward price displacement.
This is a strictly stronger and more falsifiable claim than either track's own standalone hypothesis. IFAE's governing hypothesis only requires that Xt carry some information about Yt. Regime Intelligence's governing hypothesis only requires that Rt be structured and persistent in its own right. Eq. 1 additionally requires an interaction: that the function mapping Xt to displacement probability itself changes shape across values of Rt, rather than Rt simply adding an independent, additive adjustment. Sections 04–06 report the empirical program built to test this, and Section 05 shows that the interaction is, in specific and reproducible cases, large.
Both Xt and Rt are computed as independent, live per-bar reconstructions inside Narion Execution Research Infrastructure, then combined downstream rather than merged into a single model. This separation is deliberate: it keeps each state independently testable (as documented in their own research modules) while still allowing their joint effect on displacement probability to be studied directly.
Neither Xt nor Rt is discarded or overwritten by this integration; both remain independently published and independently validated. What is new in this paper is the joint outcome statistics computed at their intersection.
Testing Eq. 1 required moving well beyond a single backtest. Narion's internal validation program proceeds through thirteen sequential stages, each reconciled independently against the raw output of its underlying audit scripts rather than against narrative summaries — a discipline maintained specifically because, as Section 07 discusses, several of the program's own audit scripts were found to disagree with each other on the same underlying statistics.
Five findings, drawn from the reconciled research program, bear most directly on Eq. 1. Each is reported here at the level of a research result — what was found and how strongly — consistent with how Narion discloses validation outcomes elsewhere; the internal scoring formulas and thresholds that produced them remain proprietary (Section 11).
Regime Intelligence's expansion-state dimension — classifying the market as compressing, trading normally, or actively expanding on a short- versus long-window volatility ratio — offers a direct, isolable test of Eq. 1, because it can be held out from the other five regime dimensions and examined as a single conditioning variable against the identical execution-qualified signal population.
Within the execution-qualified structural-location breakdown, the expansion sub-split shows a clear divergence in realized hit rate for the identical underlying signal population:
Read plainly, this says the same order-flow displacement signal is a meaningfully better bet when the market is trading at a normal volatility regime than when it is already in active volatility expansion — plausibly consistent with the model identifying candidates closer to the start of a genuine move under normal conditions, and later, into an already-expanding and more exhausted move, under active expansion. This is a single-dimension result, correlated with the other five regime dimensions rather than fully isolated from them (Section 03 of Narion's Stage 3 market-context research), and should be read as illustrative support for Eq. 1 rather than as a standalone production filter.
A joint hypothesis over a six-dimensional regime space multiplies the number of cells being tested far faster than forward data accrues. Narion's research program applies four specific statistical disciplines to keep regime-conditional claims honest.
Rather than a binary confirmed/rejected label, regime-conditional findings in this program are classified into four evidentiary types, extending the three-type taxonomy used in Narion's earlier regime research to a fourth type the joint (Xt, Rt) research specifically surfaced.
| Type | Definition | Policy Implication |
|---|---|---|
| A | Established historical evidence, calibrated across a large, statistically adequate sample. | Eligible for SECONDARY-tier research status; still not a live-execution approval on its own. |
| B | Established negative evidence — a regime with adequate sample size and a confirmed unfavorable outcome. | WATCHLIST — retained as an active negative control, not silently dropped from the taxonomy. |
| C | Underpowered forward evidence that is directionally consistent with the hypothesis. | RESEARCH — continued observation; explicitly not distinguishable from noise at conventional significance. |
| D | Forward evidence genuinely insufficient to distinguish the candidate from noise in either direction. | RESEARCH, null-not-refuted — must not be reclassified as a negative finding on the same evidence that would be too thin to promote a positive one. |
The combined evidence supports Eq. 1 directionally, and in specific, well-sampled cases — extreme trend strength, VWAP positioning, and volatility-expansion conditioning — decisively. It does not yet support treating any single regime-conditioned configuration as production-ready. The binding constraint, consistently across every stage of this program, is independent forward-episode count, not model or regime-taxonomy design.
The recommended research agenda accordingly prioritizes accrual over further parameter search: freeze the regime taxonomy and candidate configuration for any regime under active forward study, apply mandatory episode-level de-duplication at collection time rather than retrospectively, pre-register a stopping rule (a minimum independent-episode count, tested once, with a single interim futility check) before evaluating significance, and apply multiplicity correction across every regime candidate tested concurrently. Further sensitivity-grid expansion over thresholds, horizons, or payoff structures is explicitly de-prioritized until episode count itself has grown, since the current program's own degeneracy analysis shows that expanding the grid without expanding independent episodes adds configurations without adding information.
Model_D's own IFAE validation program remains, independently, at the same disciplined stage — shadow-validated, not production-promoted, pending forward unseen-data confirmation. This paper's contribution is to show that the two tracks' combination is not merely additive: the regime dimension changes what a given microstructure reading should be taken to mean, in directions and magnitudes that are measurable today, even where the confidence to act on any single instance of that conditioning is not yet.
This paper states the combined governing hypothesis, the integration architecture, the structure of the empirical research program, and headline findings and limitations because that is what an academic-grade research publication owes its readers. It withholds the closed-form feature and regime-scoring mathematics behind Xt and Rt, and the exact internal thresholds and configuration parameters behind every reported result, because those remain the proprietary research infrastructure each standalone module was built to produce.
The governing microstructure hypothesis, seven-signal family, and Model_D validation record.
PUBLIC MODULEThe standalone regime-persistence hypothesis and six-engine composite classification architecture.
PUBLIC MODULEHow live signal intelligence, regime research, and policy combine into an auditable, fail-closed execution-decision pipeline.
PUBLIC MODULEResearch briefings for qualified partners and engineering collaborators, including the full reconciled empirical corpus withheld from this page, are reviewed within 48 hours.