Research and Engineering at Vyreon Labs

Vyreon Labs is the result of a multi-year independent research program into financial time series, options-market structure, signal processing, probabilistic estimation, uncertainty, and live analytical systems.

This page documents how the research developed. It includes methods that reached production, systems that were live-tested and retired, hypotheses that failed, implementation errors that changed the interpretation of earlier results, and the technical lessons that led to the current SPY Financial Telemetry platform.

The purpose is not to present every experiment as a discovery.

The purpose is to show the engineering record:

Form a hypothesis, implement it, test it under causal and live conditions, investigate failure, and retain only what survives.

For the current production system, see the Methodology. For the professional background of the developer and operator, see About Sheldon Glowatski.

How To Read This Page

Historical research notes are not automatically descriptions of the current system.

The archive contains provisional interpretations, discarded models, enthusiastic working hypotheses, assistant-generated suggestions, debugging records, and results that later failed under walk-forward or live testing. Those materials are useful because they show the development process, but they should not be treated as present-day production truth.

Vyreon follows five source-of-truth rules:

1. Current production code takes priority over historical notes.
2. A backtest result is not treated as live evidence.
3. A visually convincing chart is not treated as proof of causality or tradability.
4. A research hypothesis remains provisional until its definitions, timing, and validation are checked.
5. When a material implementation error changes an interpretation, the corrected result replaces the earlier claim.

Research Principles

Preserve Causality

A model that uses future information is not a forecasting model, even when the leakage is subtle.

Causality applies to:

• Feature construction.
• Filter windows.
• Normalization statistics.
• Rolling return baselines.
• Delayed target maturation.
• Model prediction order.
• State updates.
• Uncertainty scoring.
• Validation joins.

This principle was learned through direct failure. Early forecasting experiments included lag-direction and alignment mistakes that made historical results look stronger or temporally cleaner than they were. Those mistakes changed the project permanently. Later architectures were designed around prediction-first order, frozen prior states, delayed labels, and explicit warmup.

Separate Structure, Direction, And Execution

A model can correctly detect that a system has changed without predicting how the new state will resolve.

It can also predict a broad direction and still fail because the instrument, entry timing, transaction mechanics, or holding period are wrong.

The 1DTE project made this distinction unavoidable:

• Structure: Did the prior market state break?
• Direction: Did the break resolve upward or downward?
• Velocity: Did the move happen quickly enough to exploit?
• Execution: Could the selected instrument capture the move after spread, decay, and fill behavior?

These are different engineering problems.

Publish Uncertainty, Not Only A Central Estimate

A single expected value is incomplete.

The current research program emphasizes predictive ranges, covariance, innovation variance, coverage, error stability, and visible bias. This is a direct response to earlier systems that produced visually compelling point forecasts without an honest representation of uncertainty.

Treat Live Operation As An Experiment

Live deployment is not only a final release step. It is a source of evidence.

A production or paper system reveals:

• Data-provider behavior.
• Market timing constraints.
• Platform resets and warmup requirements.
• Missing or delayed fields.
• Fill and execution assumptions.
• Edge effects.
• Model-state discontinuities.
• Operational monitoring gaps.
• Failure modes that a notebook does not reproduce.

Keep Failures Legible

A failed approach can still produce durable technical knowledge.

The research archive preserves why systems were changed, not only what replaced them. That record helps prevent the same attractive mistake from being rediscovered under a new name.

Development Timeline

2022 To 2023: Independent Systems Research Begins

The work that became Vyreon began as independent research into market structure, automated trading, and whether noisy financial data could be treated as a systems-engineering problem.

The early objective was broad: build software capable of ingesting market data, generating analytical features, forecasting future behavior, and eventually supporting automated portfolio decisions.

At this stage, the project was not yet the current public telemetry platform. It was an expanding experimental environment.

2024: Direct Forecasting, Feature Engineering, And Wavelet-RNN Research

The 2024 work focused on time-series forecasting.

Architectures and tools explored included:

• LSTM and GRU networks.
• CNN-LSTM hybrids.
• Attention mechanisms.
• DeepAR and AutoGluon forecasting.
• PatchTST and transformer-based concepts.
• Wavelet decomposition and wavelet reconstruction.
• Fourier and Hilbert-domain features.
• Technical indicators and lagged variants.
• LightGBM feature ranking.
• PCA and dimensionality reduction.
• Probabilistic mean and variance outputs.
• Walk-forward and out-of-sample testing.

The central question was whether multiscale and engineered features could make financial time series sufficiently stationary and informative for useful forecasting.

Several results appeared excellent in static plots. Some showed high apparent fit, narrow intervals, or visually convincing trend reconstruction.

Then the validation process exposed fundamental problems.

One major issue was a lag-direction error. Features intended to represent historical lag values were shifted in the wrong direction, which introduced future values. Another issue misaligned walk-forward predictions and realized data by the prediction-window length. These mistakes explained why some predictions appeared reversed or unusually well aligned.

The project did not hide those failures. They became part of the operating methodology.

The main lessons from this phase were:

• Feature timing must be audited independently from model architecture.
• A deep model cannot rescue a contaminated target or leaked feature set.
• Increasing layer width or depth does not solve poor information content.
• Feature importance can identify useful transforms, but importance inside one sample is not proof of generalization.
• Wavelets are useful for multiscale representation, but decomposition order, boundary effects, symmetry, phase behavior, and reconstruction definitions matter.
• Walk-forward validation is more informative than one static train-test plot.
• Direct price forecasting was not a stable foundation for the system Vyreon was becoming.

Early 2025: Market Physics And Probability-Mass Structure

The research shifted from forecasting price directly toward representing the options market as structured probability mass across strike and expiry.

The options chain was treated as a matrix rather than a collection of individual contracts. The work explored:

• Open-interest distributions.
• Implied-volatility distributions.
• Call and put structure.
• Strike-weighted centroids.
• Sensitivity-weighted centers.
• Distribution width and curvature.
• Centroid movement through time.
• Relationships between price and structural centers.
• Differences across expiration horizons.

Physical analogies were used to generate hypotheses about field behavior, equilibrium, damping, resonance, energy, phase, and bifurcation.

Those analogies were exploratory tools. They did not establish that markets obey literal mechanical laws.

Some early mathematical interpretations did not survive.

The most important example was the damping-ratio work. Apparent negative damping behavior was traced to a calculation error involving an omitted absolute value. A damping ratio does not become negative under the intended definition. The broader damping interpretation was then retired as a primary signal.

The useful knowledge survived:

• Distribution shape matters.
• A discrete probability-mass centroid can be more stable and computationally practical than an elaborate fitted distribution.
• Multiple maturity buckets describe different market processes.
• Structural relationships should be measured before a directional story is assigned.
• A physically inspired hypothesis must still be validated statistically and causally.

January To July 2025: Real-Time Wavelet Bifurcation Detection

During this period, Sheldon built a system intended to detect real-time structural transitions in intraday markets.

The project combined high-frequency data engineering, multiscale signal processing, higher-order derivatives, machine learning, live observability, and automated options execution.

The resulting system became the clearest example of a model solving one problem and being deployed into another.

August To November 2025: Live Deployment, Phase Research, And Architecture Pivot

The bifurcation system continued through live paper operation and expanded into additional research on:

• VWAP and options-derived centroids.
• Continuous wavelet transform power.
• Spectral entropy and bandwidth.
• Phase-locking value.
• Lead-lag relationships.
• Intraday resonance and coherence.
• Regime gating.
• Mean-reversion and momentum separation.

The work produced useful event-level diagnostics, but it also exposed a structural limit: intraday open interest was effectively static and could not provide the dynamic flow state the regime model expected.

The project began moving toward daily options-flow measurements, expiry-bucketed targets, and slower market-state priors.

SimBot was shut down on November 2, 2025. The final notes recorded four core reasons:

• The original CNN did not generalize out of sample.
• A TCN-GRU fusion model improved generalization but still lacked a reliable intraday directional measure.
• Open interest was static intraday, removing a key expected input.
• Calls and puts were not symmetric in the way the strategy assumed.
• Overnight theta imposed an approximately 15 percent headwind on the 1DTE instrument under the tested setup.

That retirement was the practical end of the intraday 1DTE path.

Late 2025 To 2026: Flow Barometer And Production Telemetry

The architecture pivoted toward a slower, multi-horizon system.

The new research program emphasized:

• Expiry-bucketed options structure.
• Open-interest and sensitivity centers.
• Causal forward-return targets.
• Rolling baselines built from matured outcomes.
• Regularized learned relationships.
• Recursive state estimation.
• Predictive covariance.
• Innovation variance.
• Delayed-label processing.
• Calibration and drift monitoring.
• Public reporting.

This became the current Vyreon platform.

The objective changed from producing a binary trade trigger to maintaining a stable market-state estimate with explicit uncertainty and continuous model-health evidence.

February To April 2026: Expiry Structure, Public Reports, And Backend Engineering

After the public telemetry platform launched, the research split into two connected paths.

The first path studied how the options matrix changes around known expiration boundaries. The work examined open-interest centroids by expiry, aggregate KOI movement, gamma concentration, the difference between pre-expiry drift and the observable post-expiry field step, and the delayed price response that followed the removal of option inventory.

The second path built the operational system required to preserve and query live evidence. A Gmail ingestion worker, PostgreSQL database, derived-data layer, and read-only FastAPI service were deployed. The data model retained every raw telemetry emission, deduplicated exact payloads, and selected the latest valid arrival as canonical truth for each node and date.

The public report also changed during this period. Early language sometimes converted a measured state into a stronger directional story than the system supported. The reporting method was revised to separate measured expectations, uncertainty, options structure, volatility state, and execution implications.

May To June 2026: Post-Launch Volatility And Eigenstructure Research

Once a live archive of multi-horizon expectations and matured innovations existed, the platform could be studied as a state observer rather than only as a forecasting system.

The main research questions became:

• Does innovation magnitude behave like a real-time volatility-state measurement?
• Does the change in innovation magnitude identify transition risk?
• Does cross-horizon geometry contain a separate instability measurement?
• Can any of those measurements improve risk control without pretending to predict direction?

The resulting work linked smoothed innovation magnitude to realized volatility, VIX, volatility-term-structure stress, and short-horizon movement. A separate eigenstructure study built a four-dimensional state from the horizon expectations, tracked rotation of the dominant covariance loading, and applied a stationary wavelet transform to that instability signal.

The strongest common result was not directional forecasting. It was observability of volatility, stress, structural reorganization, and movement probability.

July To August 2026: Estimator Audit And Research-Publication Expansion

The production estimator underwent a detailed timing and covariance audit.

The review found that the original uncertainty path mixed information from different points in the update cycle. Correcting that sequence exposed a second issue: the filter began with an identity coefficient covariance even though the Ridge model had been trained on thousands of observations.

A full training-derived covariance was added to the model artifact, all other artifact fields were checked for exact parity, and the node was rebuilt from a fresh warmup. The pathological startup interval disappeared while the central expectation paths and average errors remained close to the earlier production behavior.

In parallel, the public site began expanding from a weekly archive into a knowledge base with methodology, chart guidance, glossary, FAQ, author, and research pages.

Selected Research Program 1: Wavelet-RNN Forecasting

Wavelet-RNN Objective

Determine whether multiscale signal decomposition and recurrent models could improve forecasting of noisy financial time series.

Experimental Design

The work decomposed price and volume-related series into wavelet components and engineered additional features from:

• Wavelet detail and approximation levels.
• Log returns and higher-order differences.
• Rolling means and standard deviations.
• Volume differences.
• Fourier spectral summaries.
• Hilbert instantaneous frequency.
• Trend and momentum indicators.

The project compared broad feature sets with smaller, selected sets. It tested deeper and wider recurrent architectures, different context lengths, dropout, regularization, attention, PCA, and probabilistic outputs.

What Appeared Promising

Wavelet features repeatedly ranked as useful within the experimental feature-selection process. Smaller feature sets often produced more stable forecasts than large collections of correlated indicators. Denoised or multiscale representations could preserve trend structure while reducing high-frequency noise.

What Failed

Several visually strong results did not survive walk-forward testing.

The source record also contains repeated corrections to:

• Lag direction.
• Prediction alignment.
• Decomposition order.
• Interpretation of wavelet levels.
• Boundary behavior.
• Phase shifts introduced by asymmetric wavelets.
• Feature-window length.

PCA sometimes produced a clean fit and tight confidence interval while still failing walk-forward validation.

Wavelet-RNN Status

Exploratory and discontinued as a production forecasting path.

The work remains relevant as evidence of signal-processing, model-development, feature-selection, and validation experience. Wavelet methods later contributed to the bifurcation detector, but the Wavelet-RNN itself is not part of the current public system.

Selected Research Program 2: Real-Time Bifurcation Detection

Bifurcation-Detection Objective

Detect when intraday market structure becomes unstable and transitions into a new state.

The detector was designed to answer when a structural change occurred. It was not originally designed as a complete directional forecasting system.

Data Engineering

The system ingested noisy, multi-source price, volatility, volume, and options-derived measurements.

The pipeline handled:

• Data collection and synchronization.
• Cleaning and missing-value handling.
• Resampling to a common cadence.
• Causal feature construction.
• Stationary and scale-aware transformations.
• Real-time model input assembly.
• Live logging and alerting.

The final event representation used approximately five hours of market history at two-minute resolution.

Eight-Channel Event Fingerprint

The primary model input was an eight-channel by 150-timestep matrix.

The channels were built from continuous wavelet transform measurements and their temporal behavior, including:

• Spectral bandwidth.
• Resonance energy.
• Weighted wavelet scale.
• Wavelet phase mean.
• Wavelet phase standard deviation.
• Spectral entropy.
• First derivatives.
• Second derivatives.

The objective was to represent the geometry of a transition rather than only the price path.

Lightweight Convolutional Classifier

A one-dimensional convolutional neural network was trained to recognize transition motifs such as:

• Slope inversion.
• Curvature collapse.
• Spectral expansion.
• Derivative sign changes.
• Concentrated multiscale energy.

The CNN was intentionally lightweight because the input already contained engineered structural information.

Template-Matching Gate

To reduce false detections, the model output was combined with Gaussian pulse template matching across:

• The transformed signal.
• Its first derivative.
• Its second derivative.

All three representations were required to align before an event was accepted.

This created a deterministic confirmation layer around the learned classifier.

Live Observability

The detector was deployed into a live paper-trading system operating on SPY and QQQ.

The live workflow included:

• Continuous market-data ingestion.
• Real-time feature generation.
• Event classification.
• Discord notifications.
• Operator review while the system ran unattended.
• Automated option selection and execution.
• Trade logging and performance diagnostics.

The archive contains a library of live event detections and screenshots. Those observations support that the system repeatedly identified meaningful intraday transitions. They are not a substitute for a pre-registered, independently audited precision and recall study.

What The Model Knew

The detector could identify when the prior state had become unstable.

It did not reliably know whether the new state would resolve upward or downward.

That distinction became the central finding:

The model knew when to trade. It did not know which direction.

A later TCN-GRU fusion model with self-attention improved out-of-sample behavior. It did not create directional information that was absent from the inputs.

Why The Trading System Failed

The strategy deployed the detector into 1DTE options.

Three operational forces dominated the outcome.

1. Confirmation Arrived Too Late

Entries often occurred near 3:00 PM.

The model waited for the structural break to become sufficiently clear. In historical examples, bifurcation energy sometimes unfolded over one or two days. In the live market, transitions could resolve within minutes or hours.

By confirmation, much of the usable intraday movement had already occurred.

2. Overnight Theta Changed The Zero Point

The tested 1DTE options lost approximately 15 percent of their value overnight from theta under the observed setup.

The strategy therefore did not begin each directional decision at neutral expectancy. It began with a material decay headwind.

A 50/50 directional assumption with asymmetric upside can appear profitable in simplified math. It is not the same problem after an approximately 15 percent overnight cost is applied repeatedly.

3. Structural Instability Was Treated As Velocity

A bifurcation means the prior state has become unstable or changed.

It does not guarantee immediate directional acceleration.

The detector solved a structure problem. The execution vehicle required rapid and correctly directed price velocity.

The core postmortem was:

The system solved structure. It was deployed into velocity.

Final Outcome

The paper strategy began with an account of approximately $750,000 and slowly declined through repeated short-dated overnight trades. The system was retired on November 2, 2025.

It did not establish a profitable trading strategy.

What Survived From The Bifurcation Deployment

The project produced durable engineering knowledge:

• Multiscale transforms can produce compact event fingerprints from noisy intraday data.
• Higher-order derivatives and deterministic templates can create selective event gates.
• Structural-event timing and directional forecasting are different problems.
• Model generalization and execution suitability must be tested separately.
• Short-dated derivatives magnify small timing errors.
• Live operation exposes failure modes that historical simulation can hide.
• Alerting, observability, and post-trade diagnostics are part of the analytical system.
• A technically valid detector can still be economically useless in the chosen deployment environment.

The final lesson was not that the detector was worthless. It was that correct detection was insufficient for the instrument and strategy built around it.

Read the full 1DTE engineering postmortem

Selected Research Program 3: Options Probability-Mass Centroids

Centroid-Research Objective

Represent options-market structure through the location, width, and movement of weighted distributions across strike and expiry.

Why Centroids Became Important

Early approaches attempted to fit skew-normal distributions, mixtures, and other parametric forms to the options matrix.

Those methods introduced several problems:

• Slow and unstable optimization.
• Poor behavior on multimodal single-stock chains.
• Sensitivity to fit assumptions.
• Difficulty maintaining real-time performance.
• Additional transformation layers that could hide simple data problems.

The research returned to the actual objective: estimate the center of probability or structural mass.

A discrete weighted centroid was simpler, faster, and easier to audit.

Structural Meaning

A centroid is not a guaranteed target, support level, resistance level, or direct statement of trader intent.

It is a compact description of where a selected form of market structure is concentrated.

Different weighting schemes can describe different objects:

• Open-interest concentration.
• Volatility-sensitive concentration.
• Gamma and vega sensitivity mass.
• Implied-volatility basin location.

The current public system uses structural centers as features within a broader model. It does not assume that price must move toward them.

What Survived From Centroid Research

Probability-mass centroids became one of the durable foundations of Vyreon.

The surviving insight was not that one centroid predicts price by itself. It was that the options matrix contains measurable structure that can be summarized across maturity buckets and combined with price context, uncertainty, and later realized outcomes.

Centroid-Research Status

Retained and developed into the current production framework.

The exact feature definitions, trained weights, and operational thresholds remain proprietary. The public methodology explains the concepts and interpretation.

Selected Research Program 4: Multi-Horizon Probabilistic Telemetry

Telemetry-Research Objective

Build a causal system that estimates market state across several horizons, represents uncertainty, adapts as delayed outcomes mature, and publishes evidence about model health.

Horizon Design

The current public system separates the options matrix into four buckets:

• Near-term, 8 to 30 calendar days.
• Short-term, 31 to 60 calendar days.
• Medium-term, 61 to 120 calendar days.
• Long-term, 121 to 365 calendar days.

Each bucket has its own inputs, target history, coefficient state, uncertainty, and maturation delay.

The system does not force the horizons to agree. Their disagreement is part of the output.

Horizon-Averaged Targets

The target is not one terminal return on one exact day.

Several future dates are sampled across each bucket. Their returns are averaged to produce a horizon state. This reduces dependence on one endpoint and preserves the meaning of the full maturity range.

The complete target remains unavailable until the final sampled date has occurred.

This creates a delayed-label system. Near-term outcomes mature sooner. Long-term outcomes may take approximately one year.

Causal Baseline

The system models excess return relative to a rolling baseline of previously matured returns.

The baseline is calculated before the current target is added. This prevents the current outcome from influencing its own reference value.

The public total expected return adds the current baseline back to the modeled excess-return estimate.

Learned Model And Recursive State Estimation

A regularized linear model provides the initial relationship between processed structural features and excess return.

The linear base was chosen for stability, interpretability, auditability, and lower overfitting risk.

An adaptive Kalman coefficient estimator then allows the learned relationship to change gradually through time.

The production update follows prediction-first order:

1. Freeze the prior coefficient state and covariance.
2. Form the prior expected value and uncertainty.
3. Observe a newly matured outcome.
4. Calculate and score the innovation against the frozen prior.
5. Update the coefficient state and covariance.
6. Update the innovation-variance estimate for the next observation.

There is no backward smoothing pass used to rewrite a previously published live signal.

Predictive Uncertainty

The public expected range combines coefficient uncertainty, feature projection, innovation variance, and a small allowance for state evolution.

The result is closer to a prediction interval than a parameter confidence interval. It is intended to contain future realized horizon outcomes at approximately the stated frequency over many observations.

August 2026 Estimator Timing And Covariance Audit

A production review found that the original implementation did not use one consistent information set for every uncertainty quantity.

The central mean was formed from the prior coefficient state. However, the adaptive observation variance was updated with the current innovation before that same innovation was scored, and some matured diagnostics reconstructed uncertainty from coefficient covariance after the observation had already been assimilated.

The corrected order is now:

1. Form the predictive coefficient covariance from the prior state plus process covariance.
2. Read and freeze the observation variance before the current outcome is used.
3. Form the prior mean, innovation, predictive variance, standardized innovation, and interval-containment result.
4. Update the coefficient state and covariance.
5. Update the innovation-variance tracker for the next observation.

The covariance update uses a numerically stable Joseph form.

The prior-consistent calculation exposed a separate initialization defect. The earlier filter used:

P0 = identity matrix

That was a diffuse numerical prior rather than an estimate of uncertainty in the trained Ridge coefficients. When it was finally projected through the true prior interval, the first long-horizon band expanded to several hundred percent.

The replacement covariance was derived from the same standardized training matrix, Ridge penalty, and residual variance used to create the production coefficient prior:

P0 = R0 A^-1 (X'X) A^-T
A  = X'X + alpha I

The test artifact was required to reproduce the existing feature order, scalers, Ridge coefficients, intercepts, residual variance, and Kalman hyperparameters exactly. The only substantive addition was a full per-bucket coefficient covariance matrix. Each matrix was finite, symmetric, and positive semidefinite.

A full warmup showed that:

• The pathological long-horizon startup interval disappeared.
• Expected-return paths remained visually close to the earlier production paths.
• Average errors changed only modestly.
• Interval coverage remained conservative rather than saturating at 100 percent.
• The innovation-based volatility signal remained strongly related to realized volatility.

This audit did not prove that every uncertainty parameter is optimal. It did establish a coherent update order and a defensible initial covariance.

Telemetry-Research Status

Current production.

The full implementation is described on the Methodology page.

Selected Research Program 5: Innovation-Based Volatility And Validation

Innovation As A State Measurement

An innovation is the difference between a matured realized outcome and the expectation formed from the prior state.

innovation = realized outcome - prior expected outcome

Small innovations mean the market remained close to the prior estimated state.

Large innovations mean realized behavior departed more strongly from prior expectations.

Vyreon measures innovation magnitude across horizons and smooths it into a volatility-state signal.

The signal is directional only in magnitude. A large positive or negative market move can both create a large innovation.

Independent Volatility Comparison

The public validation compares the innovation-based signal with independent realized-volatility estimators, including close-to-close volatility and Parkinson volatility.

The purpose is to test whether the innovation stream corresponds with observable market instability rather than arbitrary model noise.

Calibration

The validation framework examines several separate questions:

• Do realized outcomes remain inside the expected ranges at a reasonable frequency?
• Is average central error stable?
• Is there persistent one-sided bias?
• Does performance deteriorate after out-of-sample or live boundaries?
• Does the volatility signal continue to correspond with realized volatility?

High coverage is not automatically good. It can indicate intervals that are too wide.

Coverage must be read with interval width, average error, visible bias, observation count, and the exact validation convention.

Adaptive Versus Issue-Date Validation

Two related scoring methods must remain distinct.

Adaptive-filter coverage asks whether a newly matured label was plausible under the filter state immediately before that label was assimilated.

Strict issue-date coverage asks whether the outcome later realized inside the exact interval that was published on the original issue date.

The second requires an immutable forecast ledger and a later join to matured outcomes.

Both are useful. They answer different questions.

Post-Launch Volatility-State Findings

Internal research after launch tested the volatility signal against several independent measurements and practical risk questions.

The exact correlation changes with the volatility estimator, smoothing span, sample window, and later production revisions. Across the documented studies, smoothed innovation magnitude commonly showed strong same-date correspondence with close-to-close and Parkinson realized volatility. Historical research tables also reported high correspondence with VIX and lower, but still meaningful, correspondence with VIX3M.

The more important distinction was between level and change:

• RMS level behaved like a volatility-state measurement.
• RMS change contained transition information, but was much weaker as a contemporaneous level measurement.

Top-decile transition events were associated in the research sample with changes such as:

• Backwardation frequency increasing from about 7 percent to about 21 percent.
• Short-horizon stress frequency increasing from about 24 percent to about 47 percent.
• Higher future realized volatility and larger future absolute movement.

These results support a hazard-rate interpretation. A transition event changes the distribution of future conditions. It does not specify whether price will move upward or downward.

Risk-Overlay Experiments

The research then asked whether RMS should replace realized volatility or modify it.

The best documented result did not come from using RMS as a stand-alone trading signal. It came from using realized volatility as the primary regime measurement and allowing RMS to modify exposure.

One internal strategy comparison reported:

RV-only Sharpe:        approximately 1.27
RV plus RMS Sharpe:    approximately 1.31

The drawdown improvement was small but directionally favorable.

That result is exploratory. It does not establish an investable strategy, and it does not include every implementation cost or portfolio-design choice. Its value is narrower: the innovation signal appeared to contribute information that was not completely redundant with trailing realized volatility.

Several simpler approaches failed or underperformed:

• RMS as a direct replacement for realized volatility.
• Binary trading rules based only on RMS direction.
• Stand-alone options trades triggered by large RMS events.
• The assumption that a larger shock must produce a better trade.

The current interpretation is therefore conservative:

RMS is strongest as a market-state and transition measurement, and may be useful as a risk modifier. It is not a validated directional trading signal.

Validation-Research Status

Current production monitoring, with strict origin-time validation continuing as an explicit development area.

Selected Research Program 6: Dominant Eigenstructure Instability

Eigenstructure-Research Objective

Measure whether the organization of the four horizon states changes before periods of elevated market movement.

The research began with the standardized horizon vector:

x_t = [mu_1 / sigma_1, mu_2 / sigma_2, mu_3 / sigma_3, mu_4 / sigma_4]

A rolling covariance matrix describes how the four maturity projections move together. Its dominant eigenvector represents the strongest shared organization mode inside that rolling window.

The main instability measurement is:

loading_rotation_speed_t = ||v1_t - v1_(t-1)||

where v1 is the dominant eigenvector.

High loading rotation means that the dominant cross-horizon organization is changing quickly. Low loading rotation means the latent organization is comparatively stable.

Stationary Wavelet Decomposition Of Instability

A stationary wavelet transform using a Coiflet wavelet was then applied to the instability series, not directly to price.

The decomposition produced two useful descriptive layers:

• Detail energy: localized, higher-frequency turbulence inside the instability process.
• Approximation energy: slower structural reorganization and regime consolidation.

The recurring observed sequence was:

detail energy expands
-> loading rotation rises
-> approximation energy rises
-> instability subsides as structure relocks

This is an interpretation of the measured morphology, not a claim that every event follows an identical physical law.

Historical Reconstruction And Event Study

The state vector was reconstructed from causal walk-forward outputs over 2016 through 2024. The research used the prior expected value divided by predictive standard deviation as the historical expectation-state proxy.

Threshold studies on loading rotation found a graded hierarchy:

• Around 0.10 identified broad turbulence.
• Around 0.15 identified more meaningful structural reorganization.
• Around 0.20 selected rarer, larger transitions.

At the 0.15 threshold, the notebook reported about 49 events and movement-magnitude lift of approximately:

5 trading days:   1.19 times baseline
10 trading days:  1.13 times baseline
20 trading days:  1.21 times baseline

The immediate one-day effect was weak.

What Survived And What Failed

The research supports a non-directional conclusion:

Rapid rotation of the dominant cross-horizon loading is associated with a higher probability of later structural repricing and elevated movement.

It did not support a simple trade rule.

The following ideas failed or remained too weak:

• Instability spike means buy.
• A second spike alone means buy.
• Larger instability always produces better option performance.
• Wavelet peak-to-peak exit rules.
• Wavelet approximation decay as a robust trading exit.
• A general bullish double-loading sequence.

A small four-event subset involving a downward first resolution and a later second instability event looked favorable, but the sample was too small for statistical inference. A later double-loading hypothesis did not survive broader historical testing and is not retained as a directional signal.

Eigenstructure-Research Status

Retained as a causal structural-instability diagnostic. Not deployed as a directional or execution signal.

Selected Research Program 7: Expiry-Driven KOI Step Response

Expiry-Research Objective

Determine whether scheduled removal of option inventory produces a measurable change in the aggregate open-interest centroid and a delayed response in the underlying price.

Expiration is known in advance. The contracts, open interest, strikes, and scheduled removal date are observable before the event. That makes expiry a useful natural experiment.

The Timing-Definition Correction

The first event studies classified the field step using movement into the expiry-day close:

KOI(t0) - KOI(t0-1)

That was the wrong boundary for removal of expiring inventory because the expiry-day snapshot still contained the expiring contracts.

The corrected observable step is:

Delta KOI_post = KOI(t+1) - KOI(t)

where t is the final snapshot containing the expiry and t+1 is the first trading-day snapshot without it.

This correction resolved a sign inconsistency. Removing an expiry below the aggregate centroid should raise the remaining weighted centroid. The earlier event label sometimes appeared to say the opposite because it measured pre-expiry drift instead of the post-expiry field change.

The corrected variable is still not fully ex ante. It contains the first post-expiry snapshot, including new positioning that appears before that snapshot is recorded. A purely pre-event strategy would need to estimate the first-order centroid change mechanically from the inventory scheduled to expire, then validate that estimate against the observed post-expiry state.

Multi-Symbol Pooled Event Study

After correcting the boundary and pooling the available symbols, the research archive reported:

571 expiry events
264 post-step-up events
307 post-step-down events

The pooled impulse-response functions showed:

• A clear and persistent aggregate KOI reconfiguration after the expiry boundary.
• A smaller underlying-price response that developed over roughly one to three trading days.
• A stronger and faster average response for downward field steps than upward field steps.
• A contemporaneous or smoother relationship for kstar_all, which behaved more like a state descriptor than a clear forcing variable.

The unconditioned price response was intentionally muted because upward and downward events partially cancelled each other. Directional conditioning separated the response paths more clearly.

These studies support an association and timing structure. They do not, by themselves, prove a causal market law or a profitable options strategy.

Amplitude Modeling

The next question was whether pre-event market state could estimate the size of the response.

The event generator was expanded with volatility, trend, gamma, open-interest, centroid, curvature, and price-context features. Models were trained only on features available before the response window.

The notebook's held-out results reported approximate out-of-sample R-squared values of:

Maximum absolute move over days 1 to 3:  0.44
Mean absolute move over days 1 to 3:     0.42
Sum of absolute moves over days 1 to 3:  0.41
Directional mean target:                -0.08
Directional integral target:            -0.08

The learnable targets were movement amplitude. The directional targets were not learnable from the same pre-event state.

The most important features were predominantly volatility-capacity and local-state measurements, including rolling return standard deviation, Parkinson volatility, short trend slopes, selected gamma shares, and recent KOI and kstar motion.

This supports a useful decomposition:

• The scheduled expiry defines the event.
• The field reconfiguration defines a later observed direction.
• Pre-event volatility and structural state help estimate how much movement the market can express.

Negative Results Inside The Expiry Program

Several attractive simplifications did not survive:

• gamma_0dte_share was not a reliable linear amplitude scaler by itself.
• A simple product of KOI displacement and gamma share did not cleanly separate strike-crossing events.
• Static high-gamma filtering did not produce monotonic lift.
• Larger absolute KOI displacement did not always imply a larger price response.
• Fixed-dollar strike thresholds were not portable across symbols and price regimes.
• Some apparent directional conclusions changed after the expiry boundary was defined correctly.

ATR relative to spot showed a clearer upper-envelope relationship with future percentage movement. That result suggests volatility capacity matters, but the later hand-built nonlinear composites did not yet establish a stable production rule.

What Remains Before Deployment

The public research record should not describe this as a completed trading edge.

The remaining requirements include:

• A purely ex-ante estimate of the expiry-driven centroid step.
• Chronological walk-forward testing.
• Leave-one-symbol-out testing.
• Stability of feature importance across time and assets.
• Option-chain simulation with spreads, implied-volatility changes, fills, decay, and exit rules.
• Comparison against simpler baselines.
• Clear handling of overlapping and clustered expiry events.

Expiry-Research Status

Promising multi-asset system-identification research. Not part of the production telemetry model and not a validated trading strategy.

Selected Research Program 8: Telemetry Backend And Canonical Data Lineage

Backend-Engineering Objective

Build a production system that can preserve live telemetry, recover from replay and correction events, expose stable derived outputs, and support later validation without requiring manual reconstruction.

The deployed architecture is:

QuantConnect email payloads
-> ingestion worker
-> validation and parsing
-> PostgreSQL
-> derived-data layer
-> read-only API

Append-Only Source And Canonical Truth

The telemetry mailbox is treated as an append-only event stream.

Every raw message is retained as evidence. Parsed payloads are linked to their raw source. Exact logical duplicates can be identified by payload hash, while alternate payloads for the same node and date remain available for audit.

Canonical selection uses the latest valid received_at value for each node and date.

This supports:

• Accidental duplicate handling.
• Intentional resend and correction.
• Backfill after outages.
• Model-migration replay.
• Preservation of earlier versions.
• One clean current answer per node and date.

The core relational objects are:

• raw_email
• payload
• prior
• matured
• derived_horizon
• derived_volatility
• api_key

Derived Layer And API

For each canonical payload, the backend calculates stable return-space and price-space outputs, including expected mean, uncertainty, expected ranges, and innovation RMS.

The API separates internal and customer surfaces.

Internal routes can return full metadata, prior rows, matured rows, and derived outputs. Customer routes return a smaller derived contract. API keys are scope-based and stored as hashes.

Deployment And Fault Isolation

The stack is containerized into database, initialization, API, and worker services.

The worker can fail without corrupting PostgreSQL or taking down the read-only API. The API can continue serving the last valid canonical state while ingestion is temporarily unavailable.

Scheduled ingestion uses exact-time cron jobs plus a periodic fallback loop. This provides deterministic timing and recovery without introducing a larger orchestration platform.

Why The Backend Is Part Of The Research Method

The backend is not only product infrastructure.

It provides the data lineage needed to answer questions that the model alone cannot answer:

• What was emitted for a specific date?
• Was a date later corrected?
• Which payload was canonical at the time?
• Which issue-date prior later matured?
• Did a validation result use the exact forecast that existed before the outcome?

That makes the backend foundational to strict origin-time validation and reproducible post-launch research.

Backend-Engineering Status

Current production engineering. The system is live, replay-safe, canonicalized, and API-accessible.

Selected Research Program 9: Market-State Synthesis And Reporting Discipline

Reporting-Research Objective

Translate a dense probabilistic state into readable language without turning measurement into unsupported narrative.

The first public reports exposed a recurring failure mode. A mechanically correct measurement could be described too strongly.

Examples included:

• Treating a negative near-term expectation as a deterministic downside prediction.
• Treating an open-interest concentration as a price magnet.
• Treating a smoothed volatility measure as proof that instability must persist.

Review against later market behavior showed that the model could remain calibrated while the written interpretation overreached.

State Versus Execution

The reporting framework now separates:

• State: what the model measures.
• Trajectory: how the measured state is changing.
• Uncertainty: how wide the plausible range remains.
• Structure: where options inventory and volatility sensitivity are concentrated.
• Execution: how a user may combine that context with an independent trading process.

The report does not claim that state alone determines path.

First-Order Abstractions

The public reporting system favors first-order abstractions that can be traced directly to measurements.

Examples include:

• Expectation label.
• Directional slope.
• Strengthening or weakening state.
• Cross-horizon agreement.
• Expected-range width.
• Innovation expansion or compression.

Higher-order composite scores are treated cautiously because each additional layer can hide assumptions and encourage narrative drift.

The preferred reasoning path is:

measurement -> interpretation

not:

measurement -> opaque compression -> interpretation

Reporting-Research Status

Current publication methodology. The weekly report is treated as a measured state synthesis, not a trade recommendation or market narrative engine.

Failed Or Revised Hypotheses

Direct Price Prediction

Deep learning models repeatedly produced visually plausible forecasts. Many did not survive walk-forward validation or changing market regimes.

Conclusion: Direct point prediction was not a sufficiently stable foundation. The project moved toward state estimation and uncertainty.

Negative Damping Ratio

The apparent negative damping behavior came from a mathematical implementation mistake involving an omitted absolute value. The broader signal interpretation did not survive.

Conclusion: Invalidated as a primary signal. The surrounding lessons about careful definitions, dynamic state, and structural transitions remained useful.

Bifurcation Equals Direction

The detector identified structural instability but did not reliably encode whether the new state would resolve upward or downward.

Conclusion: Event timing and direction must be modeled separately.

Intraday Open Interest As A Dynamic Flow State

Open interest was effectively static during the session.

Conclusion: It could not act as the intraday regime variable originally expected. Daily changes were more appropriate for capital-flow measurement.

Calls And Puts As Symmetric Expressions

The live strategy treated calls and puts too symmetrically.

Conclusion: Option-side behavior, decay, skew, and execution cannot be collapsed into a simple mirrored rule without evidence.

Intraday Phase-Lock As The Regime Oracle

Phase and resonance metrics produced useful event diagnostics, but they were too noisy and context-dependent to define the complete market regime.

Conclusion: Retained as research context and possible tactical input, not the current slow state estimator.

More Features And More Model Capacity

Large feature sets and deeper models often produced better in-sample appearance but weaker generalization.

Conclusion: Complexity must be justified by out-of-sample lift, not architectural ambition.

PCA As A Universal Solution To Dimensionality

PCA sometimes created clean-looking fits and narrow intervals but did not consistently improve walk-forward behavior.

Conclusion: Explored and not retained in the current public methodology.

Pre-Expiry KOI Drift As The Expiry-Step Direction

The first directional split used KOI movement into the expiry-day close. That measurement occurred before the expiring contracts disappeared from the daily chain.

Conclusion: Timing definition corrected. The observable post-expiry step is measured between the final snapshot containing the expiry and the first snapshot without it.

Static High-Gamma Regimes As A Strike-Crossing Filter

High gamma and high 0DTE gamma share looked like intuitive conditions for a large move. The filtered samples did not produce reliable monotonic lift, and extreme gamma often described a strongly pinned environment rather than a freely propagating one.

Conclusion: Rejected as a simple static rule. Gamma remains a state variable, not a stand-alone trade filter.

A Larger Eigenstructure Event Must Imply A Better Trade

Moderate instability sometimes performed better than extreme instability, and simple option-entry rules were not stable across periods.

Conclusion: Unsupported. Eigenstructure instability is retained as a movement and transition diagnostic, not as a monotonic trade-strength score.

Double-Loading Eigenstructure As A Bullish Sequence

Two live examples suggested that a second loading period after an unresolved first period might precede an upward resolution. Later historical testing did not establish a reliable directional effect.

Conclusion: Not retained. The pattern remains an observation, not a production or research signal.

What The Current Platform Retained

The present Vyreon architecture is not every idea explored in the archive. It is the subset that remained useful after repeated correction and validation.

The current foundation includes:

• Options-market structure measured across strike and expiry.
• Probability-mass and sensitivity centers.
• Four distinct forward horizons.
• Causal feature processing.
• Horizon-averaged delayed targets.
• A causal rolling return baseline.
• Regularized linear estimation.
• Adaptive multidimensional Kalman coefficient states.
• Training-derived initial coefficient covariance.
• Prior-consistent predictive uncertainty.
• Innovation-based volatility.
• Coverage, error, bias, and drift monitoring.
• Replay-safe raw data retention and canonical payload resolution.
• A derived-data and read-only API layer.
• Daily production operation.
• Recurring public reports with traceable first-order interpretation.

The current platform does not depend on the discarded negative damping signal, the retired 1DTE trader, the original CNN direction assumption, or the historical Wavelet-RNN forecasts.

Research Method And Evidence

The project follows an iterative experimental loop:

1. Define a measurable technical uncertainty.
2. Build the smallest implementation capable of testing it.
3. Run historical, walk-forward, out-of-sample, or live experiments.
4. Record metrics and observed failure modes.
5. Inspect timing, data lineage, causality, and implementation definitions.
6. Retain, revise, or retire the approach.
7. Rebuild affected state when the implementation changes.

Evidence produced across the project includes:

• Chronological research notes.
• Mathematical derivations.
• Feature definitions.
• Architecture comparisons.
• QuantConnect research projects.
• Backtests and live deployments.
• Walk-forward studies.
• Pooled multi-symbol event studies and impulse-response functions.
• Model artifacts and validation outputs.
• Artifact-parity and covariance checks.
• Fresh-state production warmups after material estimator changes.
• Canonical raw payload history and correction records.
• Data-pipeline and infrastructure records.
• Discord alerts and live screenshots.
• Public telemetry reports.
• Calibration and volatility-validation charts.
• Failure postmortems.

The private whitepaper archive contains twenty sequential parts covering the development from early Market Physics hypotheses through the final SimBot and Flow Barometer pivot. Selected parts may be rewritten into public case studies after their claims are checked against the current implementation and surviving evidence.

Use Of Language Models In The Research Process

Language models have been used as engineering assistants for tasks such as:

• Organizing long research notes.
• Drafting code modules.
• Comparing possible mathematical approaches.
• Generating test ideas.
• Reviewing technical explanations.
• Converting structured outputs into readable prose.

The archive also documents a limitation of this workflow: generated suggestions can sound certain before they have been validated.

For that reason, assistant-generated interpretation is not treated as evidence by itself. Numerical claims must come from code, data, live output, or a documented experiment. Current production code and verified system behavior take priority over earlier conversational confidence.

Current And Future Research Directions

Strict Issue-Date Forecast Validation

Build and maintain an immutable forecast ledger that stores the exact mean and uncertainty published on each issue date, then joins those records to fully matured horizon outcomes.

The canonical backend now provides much of the data lineage required for this work. The remaining task is to preserve the exact issue-time interval as an immutable research object and score it only when the corresponding horizon has fully matured.

Expiry-Step Estimation And Amplitude Validation

Replace the observed post-expiry KOI step with a strictly pre-event estimate calculated from the inventory scheduled to expire. Compare the predicted field change with the first post-expiry snapshot, then test whether the amplitude model remains stable under chronological and leave-one-symbol-out validation.

Only after that should the research proceed to option-level execution simulation.

Volatility-State Risk Control

Test continuous exposure rules that combine realized volatility and innovation RMS. The objective is not to create a new directional signal. It is to determine whether the model can de-risk earlier, re-risk more coherently, or identify periods when short-volatility exposure is unusually hazardous.

Eigenstructure Hazard Modeling

Continue testing loading rotation as a non-directional transition and movement-risk measurement. Future work should emphasize threshold stability, event dependence, cross-asset behavior, and whether it adds information beyond RMS and ordinary realized-volatility features.

Additional Assets

Extend the architecture only to assets with sufficient options liquidity, strike coverage, expiry coverage, and data quality. Each asset requires separate validation.

Portfolio And Execution Layers

Investigate reinforcement-learning and adaptive-control approaches for portfolio allocation, position sizing, execution selection, and risk management.

These layers are future research. They are not part of the current public telemetry model.

Open-Source Engineering Work

Publish selected non-proprietary components, simulations, validation utilities, and documentation that demonstrate software and systems-engineering practice without exposing production coefficients, artifacts, or operational thresholds.

Research Case Studies

Potential future pages include:

• Wavelet-RNN forecasting and the discovery of leakage.
• Real-time bifurcation detection.
• The 1DTE execution postmortem.
• Probability-mass centroids.
• From direct forecasting to state estimation.
• Causal validation for delayed financial targets.
• Innovation magnitude as a volatility signal.
• Dominant eigenstructure rotation as a non-directional transition detector.
• Expiry-boundary KOI step-response research and its timing correction.
• The August 2026 Kalman timing and covariance audit.
• Building a replay-safe canonical telemetry backend.

Limitations

The research archive is extensive, but volume is not proof.

Important limits include:

• Many early experiments were exploratory rather than pre-registered.
• Some historical notes contain provisional or assistant-generated interpretations.
• Several research exports use stronger language than the preserved experiment can support. The public page deliberately downgrades those statements to the verified result.
• Live screenshots provide case evidence, not a complete statistical benchmark.
• Financial data can contain missing, stale, sparse, or vendor-dependent values.
• Overlapping forward targets and clustered expiry events create serial dependence.
• Pooled multi-symbol event counts do not imply independent observations.
• The observed post-expiry KOI step is not the same as a fully ex-ante step estimate.
• Held-out machine-learning results require stronger chronological and symbol-holdout replication before deployment.
• Option-level profitability has not been established from the expiry-response study.
• Long-horizon validation develops slowly.
• Historical calibration does not guarantee future calibration.
• A calibrated estimator is not automatically a profitable trading strategy.
• The public methodology does not expose all proprietary implementation details.

These limits are part of the research record. They are not exceptions hidden from the reader.