Institutional order flow leaves a trace before price moves.

A research question, stated once and tested repeatedly.

Visible order-book imbalances are effects, not causes. Narion's research program treats them that way: the organizing question is whether the surge in aggressive order execution that precedes a directional price move is measurable — in the arrival rate, size, and timing structure of trades and depth changes — before a conventional imbalance pattern has fully formed. The Institutional Flow Anticipation Engine (IFAE) is the operational system built to test that question against BTC/USDT perpetual futures at tick resolution, and Model_D is the current machine-learning candidate trained on its output.

What follows is organized the way a research paper is organized: a governing hypothesis, the system that operationalizes it, the signals derived to test it, the labeling and modeling philosophy used to learn from them, the full empirical results produced to date, and an explicit accounting — in Section 09 — of what has been independently validated versus what remains open. Readers assessing production readiness should read Section 09 before treating the results in Section 07 in isolation.

Price velocity responds to net aggressive flow, damped by resting liquidity.

The market is modeled as a stochastic dynamical system in which price does not move on its own — it is pushed by net aggressive order flow and resisted by the depth of liquidity standing in its way. This is the governing hypothesis the rest of the research program exists to test, and it is the one piece of the underlying mathematics Narion publishes in closed form, in keeping with standard practice among microstructure researchers who state their governing model openly while keeping estimation detail proprietary.

Q_t and L_t are theoretical constructs, not single observable numbers. IFAE's signal family (Section 03) is the operational reconstruction of this pair from seven microstructure metrics computed directly off live order-book and trade streams, and Section 02 describes the system that performs that reconstruction end to end. How each metric is combined into a forecast, and the exact statistical machinery used to estimate α, β, and the resulting probability of displacement, are described qualitatively in Sections 04–06 and are not published in closed form.

Five layers, each independently testable.

IFAE is organized as a layered pipeline rather than a single monolithic model. Each layer has one responsibility, is versioned independently, and can be validated on its own terms before the layer above it is trusted. This separation is what makes the empirical results in Section 07 attributable to a specific stage of the pipeline rather than to the system as an undifferentiated whole.

Layers 1–2 supply the data contract Section 03 describes; Layers 2–3's labeling and verification handoff is the subject of Section 05; Layer 3's model is the subject of Section 06. Layers 4–5 are execution-engineering concerns, described here only at the level needed to understand where the empirical evaluation loop closes.

Seven metrics, each answering a different question about flow.

IFAE reconstructs Q_t and L_t from live L2 order-book depth and tick-by-tick trade executions using seven purpose-built microstructure metrics. Each is described below by function — what it measures and what it is designed to detect — together with the qualitative interpretation bands used for human monitoring, rather than by its closed-form statistical definition.

Each signal is read at three different time-scales.

A single instantaneous reading of any of the seven metrics is a noisy quantity on its own. IFAE additionally summarizes every metric over multiple trailing windows using a small set of standard rolling statistics, giving the model a view of not just the current value of a signal but its recent trend, stability, and directional consistency. The statistics themselves — mean, dispersion, trend, and directional persistence — are named here by what they measure; their exact window lengths and closed-form derivation are withheld as labeling/feature-engineering detail.

Every one of the seven core metrics receives this treatment, so the model's effective input space is substantially larger than seven numbers per tick — it is each metric's current reading plus its short-, medium-, and longer-horizon behavior. The exact horizon lengths, the closed-form rolling-statistic formulas, and the resulting feature count are model-training detail and are not published on this page.

From a raw tick stream to a tradeable label, in outline.

Two design commitments shape how IFAE turns raw microstructure signal into a training target, described here at the level of research philosophy rather than implementation.

Across the 2,057,101-row Model_D master dataset, this two-stage construction produced a substantially imbalanced target: positive long and short labels together account for under 4% of all rows, with the overwhelming majority of ticks correctly carrying no signal at all. This imbalance is treated as a first-class modeling constraint (Section 06.3), not an artifact to be resampled away.

The exact displacement threshold, continuation horizon, and take-profit/stop-loss architecture used by the verifier are labeling-logic detail and are not published on this page.

Why the signal family needs a nonlinear learner.

Model_D is a gradient-boosted tree ensemble, chosen over a linear baseline for a specific structural reason: microstructure signals are conditional. Aggressive-buyer dominance is only informative in the context of concurrent volatility; a persistence reading can reflect genuine continuation or a stop-hunt depending on liquidity behavior occurring at the same time. A single linear decision boundary cannot represent that conditionality — it can only learn one global direction of effect per signal.

The exact tree configuration, loss-reweighting formula, and training hyperparameters are model-training detail and are not published here.

What the retraining study actually found.

Framework A is Narion's standing retraining protocol: drift analysis, retraining, shadow comparative replay, confidence calibration, and — pending — shadow promotion. Model_D has completed the first four phases against BTC/USDT perpetual futures, evaluated August 2026. The results below are the headline empirical findings from that study, presented as results, the way a publication reports outcomes, not as a description of the estimation method that produced them.

7.1 — Phase 1: Why Retraining Was Triggered

The production model showed declining signal density in the most recent market regime — at the ≥0.70 confidence threshold, production signal density in the Recent regime had fallen to 0.196%, prompting a formal retraining decision.

7.2 — Phase 2: Dataset & Cross-Validation

The Model_D master dataset spans 2,057,101 chronologically ordered rows across 67 model features, from 2025-11-13 to 2026-08-18, split into Historical, Transition, and Recent periods reflecting the maximum feasible temporal composition of the available data rather than a fixed target ratio. Model_D was evaluated with chronological (time-series) five-fold cross-validation over the most recent 800,000 rows, respecting temporal ordering rather than shuffling the sample.

7.3 — Phase 3: Signal Density Recovery by Regime

Production and Model_D were scored against the identical master dataset and identical 67-feature vectors. Retraining recovered signal density across every regime tested, with the largest relative improvement precisely in the Recent regime the retraining was intended to address.

7.4 — Phase 4: Confidence Calibration

Precision, recall, F1, and signal density were computed at six candidate thresholds against the full dataset. The transition from 0.50 to 0.70 is the key inflection point: precision roughly doubles while recall falls only marginally.

p_long ≥ 0.70 is recommended as the general operating threshold: it is the strongest balance of precision, recall, F1, and signal density both on the full dataset and on the Recent regime specifically. p_long ≥ 0.90 is retained as a candidate high-confidence tier — excellent precision, but recall too low for a general production threshold — pending forward validation.

Reading model confidence as a market phase, not just a probability.

Beyond the discrete precision/recall calibration in Section 07, Narion's research notes attach a qualitative phase-space interpretation to model confidence bands — reading p_long not just as a probability but as a rough marker of where the market sits in a displacement cycle, from noise through accumulation, ignition, propagation, and exhaustion.

The research notes report that realized forward returns rise through the middle of this range and weaken or reverse at the extreme high end, consistent with the model most reliably catching a move once liquidity absorption is well underway but before it is fully exhausted — a reading that argues for treating the extreme-high band as a structural confirmation filter rather than an entry trigger. This differs from the single-threshold calibration in Section 07.4 and the two have not been formally reconciled against a common evaluation protocol; this is recorded as an open item in Section 09.

Seven scenarios human reviewers watch for.

What the validation record does not yet establish.

A results section is only as trustworthy as the caveats published alongside it. The items below are read before, not after, the empirical results in Section 07 — they are the reason Model_D remains a shadow-validated candidate rather than a production system.

Two further items are already flagged in the underlying Model_D evaluation: the Phase 4 calibration figures reflect precision/recall against labels drawn from the same construction window as the training data — a form of the same-dataset issue above, specific to calibration; and no audit script in the reviewed evaluation independently tests for overlapping or duplicate signal windows beyond the verifier's trade-lock mechanism, so signal independence should be treated as designed-for rather than confirmed.

A legitimate shadow candidate, not yet a production system.

On the balance of the evidence in Sections 07–09, Model_D is a legitimate shadow-deployment candidate. It has not been promoted to production, and this module concurs with that decision on the grounds already stated: the calibration was performed on the same dataset used to train the model rather than an untouched forward holdout, and cross-validation variance across folds indicates the underlying feature–return relationship is non-stationary.

The correct next step is not a further retraining cycle. It is a forward shadow deployment: run production and Model_D simultaneously against newly arriving market data that neither model has been trained or calibrated on, at their respective operating thresholds, logging both models' outputs and the realized future outcome for each. The promotion question to answer is whether Model_D retains its recovered signal density while maintaining calibrated precision on genuinely unseen data — not whether it continues to outperform on the dataset it was calibrated against.

Ahead of or alongside that deployment, the open items from Section 09 are prioritized: reconciling the single-threshold-versus-banded-entry discrepancy on a common evaluation protocol; an explicit overlap/independence audit of the signal sample beyond the existing trade-lock mechanism; and reconciling the causal and phase-space interpretive material against audit-script output with the same rigor applied to the Framework A results.

Extended research notes, for qualified readers.

Longer-form methodology papers are in preparation. Full versions — including the reconciled feature mathematics and labeling specification withheld from this page — are available to research partners under a research-access agreement.