Methodology | Vyreon Labs

Vyreon Labs measures financial market structure through options data, price behavior, and probabilistic state estimation. The current public system is designed to describe the market as it exists now, estimate a range of possible future outcomes, and measure how closely realized behavior agrees with prior expectations.

The system does not try to name one exact future price. It does not treat one option contract as a complete signal. It does not assume that call activity is automatically bullish or that put activity is automatically bearish. Instead, it measures how information is distributed across many strikes, expiration dates, and time horizons.

This page explains the current production methodology in plain language. Some implementation details remain proprietary. These include exact feature weights, model coefficients, operational thresholds, and internal state values. The goal here is to explain what the system measures, how the measurements are processed, and how the public reports should be interpreted.

For a shorter introduction, see About Vyreon Labs. Definitions of recurring terms are available in the Glossary. A chart-by-chart guide is available at How to Read the Charts.

What The System Is Designed To Measure

A market price is the final result of many interacting decisions. It does not show the full structure that produced it.

The options market adds another layer of information. Each option contract has a strike price and an expiration date. It also has measurements such as open interest, implied volatility, trading volume, gamma, and vega. When these contracts are viewed together, they form a structured matrix across price and time.

Vyreon uses this matrix to estimate several parts of the current market state:

Where options inventory is concentrated across strikes.
Where volatility-sensitive exposure is concentrated.
How wide or narrow that structural field is.
How the implied volatility surface changes across strike prices.
How recent price behavior compares with the options structure.
How these relationships differ across short and long time horizons.
How uncertain the current estimate is.
How much realized market behavior differs from what the model previously expected.

These measurements are used to produce financial telemetry. The word telemetry is intentional. A telemetry system observes a process, updates its estimate as new information arrives, and reports both the estimate and its uncertainty.

Current Public Scope

The current public implementation measures SPY at a daily cadence.

SPY is used because it has a deep and active options market across many strikes and expiration dates. This provides enough structure for stable daily measurements. It also makes SPY useful as a broad reference for United States equity market conditions.

The system separates the options matrix into four forward horizons:

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 horizon has its own model state. This is important because a contract expiring in two weeks does not describe the same market process as a contract expiring in nine months.

Shorter horizons react more quickly to current positioning, event risk, and changing volatility. Longer horizons usually move more slowly and represent broader structural expectations. The horizons can agree, disagree, or change at different speeds.

Vyreon does not force them into one combined directional answer.

Step 1: Building The Options Matrix

The first stage begins with the daily SPY options chain.

The chain contains many contracts. Each contract is associated with a strike price, an expiration date, and a call or put classification. It can also contain open interest, implied volatility, volume, and sensitivity measurements such as gamma and vega.

The production pipeline applies quality filters before the chain is measured. Contracts with invalid values are removed. The system also limits the range of strikes around the current market price. This reduces the influence of extreme contracts that may have poor data quality or little practical relevance to the current field.

Weekly expiration dates are included. They add information to the shorter horizons and improve the time resolution of the matrix.

The remaining contracts are then divided into the four horizon buckets.

This creates four related views of the same underlying market. Each view contains a different set of expirations and a different distribution of strikes.

Step 2: Measuring Structural Centers

A large options chain is difficult to interpret contract by contract. Vyreon reduces the matrix into a smaller set of structural measurements.

One important measurement is the open-interest-weighted strike center. This is the weighted average strike of the contracts in a horizon, where contracts with more open interest have more influence.

This center describes where existing options inventory is concentrated. It does not prove that price will move toward that level. It does not prove support or resistance. It does not reveal whether the holder bought or sold the contract. It is a structural location, not a statement of intent.

Vyreon also calculates a volatility-sensitive center. This measurement uses option sensitivity and implied volatility to estimate where the volatility structure is concentrated across strikes.

The system also measures a combined sensitivity field using gamma and vega. In simple terms:

Gamma measures how quickly an option's directional sensitivity can change.
Vega measures how strongly an option responds to changes in implied volatility.

The combined field has a center and a width.

The center shows where the sensitivity mass is concentrated. The width shows whether that mass is tightly grouped or spread across a wider range of strikes.

A narrow field and a wide field can represent very different market structures, even when the current SPY price is the same.

Step 3: Measuring The Volatility Surface

Implied volatility is not the same at every strike.

The shape of implied volatility across strikes contains information about how the market is pricing uncertainty. Vyreon summarizes this shape within each horizon.

The current production feature set includes measurements of:

The slope of the fitted volatility surface.
The general level of the fitted surface.
The quality of the fit.

These measurements do not assume that the volatility surface is a perfect mathematical curve. The fit is a compact description of the available structure. Its quality measurement helps distinguish a clear surface from a noisy or fragmented one.

Step 4: Adding Price Context

Options data is not interpreted in isolation.

The system also measures recent SPY price behavior. This includes a relative true-range measurement and a regularized quadratic description of recent log-price history.

The price-history fit provides several simple measurements:

Curvature.
Slope.
Level.
Fit quality.

These measurements give the model context. For example, the same options structure may have a different meaning during a smooth trend than during an unstable reversal.

The price fit is not used as a conventional chart pattern. It is one input within a larger structural model.

Step 5: Forming The Constraint-Field View

The phrase constraint field describes the combined structure imposed by the options matrix.

This is not a literal physical force. It is a practical way to describe how open interest, volatility, sensitivity, strike placement, and expiration timing restrict the market's available paths.

Two markets can have the same current price but different constraint fields.

One may have tightly grouped positioning, low uncertainty, and aligned horizons. Another may have wide structural dispersion, conflicting horizons, and rapidly changing volatility.

Vyreon measures these differences through the relationships between:

Current price.
The positioning center.
The volatility-sensitive center.
The gamma and vega field.
The field width.
The implied volatility surface.
Recent price behavior.

No single distance is treated as a direct trade signal. The model learns from the complete feature vector for each horizon.

Step 6: Causal Feature Processing

Financial data is noisy. Raw daily measurements can change because of genuine market movement, contract availability, expiry rotation, or data variation.

The feature pipeline therefore smooths and standardizes the measurements before they enter the model.

The order is important.

First, the system calculates all structural measurements in their natural units. Ratios and differences are formed before normalization. This preserves their meaning.

Second, the system applies a causal Savitzky-Golay filter. This filter estimates the local level and motion of a series using only current and past observations. It does not use future values.

Third, the filtered output passes through an exponential moving average. The current production system uses the same smoothing rate across all four horizons. This shared setting was selected because it produced the strongest results while reducing the number of free parameters.

Fourth, the system applies a causal exponential Z-score. The current observation is compared with the mean and variance that existed before that observation. The running statistics are updated afterward.

This sequence produces model inputs that are smoother, more stationary, and easier to compare through time.

The full sequence is:

Raw options and price measurements
        ↓
Natural-unit structural features
        ↓
Causal local smoothing
        ↓
Exponential smoothing
        ↓
Causal exponential normalization
        ↓
Standardized model input

The causal order is a core requirement. A value used in a historical prediction must be based only on information available at that time.

Step 7: Defining The Forward Outcome

The model is trained against a horizon-averaged forward return.

This is different from using one terminal price at one exact date.

For each horizon, the system samples several future dates across the full bucket. It calculates the SPY return from the issue date to each sampled date. It then averages those returns.

This creates a smoother target that represents the general path of the horizon rather than one isolated endpoint.

For example, the 31 to 60 day horizon does not ask only what SPY did on day 60. It summarizes realized returns across several points between day 31 and day 60.

The full outcome is not known until the final sampled date has occurred. This creates a delayed-label system.

A near-term outcome becomes available sooner than a long-term outcome. The long-term horizon may require up to one year before the complete realized target is known.

This delay is expected. It is part of the horizon definition.

Step 8: Removing The Historical Return Baseline

Market returns have a persistent long-run component. A model can appear useful simply because broad equity markets often rise over long periods.

To reduce this effect, Vyreon models excess return rather than raw return alone.

For each horizon, the system maintains a causal rolling baseline of previously matured horizon returns. The current realized target is compared with this prior baseline.

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

The model target is therefore:

Realized horizon-averaged return
minus
Prior rolling mean of matured returns for that horizon

This excess-return target asks whether the current structure is associated with an outcome above or below the recent normal level for that horizon.

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

Step 9: The Learned Statistical Model

Each horizon uses a regularized linear model.

The model begins with a fixed relationship between the standardized structural features and the excess-return target. Regularization limits the size of the coefficients. This reduces sensitivity to noise and correlated inputs.

A linear model was selected for several reasons:

It is stable.
It is easy to audit.
It is fast enough for daily production use.
It reduces overfitting risk.
It produces a clear starting state for recursive estimation.

The model is trained separately for each horizon. This allows the same feature to have a different role at different maturities.

The trained model artifacts include the ordered feature list, feature scalers, initial coefficients, residual variance estimates, and recursive-filter settings.

These artifacts are frozen before live operation. The live system does not refit the historical model every day.

Step 10: Recursive State Estimation

A fixed regression assumes that the relationship between market structure and future outcomes never changes.

Markets do change. The same structural measurement can behave differently during a calm trend, a volatility shock, or a major repricing event.

Vyreon therefore places a recursive state estimator on top of the learned model.

A trained Ridge feature-to-return mapping initializes a multidimensional Kalman filter, whose coefficient state and covariance update recursively as delayed outcomes mature. This architecture can be described as a Ridge-initialized multivariate Kalman state-estimation system.

The estimator maintains two main objects:

A current estimate of the feature coefficients.
A covariance matrix that represents uncertainty in those coefficients.

Before a new matured observation is processed, the system creates a prior state. It allows uncertainty to grow slightly through a small process-noise term. This represents the possibility that the market relationship has changed since the previous update.

The system then calculates the expected excess return for the frozen feature vector.

When the realized horizon outcome becomes available, the system compares it with the prior estimate.

The difference is called the innovation.

Innovation = realized excess return - prior expected excess return

A small innovation means realized behavior stayed close to the prior model state.

A large innovation means the market behaved differently from what the prior state described.

The estimator then updates the coefficients and their covariance. The amount of adjustment depends on the uncertainty in the state and the estimated noise of the observation.

The current production implementation follows a strict prediction-first order:

Freeze the prior coefficient state and uncertainty.
Calculate the prior expected value.
Observe the matured outcome.
Calculate and score the innovation against the frozen prior uncertainty.
Update the coefficient state.
Update the innovation-variance estimate for the next observation.

There is no backward smoothing pass. Future observations are not used to revise a past live signal before it is published.

Step 11: Linking The Feature And Model Time Constants

The feature pipeline and recursive estimator both need a rate of adaptation.

Using many independently tuned settings would increase the risk of overfitting. The current production design therefore uses one shared exponential smoothing rate across the feature pipeline.

The recursive-filter parameters are then derived from that same rate.

This reduces the number of free parameters. It also keeps the input smoothing and model adaptation on related time scales.

The system still produces different behavior in each horizon because the underlying options structure, trained coefficients, residual variance, price context, and matured outcomes are different.

The horizons are not made different through arbitrary manual tuning.

Step 12: Expected Return And Predictive Uncertainty

The public expected return has two parts:

Expected total return
=
Current excess-return estimate
+
Current causal baseline

The system also produces a predictive uncertainty estimate.

This uncertainty combines:

Uncertainty in the current coefficient state.
The projected effect of that uncertainty on the current feature vector.
The estimated variance of model innovations.
A small allowance for state evolution before the next update.

The shaded 95% expected range is formed around the central expected return.

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

It is not a guarantee for one forecast.

A well-calibrated 95% range should still be exceeded sometimes. If almost no outcomes ever leave the range, the interval may be too wide. If many outcomes leave it, the interval may be too narrow.

Step 13: Positive, Negative, And Mixed States

The weekly report converts the expected range into a simple state label.

Positive means the full expected range is above zero.
Negative means the full expected range is below zero.
Mixed means the range crosses zero.

Mixed does not mean the model has no information. It means the estimated range includes both positive and negative outcomes.

A central mean can be positive while the state remains Mixed. This occurs when uncertainty is wide enough to include negative outcomes.

The report also compares the four horizons with one another. This produces descriptions such as aligned, partially aligned, fragmented, or unresolved.

These descriptions are generated by deterministic rules applied after the model output. They are not separate machine-learning predictions.

Step 14: The Volatility Signal

The innovation stream contains more information than simple forecast error.

When innovations remain small, realized market behavior is staying close to prior expectations. When innovations become larger, the market is departing more strongly from the prior model state.

Vyreon measures the magnitude of these innovations across horizons. The signal is then smoothed to show the broader trend.

This produces the volatility signal used in the public reports.

The volatility signal reflects:

Model innovation magnitude.
The alignment between realized market behavior and prior expectations.
Whether structural mismatch is expanding or compressing.

It does not indicate direction by itself.

A high value means the market is behaving less consistently with prior expectations. A low value means realized behavior is more closely aligned with the current model state.

The validation chart compares this signal with independent measures of realized volatility, including close-to-close volatility and Parkinson volatility.

Step 15: Validation And Calibration

Vyreon publishes model-health measurements because a probabilistic system should be evaluated against realized outcomes.

The validation process examines several questions:

Do Realized Values Remain Inside The Expected Range?

The system measures how often matured outcomes fall within the prior expected range used by the adaptive estimator.

Coverage is evaluated separately for each horizon.

Coverage above 95% does not automatically mean the model is better. It may mean the uncertainty estimate is conservative. Coverage below 95% may mean the range is too narrow, the central estimate is biased, or the market has changed faster than the model can adapt.

Are The Central Errors Stable?

The system compares realized returns with expected returns and measures the typical error size.

Stable errors suggest that model behavior has not changed sharply. Rising errors can indicate a regime change, a data problem, or weakening alignment between the model and the market.

Is There Visible Bias Or Drift?

A model can remain inside wide intervals while still being systematically too high or too low.

The validation chart therefore compares the direction and shape of realized and expected returns through time.

Persistent separation would indicate bias or drift.

Does The Innovation Signal Correspond With Realized Volatility?

The volatility validation compares the smoothed innovation signal with realized-volatility estimators.

Strong alignment supports the interpretation that the innovation stream is measuring a real volatility state rather than arbitrary model noise.

What The Current Validation Does Not Prove

Validation is a model-health measurement. It does not prove that every central expectation is directly tradeable. It does not prove that one forecast will be correct. It does not remove event risk or regime risk.

The adaptive coverage statistic measures the state immediately before a matured observation is assimilated. A separate origin-time ledger is required to test the exact interval published on every historical issue date against its later realized outcome.

These are related tests, but they answer different questions.

Step 16: Warmup And Live Operation

The model requires historical context before it can produce a mature state.

On startup, the production node runs a multi-year warmup. During warmup, it processes historical daily options matrices in chronological order.

This populates:

Causal smoothing windows.
Exponential means and variances.
Frozen feature histories.
Matured return labels.
Rolling baselines.
Recursive coefficient states.
Innovation variance estimates.
Coverage histories.

Telemetry is not sent during the historical warmup. Computation continues normally, but outbound reporting is disabled.

After warmup, the system begins sending daily priors. It also processes newly matured outcomes and uses them to update the state.

The daily cycle is:

Receive the current options matrix
        ↓
Build horizon-specific measurements
        ↓
Update causal feature streams
        ↓
Process any newly matured historical outcomes
        ↓
Update recursive model states
        ↓
Issue new expected returns and uncertainty ranges
        ↓
Store telemetry for reporting and validation

Step 17: From Telemetry To The Weekly Report

The weekly report is a communication layer built on top of the daily measurements.

It does not retrain the model. It does not create new market data. It translates the current state into readable language.

The public report includes:

A summary of the current market environment.
The current volatility regime.
Forward expectations across four horizons.
Changes from the previous reporting period.
Cross-horizon coherence.
A current options-market structure snapshot.
Model validation and volatility-signal validation.

The report uses fixed rules to reduce ambiguity. For example, a horizon is not called Positive unless its full expected range is above zero.

This prevents the narrative from becoming more certain than the underlying probability estimate.

Important Limitations

Every measurement system has limits. Vyreon is no exception.

Open Interest Does Not Reveal Intent

Open interest shows the number of outstanding contracts. It does not show whether one participant is bullish, bearish, hedging, providing liquidity, or offsetting another position.

A put-heavy options matrix is not automatically bearish. A call-heavy matrix is not automatically bullish.

Vyreon measures the structure and movement of the inventory. It does not claim to know the private intention behind every contract.

Structural Centers Are Not Guaranteed Price Targets

A positioning center or volatility center is a weighted summary of the options matrix.

It is not guaranteed support. It is not guaranteed resistance. It is not a deterministic attractor. Price can move away from every measured center.

The Current Core Model Is Linear

The production estimator uses a regularized linear relationship that adapts through time.

This improves stability and interpretability. It also means the model may not capture every nonlinear interaction or sudden discontinuity.

The Uncertainty Model Is An Approximation

The expected ranges use a linear state model and an adaptive innovation-variance estimate. Financial returns can be skewed, heavy-tailed, and affected by jumps.

A 95% range should be understood as a calibrated operating estimate, not a physical boundary.

Forward Targets Overlap

Nearby issue dates share some of the same future price path. This creates serial dependence in matured outcomes and innovations.

The model accounts for changing innovation variance, but the observations are not fully independent.

Long Horizons Update Slowly

A long-term forecast takes much longer to mature than a near-term forecast.

This means long-horizon live validation develops slowly. Recent long-term expectations cannot be judged immediately.

The Model Does Not Read News

The production system is based on market data. It does not directly read earnings releases, economic reports, political events, or news headlines.

Unexpected events appear through changes in market data and through later innovations.

Data Quality Matters

Options chains can contain missing values, stale measurements, changing contract availability, and differences caused by expiry rotation.

The system applies filters and quality checks, but no data pipeline is perfect.

The Current Public Implementation Covers SPY

The architecture is designed to be extended to other assets. The public production record currently focuses on SPY.

Results from SPY should not be assumed to transfer automatically to every stock, index, commodity, or volatility product.

This Is Not Financial Advice

The reports describe measured market conditions. They do not provide personalized investment advice, guaranteed returns, or instructions to buy or sell a security.

Readers remain responsible for their own decisions, risk controls, and due diligence.

Why The Methodology Is Structured This Way

The system follows several design principles.

Measure Before Predicting

The main product is a market-state measurement system. Expected returns are one output within that system.

Preserve The Time Horizon

Near-term and long-term options describe different processes. They remain separate throughout the pipeline.

Use Only Information Available At The Time

Causal processing is required from the first filter through the recursive state update.

Reduce Unnecessary Parameters

The system uses one shared feature-smoothing rate and derives recursive-filter settings from it. This limits manual tuning.

Adapt Without Constant Retraining

The regularized model provides a stable starting structure. Recursive state estimation allows that structure to change as new outcomes mature.

Publish Uncertainty And Validation

A central estimate without uncertainty is incomplete. A model without realized-outcome checks is difficult to trust.

Vyreon therefore publishes expected ranges, error behavior, coverage, and volatility-signal comparisons alongside the market report.

Summary

The current Vyreon methodology can be summarized as follows:

Build a daily SPY options matrix across strikes and expirations.
Divide the matrix into four forward horizons.
Measure positioning, volatility, sensitivity, field width, implied-volatility shape, and price context.
Process those measurements through a causal smoothing and normalization pipeline.
Use a regularized model to estimate horizon-specific excess-return structure.
Use recursive state estimation to maintain temporal consistency and uncertainty.
Add the causal return baseline back to form total expected returns.
Measure innovations as realized outcomes mature.
Use innovation magnitude as a volatility-state signal.
Publish the current state, expected ranges, structural context, and model-health diagnostics.

The result is not a single buy or sell signal. It is a continuously updated description of market structure, forward expectations, uncertainty, and model alignment across multiple time horizons.