How a Live 1DTE Bifurcation System Failed

Research Status

Live-tested and retired. The detector and strategy described here are not part of the current production telemetry system.

This project attempted to detect intraday market-state breaks and trade them with one-day-to-expiry options.

The detector was technically sophisticated.

It processed high-frequency data, wavelet structure, higher-order derivatives, a compact convolutional model, deterministic template matching, live alerts, and automated option execution.

It still failed as a trading system.

The reason was not one bug.

The detector solved a different problem from the one required by the execution vehicle.

The model detected structure. The trade required immediate, correctly directed velocity.

The Original Objective

The system was designed to detect bifurcations.

A bifurcation was defined operationally as a point where the previous local market state became unstable and reorganized.

The detector was intended to answer:

Is the market undergoing a meaningful structural transition now?

It was not initially built to answer:

Will the next move be up or down, and will it happen fast enough to overcome one-day option decay?

That distinction became decisive.

Live Data Pipeline

The system combined multiple intraday measurements.

The pipeline included:

• Data collection.
• Timestamp synchronization.
• Resampling to a common cadence.
• Cleaning and missing-value handling.
• Causal rolling transforms.
• Wavelet features.
• Higher-order derivatives.
• Model inference.
• Deterministic confirmation.
• Alerts.
• Option selection.
• Paper execution.
• Trade diagnostics.

The final event representation used approximately five hours of history at a two-minute cadence.

Eight-Channel Event Fingerprint

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

The channels represented features such as:

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

The model was not asked to infer all structure directly from raw price.

The input had already been transformed into a compact event fingerprint.

Lightweight CNN

A one-dimensional convolutional neural network was trained to recognize transition motifs.

The motifs included:

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

The CNN was intentionally small.

The research hypothesis was that feature engineering had already compressed the important geometry into a form that a lightweight classifier could recognize.

Deterministic Template Gate

The learned model was combined with Gaussian pulse template matching.

The transformed signal, first derivative, and second derivative were compared with expected transition shapes.

An event was accepted only when the model and the deterministic geometry agreed.

This gate increased selectivity.

It also delayed confirmation.

That tradeoff became important in live execution.

Live Paper Deployment

The system ran in live paper operation on SPY and QQQ.

It produced:

• Real-time detections.
• Discord alerts.
• Operator-visible diagnostics.
• Automated option selections.
• Paper trades.
• Post-trade records.

The archive contains many event screenshots that appear to align with meaningful intraday transitions.

Those examples support that the detector was measuring something real.

They do not prove a profitable or complete trading system.

What The Detector Did Well

It Detected Structural Change

The system often identified points where the recent market state stopped behaving normally.

It Created A Compact Multiscale Representation

The eight-channel fingerprint converted noisy price and volatility behavior into a manageable model input.

It Operated Live

The system survived real platform constraints, data irregularities, scheduling, and unattended operation.

It Produced Observability

The alerts and diagnostics made it possible to inspect the detector rather than treat it as a hidden black box.

It Separated Event Gating From Execution

The architecture could be analyzed layer by layer.

What The Detector Did Not Know

The detector did not reliably encode direction.

An instability event could resolve:

• Upward.
• Downward.
• Through a fast reversal.
• Through extended chop.
• Through a slow transition that outlasted the option.

A later TCN-GRU fusion model with self-attention improved generalization.

It did not create directional information that was absent from the inputs.

This was the central model limitation.

Why 1DTE Options Were Unforgiving

One-day options amplify small mistakes.

They are highly sensitive to:

• Entry time.
• Direction.
• Overnight decay.
• Implied-volatility changes.
• Spread.
• Gamma path.
• Gap risk.
• Whether movement occurs before expiry.

A broad structural detector can be useful and still be economically incompatible with this instrument.

Failure 1: Confirmation Arrived Late

The template gate and classifier required enough evidence for the transition to become clear.

Entries often occurred late in the session, near 3:00 PM.

In historical charts, some bifurcations unfolded over one or two days.

In live markets, many transitions resolved within minutes or hours.

By the time the detector confirmed the event, much of the usable price movement had already occurred.

The model was correct too late.

Failure 2: Overnight Theta Changed The Baseline

The tested short-dated options lost approximately 15 percent overnight in the observed setup.

This is an empirical approximation from the specific strategy and instruments. It is not a universal constant for every 1DTE option.

The important point is that the position did not begin the next day from neutral value.

A simplified strategy might assume:

correct direction -> gain
wrong direction -> loss

The real position began with a large decay headwind before the directional question was settled.

Failure 3: Structure Was Treated As Velocity

A bifurcation means the previous state became unstable.

It does not guarantee:

• Immediate acceleration.
• Persistent trend.
• Sufficient displacement.
• A move before option decay dominates.

The detector answered:

the state changed

The instrument required:

price will move far and fast in one chosen direction

These are different targets.

Failure 4: Open Interest Was Static Intraday

The research expected open interest to contribute to a dynamic intraday regime state.

In practice, open interest was effectively static during the session.

That removed one of the anticipated live flow measurements.

Daily open-interest changes were more suitable for tracking slower capital structure.

Failure 5: Calls And Puts Were Not Symmetric

The strategy treated calls and puts too similarly.

In practice, their behavior differed through:

• Skew.
• Implied volatility.
• Decay.
• Liquidity.
• Downside volatility expansion.
• Upside overwrite and supply.
• Path dependence.

A mirrored rule did not capture those differences.

Paper-Strategy Outcome

The paper account declined gradually through repeated short-dated overnight trades.

The system was shut down on November 2, 2025.

The strategy did not establish positive expectancy.

The exact paper-account starting value should be verified from preserved records before publication.

What The Postmortem Changed

Structure, Direction, Velocity, And Execution Became Separate Layers

The project stopped treating one event score as a complete trade.

Live Validation Became Mandatory

Historical event alignment did not reproduce the full economics of short-dated options.

Instrument Choice Became Part Of The Model

A signal is not useful in isolation from the vehicle used to express it.

Timing Error Became An Economic Variable

A detector that is several hours late can be functionally wrong even when its classification is correct.

State Estimation Became More Attractive Than Trade Triggering

The next architecture focused on stable multi-horizon state, uncertainty, and model health.

What Survived

The project produced reusable technical knowledge.

Multiscale Event Fingerprints

Wavelets, spectral summaries, and derivatives can create compact representations of noisy transitions.

Selective Deterministic Gates

Template matching can reduce false positives when the target geometry is known.

Live Observability

Alerts, logs, and post-trade diagnostics are part of the analytical system.

Model And Execution Validation Must Be Separate

A detector can generalize while a strategy loses money.

Missing Directional Information Cannot Be Created By Model Complexity Alone

The TCN-GRU result reinforced this point.

Short-Dated Derivatives Magnify Hidden Assumptions

Theta, skew, fills, and timing can dominate a modest statistical edge.

Relationship To Current Vyreon

The current telemetry system does not use the retired 1DTE strategy.

It retained the engineering lessons:

• Causal processing.
• Explicit uncertainty.
• Delayed-label handling.
• Live observability.
• Separation of state and execution.
• Conservative interpretation.
• Visible model-health monitoring.

The current system describes market state.

It does not automatically convert every state into an options trade.

Research Sequence

The system developed through several iterations.

Event-Shape Discovery

The first phase focused on identifying repeated multiscale transition shapes in historical intraday data.

The research examined:

• Higher-order price derivatives.
• Continuous wavelet-transform power.
• Spectral entropy.
• Spectral bandwidth.
• Phase behavior.
• Curvature and slope changes.
• Relationships among price, VWAP, and options-derived structure.

The goal was to describe the shape of a break before deciding how to trade it.

Learned Classifier

A compact CNN was then trained on the engineered event fingerprints.

The model output was treated as one component of the gate, not as sufficient proof by itself.

Live Confirmation Layer

Template matching was added because historical review showed that classifier confidence alone could accept shapes that were not complete bifurcations.

The combined gate required agreement among:

• Learned event probability.
• Signal template.
• First-derivative template.
• Second-derivative template.

Directional Augmentation

After the missing-direction problem became clear, additional recurrent and attention-based models were tested.

The TCN-GRU fusion model improved some out-of-sample classification behavior, but it did not create reliable directional resolution.

Execution Retirement

The project was retired when the live evidence showed that the combined detector, option vehicle, and holding logic did not form a profitable system.

This sequence is important because it prevents a misleading summary such as "the CNN failed."

The failure occurred at the system level.

Structural Classifier Versus Trading Forecast

A classifier can be useful when it identifies a condition such as:

the local state no longer resembles the preceding stable state

A trading forecast requires more.

For the tested 1DTE strategy, the required output was closer to:

the market will move in direction D
by magnitude M
before time T
with enough option value remaining after costs

The detector did not provide that complete output.

The distinction can be expressed as four layers.

Layer	Question	Result
Detection	Did a structural break occur?	Often useful
Direction	Which branch will resolve?	Unreliable
Velocity	Will sufficient movement occur quickly?	Unreliable
Execution	Will the option capture the move after decay and spread?	Failed

The detector's apparent success in the first row could not compensate for missing information in the remaining rows.

Why Live Screenshots Were Not Enough

The archive contains many screenshots where an alert occurred near a meaningful market move.

Those examples are valuable.

They show:

• The live system was operating.
• The event morphology was not invented after the entire history was visible.
• Some detections aligned with real transitions.
• The alerts were sufficiently selective to inspect individually.

They do not establish:

• Precision across all alerts.
• Recall across all true transitions.
• Profitability.
• Stable direction.
• Correct option selection.
• Independence from operator interpretation.

A live screenshot library is case evidence.

A production trading claim requires a complete event ledger and outcome table.

Economic Decomposition Of A 1DTE Trade

The value change of a short-dated option depends on several components.

A simplified decomposition includes:

option value change
approximately
delta contribution
+ gamma contribution
+ vega contribution
- theta
- spread and execution cost

For a one-day option, each term can be large relative to premium.

A correct directional move can still lose if:

• The move is too small.
• The move arrives too late.
• Implied volatility contracts.
• The entry spread is poor.
• Overnight decay is large.
• The contract starts too far out of the money.

This is why a price-event detector cannot be validated only against the underlying chart.

The contract path is the economic target.

Why The System Was Not Simply Retuned

Several possible rescue paths existed:

• Lower the confidence threshold.
• Enter earlier.
• Use longer-dated options.
• Add more directional features.
• Remove the template gate.
• Trade both directions.
• Use straddles.
• Change the holding period.

Each change would define a different strategy.

The project did not have enough evidence that one adjustment would solve the missing directional and velocity information.

Continuing to tune the same historical event library would also increase overfitting risk.

Retirement was the disciplined choice.

Engineering Lessons For Later Systems

Define The Economic Target Before The Model Target

The detector was trained around structural-event classification.

The strategy needed option-return expectancy.

Those are not interchangeable labels.

Preserve A Full Event Ledger

Every alert should store:

• Issue timestamp.
• Input state.
• Classifier score.
• Template score.
• Selected contract.
• Bid and ask.
• Fill assumption.
• Underlying path.
• Option path.
• Exit reason.

Test The Vehicle Independently

Before connecting a signal, the selected option rule should be tested against known underlying paths and realistic Greeks.

Use Live Operation To Reject Attractive Stories

The system produced enough impressive examples that it would have been easy to keep rationalizing.

The gradual paper loss forced the architecture to confront the full economics.

Current Conclusion

The bifurcation detector was not meaningless.

It repeatedly identified structural events.

The trading system failed because the output did not contain reliable direction or immediate velocity, and the 1DTE vehicle imposed severe timing and decay constraints.

The correct postmortem is:

A valid event detector was deployed as though it were a complete directional execution model.

That is a systems-engineering failure worth documenting.

It is also one of the reasons the current platform is more disciplined.