Innovation Magnitude as a Volatility-State Signal

Research Status

Current production measurement. Publicly monitored. Risk-overlay and trading uses remain exploratory.

The Vyreon model produces an expectation before a future horizon outcome is known.

When the outcome later matures, the difference between the realized value and the prior expectation is an innovation.

innovation
= realized horizon outcome
- prior expected horizon outcome

Innovation is a standard state-estimation concept.

It answers:

How different was realized behavior from what the model expected before the outcome was observed?

Vyreon combines innovation magnitude across horizons to create a volatility-state measurement.

The signal does not predict direction.

A large upward surprise and a large downward surprise both increase innovation magnitude.

Why Innovation Magnitude Is Useful

Price-based realized volatility measures how much price moved over a historical window.

Innovation magnitude measures how far realized behavior departed from the model's prior state.

The two should be related if the model is tracking a meaningful market state.

During orderly periods:

• Realized outcomes remain near expectations.
• Innovations are smaller.
• Innovation dispersion is lower.

During unstable periods:

• Realized behavior departs more strongly from prior expectations.
• Innovations become larger.
• Innovation dispersion rises.

This relationship creates a model-native volatility diagnostic.

From Individual Innovations To One Signal

Vyreon has multiple horizons.

Each matured horizon produces its own innovation.

A daily aggregate can be constructed from the available horizon innovations.

A common form is root mean square:

innovation RMS
= sqrt(mean(innovation_i^2))

RMS gives large deviations more weight than small deviations.

The raw series can be noisy, so the public validation also uses a smoothed version.

The report distinguishes:

• Raw innovation dispersion.
• Smoothed innovation trend.
• Change in the smoothed trend.
• Standardized comparisons with independent volatility measures.

The raw signal is closer to the primitive measurement.

The smoothed signal is easier to compare through time.

What The Signal Measures

The volatility signal reflects two connected ideas.

Innovation Magnitude

This is the direct size of the mismatch between expectation and outcome.

Alignment With Prior Expectations

The same realized return can create different innovations under different prior states.

A move that was already inside a broad expected distribution may be less surprising than a smaller move that occurred during a tightly estimated state.

Therefore, the signal is not simply another transformation of absolute price return.

It measures market movement relative to model state and uncertainty.

Public Validation

The public volatility-validation chart compares the smoothed innovation signal with independent realized-volatility measures.

The current comparison includes:

• Close-to-close realized volatility.
• Parkinson realized volatility.

All series are standardized before comparison because they use different raw units.

Recent public reports have generally shown same-date correlations in the low-to-mid 0.8 range.

Earlier internal variants, using different smoothing spans and samples, reported correlations ranging approximately from the high 0.8s into the low 0.9s.

The exact value changes with:

• Warmup treatment.
• Smoothing length.
• Validation window.
• Realized-volatility estimator.
• Kalman and covariance conventions.
• The period included in the chart.

The important result is not one record correlation.

The important result is that the relationship remains strong across in-sample, out-of-sample, and live portions of the available history.

[Insert volatility-validation chart]

Suggested caption: The innovation-based volatility signal is standardized and compared with close-to-close and Parkinson realized volatility. The chart evaluates shape and regime correspondence, not equality of raw units.

No Retained Lead-Lag Claim

The research tested whether the volatility signal consistently led or lagged realized volatility.

No stable lead or lag survived historical testing.

This point is important.

Realized volatility itself is computed over a trailing window, so a same-date comparison can contain apparent timing differences caused by the target construction.

The current research does not claim that innovation RMS predicts realized volatility a fixed number of days in advance.

The defensible claim is:

Innovation magnitude is a causal, contemporaneous measurement of market-state mismatch that strongly corresponds with realized volatility regimes.

Level And Derivative

The research separates the level of the signal from its change.

Signal Level

A high smoothed level means innovations have recently been large.

This describes an elevated volatility or instability state.

Signal Derivative

A rapid increase means innovation magnitude is expanding.

This can describe a state transition.

Internal tests found that the derivative had weak same-date correlation with changes in realized volatility, but stronger relationships with some future stress outcomes and movement measures.

This suggests a useful distinction:

RMS level -> current volatility state
RMS change -> transition or shock state

This distinction is promising, but it is not a stable directional forecast.

Stress-Event Studies

One internal event study examined top-decile telemetry shock days.

The reported conditional rates included:

• Backwardation rising from approximately 7.1 percent at baseline to 21.5 percent after a telemetry shock.
• Short-horizon stress rising from approximately 24.3 percent at baseline to 47.5 percent after a telemetry shock.

These results are large enough to justify continued research.

They should not be treated as final probabilities for future events.

The study still requires:

• Exact event-definition verification.
• Dependence-aware confidence intervals.
• Multiple subperiod tests.
• Asset expansion.
• Sensitivity analysis across thresholds.
• A locked definition of backwardation and short stress.

The result supports a hazard interpretation:

A telemetry shock marks a period in which the distribution of future volatility conditions has changed.

It does not identify the direction of price.

Volatility Term Structure

The research compared innovation measurements with:

• VIX.
• VVIX.
• VIX9D.
• VIX3M.
• Term spreads.
• Backwardation.
• Short-volatility stress.

Innovation RMS correlated strongly with VIX-level behavior in several internal tests.

This does not make it a replacement for VIX.

VIX is derived from option prices and represents a market-implied variance measure over a defined horizon.

Innovation RMS is derived from errors in the Vyreon expectation system.

The signals view the market through different mechanisms.

A useful framing is:

• VIX measures the option market's current implied variance.
• Realized volatility measures recent price movement.
• Innovation RMS measures the magnitude of mismatch between model state and realized behavior.

Their agreement provides evidence that the innovation stream is connected to real market instability.

Their disagreement may also be informative.

Risk-Overlay Experiments

The research tested whether innovation RMS could improve existing volatility-aware allocation rules.

The cleanest comparison used realized volatility as the primary regime measurement and RMS as a secondary exposure modifier.

One internal result reported:

• Sharpe of approximately 1.27 for a realized-volatility-only rule.
• Sharpe of approximately 1.31 for a realized-volatility plus RMS rule.
• Slightly improved drawdown behavior.

The improvement was modest.

That is preferable to presenting a small study as a new standalone trading system.

The result suggests:

RMS may add incremental state information when used as a modifier, even when realized volatility remains the primary allocation input.

The following ideas generally failed or remained weak:

• Using RMS alone as a directional strategy.
• Replacing realized volatility directly with RMS.
• Simple long-volatility trades after every shock.
• Wavelet exits based only on instability peaks.
• Assuming larger RMS shocks always create better option trades.

Why Conservative Coverage Matters

Innovation is scored against the predictive uncertainty that existed before the observation was assimilated.

A standardized innovation is:

z = innovation / predictive_sigma

If the uncertainty model is well calibrated, most standardized innovations should lie inside the expected range.

The current production system is conservative. Recent 95 percent intervals have covered approximately 99 percent of observations in several buckets.

That does not automatically invalidate the volatility signal.

It means:

• The predictive intervals are wider than a nominal 95 percent target would require.
• Raw innovation magnitudes remain meaningful.
• Standardized innovations should be interpreted with the conservative interval calibration in mind.

A signal derived from raw innovation RMS and a signal derived from standardized z are not identical.

The August 2026 Estimator Audit

A production audit corrected two uncertainty issues.

First, the current innovation had been used to update the adaptive observation variance before that same innovation was scored.

Second, matured uncertainty had been reconstructed from a posterior covariance that had already assimilated the outcome.

The corrected sequence is:

1. Freeze prior coefficient state and covariance.
2. Freeze prior observation variance.
3. Form the prior expected value and predictive variance.
4. Observe and score the innovation.
5. Update the coefficient state.
6. Update observation variance for the next observation.

The audit also replaced an arbitrary identity coefficient covariance with a covariance derived from the trained Ridge model.

After burn-in:

• The pathological startup interval disappeared.
• Central expectation paths remained stable.
• Average errors remained close.
• Volatility correlation remained strong.
• The innovation signal became slightly more reactive in some periods.
• Coverage remained conservative.

This strengthened the meaning of the innovation signal because each scored row now uses one coherent prior information set.

What The Signal Does Not Say

A high innovation signal does not mean:

• Price will fall.
• Price will rise.
• A crash is imminent.
• Implied volatility is cheap.
• Implied volatility is expensive.
• A long straddle has positive expectancy.
• A short-volatility position must be closed.
• One exact future volatility level will occur.

It says:

Realized market behavior has recently departed more strongly from the prior estimated state.

That is useful context.

It is not a complete trade.

Current Product Role

Innovation RMS is one of the strongest current Vyreon measurements because it is:

• Causal.
• Directly connected to the estimator.
• Easy to calculate.
• Interpretable.
• Comparable with independent volatility measures.
• Useful across model-health and market-state analysis.

The public report uses it to describe:

• Expanding instability.
• Compressing instability.
• Current innovation magnitude.
• Recent trend in the volatility state.
• Model alignment with realized behavior.

The primitive signal is more useful than forcing it into one opaque composite score.

Next Experiments

Dependence-Aware Shock Statistics

Repeat the stress-event study with block bootstrap or non-overlapping events.

Cross-Asset Validation

Test whether the same RMS construction remains useful after each asset's underlying expectation model is separately validated.

Continuous Exposure Models

Test smooth allocation functions rather than binary regime switches.

Comparison With Implied-Volatility Risk Premium

Evaluate whether RMS identifies periods when short-volatility carry becomes less attractive.

Calibration-Aware Standardization

Compare raw RMS, RMS divided by predictive sigma, and robust percentile transformations.

Stable Feature Release

If an API product is developed, expose:

• Raw innovation RMS.
• Smoothed RMS.
• RMS percentile or Z-score.
• RMS change.

Do not expose one proprietary composite as though it were the underlying observable.

Raw, Smoothed, And Standardized Forms

Several versions of the innovation measurement can be useful.

Raw RMS

The raw RMS reacts quickly.

It is also noisy because:

• Different buckets mature on different dates.
• Individual innovations can be large.
• The available horizon set can vary.
• One event can dominate the daily aggregate.

Smoothed RMS

An EMA reduces daily noise and makes regime shape easier to compare.

The smoothing span changes the result.

Short spans are more reactive.

Long spans produce higher persistence and can correlate more strongly with slower realized-volatility windows.

No one span is universally correct.

The public chart should state the smoothing convention.

RMS Z-Score

A rolling Z-score answers:

How unusual is the current innovation state relative to its own recent history?

This improves comparability across changing regimes.

It also depends on the rolling window and warmup.

RMS Change

The first difference or slope describes transition.

It is more sensitive to noise and should not be confused with the level.

Report Interpretation

The weekly report uses the current raw value, smoothed value, and change in the smoothed trend.

A simple classification can be:

• Raw below a falling EMA, compressing.
• Raw below a rising EMA, current relief inside a still-rising background.
• Raw above a rising EMA, expanding.
• Raw above a falling EMA, renewed disturbance during broader compression.

These labels describe the relationship among measurements.

They do not guarantee persistence into the next week.

Why Visual Agreement Can Improve While Correlation Falls

A full-sample correlation can decline even when a signal appears more responsive around turning points.

Correlation is affected by:

• Amplitude.
• Smoothing.
• One influential historical spike.
• Persistence after the peak.
• Early warmup behavior.
• Standardization window.

The August 2026 correction reduced the full-sample correlation slightly while the updated signal appeared more reactive in several regimes.

That is not enough to declare improvement.

It is enough to avoid treating one scalar correlation as the complete quality measure.

Additional diagnostics can include:

• Regime classification agreement.
• Peak timing.
• Drawdown-period behavior.
• Mean absolute standardized difference.
• Subperiod correlations.
• Sensitivity to the smoothing span.

Current Conclusion

Innovation magnitude is a valid model-state observable.

Its smoothed level tracks realized volatility regimes strongly.

Its derivative contains promising transition information.

Its most defensible use is as:

• A volatility-state measurement.
• A model-health measurement.
• A risk modifier.
• A transition-hazard input.

It is not a retained directional signal.

That distinction has survived repeated testing.