Wavelet Forecasting, Leakage, and the Move Away from Direct Price Prediction

Research Status

Historical case study. The Wavelet-RNN and related direct forecasting models are not part of the current Vyreon production system.

The 2024 research program began with a conventional question:

Can multiscale signal processing and deep learning produce useful forecasts from noisy financial time series?

The work explored recurrent networks, convolutional models, attention, wavelets, Fourier features, technical indicators, feature selection, dimensionality reduction, probabilistic outputs, and walk-forward validation.

Several intermediate results looked excellent.

Some were wrong.

The project became important not because a final price-forecasting network survived, but because the failures established the causality and validation rules used by the current telemetry system.

The Original Forecasting Problem

Financial price series combine:

• Slow trend.
• Short-term noise.
• Volatility clustering.
• Event shocks.
• Regime changes.
• Non-stationarity.
• Changing market microstructure.

A raw sequence model must separate those components while avoiding future information.

Wavelets appeared useful because they decompose a signal across scales and retain time localization.

The working idea was:

1. Decompose price and related variables into multiscale components.
2. Engineer stationary or normalized features.
3. Feed those features into recurrent or convolutional models.
4. Reconstruct future price behavior.
5. Produce uncertainty or confidence bands.
6. Evaluate the result out of sample.

Architectures Explored

The archive records experiments with:

• LSTM.
• GRU.
• CNN-LSTM.
• Attention.
• DeepAR.
• AutoGluon forecasting.
• PatchTST and transformer concepts.
• Wavelet-RNN hybrids.
• PCA plus regression.
• LightGBM and tree-based feature ranking.
• Bayesian and probabilistic outputs.
• Fourier and Hilbert-domain features.

The objective was not to declare one architecture superior.

It was to determine whether any architecture could generalize under causal walk-forward testing.

Feature Engineering

The feature sets included combinations of:

• Log returns.
• Higher-order differences.
• Rolling means and standard deviations.
• Technical indicators.
• Volume features.
• Fourier summaries.
• Hilbert instantaneous-frequency features.
• Wavelet approximation and detail coefficients.
• Reconstructed denoised series.
• Trend and curvature.
• Lagged forms of the above.

In several experiments, smaller selected feature sets outperformed broad collections of correlated indicators.

This was an early lesson that survived:

More features and more model capacity do not automatically create more information.

Why Wavelets Were Attractive

Wavelets offered several practical benefits.

Multiscale Representation

A market series contains movement at different time scales.

Wavelets separate slow background movement from faster local changes.

Time Localization

Fourier methods describe frequency content over a whole window.

Wavelets describe how scale content changes through time.

Flexible Denoising

Approximation and detail components can be reconstructed or filtered separately.

Event Morphology

Higher-order derivatives and wavelet energy can describe transition shapes that are difficult to represent with ordinary moving averages.

These strengths later contributed to the intraday bifurcation detector and the eigenstructure-instability research.

They did not rescue direct price forecasting.

Failure 1: Lag Direction

One of the most important implementation errors involved lagged features.

A lag should use earlier data:

feature_lag_1(t) = feature(t - 1)

In an early implementation, the shift direction was reversed for some features.

That allowed future values to enter rows treated as historical inputs.

The model appeared to predict well because part of the answer was already present.

This is ordinary lookahead leakage, but it can be visually subtle.

A plot may still look plausible. Training loss may still decline normally. The code may run without errors.

The failure is in the information set.

Failure 2: Prediction Alignment

Another error involved the alignment between walk-forward predictions and realized outcomes.

Predictions were offset by the forecast-window length when compared with the actual series.

This could make a forecast appear delayed, inverted, or unusually aligned depending on how the plot was constructed.

The model output and the target were not describing the same date.

This led to a permanent rule:

Every prediction row must carry an explicit issue date, target date or maturity definition, and validation join key.

Failure 3: Wavelet Order And Boundary Interpretation

Wavelet decomposition introduces choices that matter:

• Wavelet family.
• Decomposition level.
• Signal length.
• Padding.
• Boundary handling.
• Reconstruction order.
• Whether approximation or detail components are retained.
• Symmetry and phase behavior.

Asymmetric wavelets can shift features in time.

Boundary coefficients can be unstable.

A clean reconstruction does not prove a future-free representation.

The archive contains repeated corrections to wavelet-level interpretation and component order.

The lesson was not that wavelets are unusable.

The lesson was:

A multiscale transform is part of the causal model and must be audited like any other feature.

Failure 4: Static Train-Test Appearance

Several models produced impressive static plots.

Some had:

• Smooth forecast paths.
• Narrow intervals.
• High apparent R-squared.
• Visually correct turning points.

Walk-forward testing told a different story.

Performance degraded when:

• The model was retrained through time.
• Regimes changed.
• The evaluation moved away from the training distribution.
• Features were recalculated causally.
• Correct alignment was enforced.

A static train-test plot can hide:

• Regime dependence.
• Repeated reuse of one test period.
• Hyperparameter selection on the test data.
• Temporal leakage.
• Unstable scaling.
• Inappropriate uncertainty.

PCA Did Not Solve Generalization

PCA was tested as a way to reduce correlated features.

Some PCA-based models produced clean fits and narrow intervals.

The walk-forward behavior was not consistently better.

PCA preserves directions of feature variance, not necessarily directions that remain predictive through time.

The current production system does not use PCA in the public path.

Why Deeper Models Did Not Help

The research increased layer width, depth, context length, and model complexity.

The result was often:

• Better in-sample fit.
• More stable training curves.
• More convincing visual output.
• Little improvement in true out-of-sample behavior.

This is expected when the limiting factor is information rather than model capacity.

A larger network cannot infer a stable direction that is not encoded in the input state.

What The Forecasting Program Learned

Causality Must Be Verified Separately

A model can be dimensionally correct and statistically impressive while using the wrong information set.

Date Alignment Is Part Of The Model

The issue date, target window, and maturity date must be explicit.

Multiscale Features Are Useful, But Not Magical

Wavelets can expose structure. They do not create stable predictive information by themselves.

Direct Price Path Is A Difficult Target

Price direction and path combine many hidden variables.

The project obtained more stable results by estimating market state and uncertainty instead of one exact future path.

Smaller Models Can Be Stronger

Regularized linear models became preferable for the production layer because they are:

• Easier to audit.
• Less prone to hidden interpolation behavior.
• Easier to initialize and validate.
• More stable under limited data.
• Compatible with recursive coefficient updates.

What Survived Into Production

The current Vyreon system retained several ideas from the forecasting phase.

Causal Feature Processing

All feature transformations must use information available at the date being scored.

Endpoint Smoothing

The current pipeline uses causal endpoint transformations rather than centered filters that require future observations.

Exponential Normalization

Adaptive EMA-based normalization allows the feature state to evolve without a fixed full-sample mean.

Walk-Forward Thinking

The model is evaluated through time, not only on one static split.

Delayed-Label Discipline

Future horizon outcomes mature only after their full window is available.

Explicit Uncertainty

The production system publishes a distribution and validation information rather than only a point estimate.

Failure Preservation

Invalidated models remain documented so their errors are not repeated.

Why The Project Moved To State Estimation

Direct forecasting asked:

What exact price path will occur?

State estimation asks:

What market state is supported by the current measurements, how uncertain is it, and how does that state change when later evidence arrives?

The second question matched the information content of the data more closely.

The current system therefore estimates:

• Horizon-specific expected return state.
• Predictive uncertainty.
• Cross-horizon agreement.
• Innovation magnitude.
• Volatility regime.
• Model calibration.

This does not make the problem easy.

It makes the output more honest.

Walk-Forward Evaluation Design

A financial time-series model should be evaluated in chronological order.

A basic walk-forward process is:

1. Train on an initial historical window.
2. Predict the next unseen period.
3. Advance the training boundary.
4. Retrain or update using only newly available history.
5. Repeat.
6. Concatenate the untouched predictions.

This design is more expensive than one train-test split.

It reveals whether performance survives:

• Different volatility regimes.
• Changing trend.
• Market shocks.
• Data-distribution drift.
• Repeated retraining.

The 2024 research increasingly moved toward this design as static results failed.

Scaling And Normalization Leakage

A scaler fitted on the full dataset contains information about future distribution.

Even when no target value is included directly, the future mean and variance can alter historical feature values.

The correct process is:

• Fit the scaler on the training window.
• Apply it to the next test window.
• Refit only after the training boundary advances.

The current production artifacts freeze per-bucket scalers trained through a declared cutoff.

Live adaptive normalization occurs in the causal feature engine before the artifact scaler.

These are separate layers.

Probabilistic Output Problems

Several historical models attempted to output mean and uncertainty.

A narrow interval can look impressive.

It can also be wrong because:

• The variance target is misdefined.
• Residuals are not independent.
• The model is overfit.
• The interval is evaluated on reused data.
• The scale transformation is inverted incorrectly.
• The mean and uncertainty refer to different dates.

This experience influenced the current requirement that the expected range be checked against later realized outcomes.

Model Comparison Discipline

The research tested many architectures.

A fair comparison requires:

• Same target.
• Same causal features.
• Same training dates.
• Same test dates.
• Same metrics.
• Same retraining schedule.
• Same transaction or execution assumptions when relevant.

Without those controls, a more complex model can appear better because it received a different problem.

Why Feature Importance Was Not Enough

Tree-based feature importance helped identify recurring useful feature classes.

It did not establish that a feature was causal or stable.

Importance can be distorted by:

• Correlated features.
• One unusual period.
• High-cardinality variables.
• Leakage.
• Repeated tuning.
• Model-specific split behavior.

The current research uses feature importance as a question generator.

It is not treated as final evidence.

Archive Interpretation

The chronological notes preserve moments when a result was described as a breakthrough before later tests were complete.

Those passages are part of the evidence record.

They show why a claim looked convincing at the time.

The public case study should preserve the correction rather than silently rewriting the past.

That is especially important for AI-assisted research, where fluent interpretation can become more confident than the evidence.

The Role Of Negative Results

Several forecasting experiments did not produce a reusable model.

They still narrowed the design space.

The project learned that:

• A high-capacity sequence model was not automatically superior to a regularized baseline.
• Multiscale representation did not remove the need for causal alignment.
• Price-level targets were sensitive to non-stationarity.
• Narrow predictive bands could be false confidence.
• One successful era could dominate the result.
• Model explanations could become more confident than the test evidence.

These are technical results.

They explain why the current system emphasizes delayed targets, explicit baselines, and model-health charts.

What Would Justify Revisiting Direct Forecasting

Direct forecasting is not forbidden permanently.

It would require a clearly different evidence base.

A future attempt should begin with:

• A fixed economic target.
• A frozen feature timestamp convention.
• Multiple chronological folds.
• Strong simple baselines.
• Strict origin-time uncertainty scoring.
• Cross-asset or symbol-holdout tests.
• A documented reason the target contains information not already captured by the telemetry state.

Without those conditions, returning to a larger network would repeat the earlier research rather than advance it.

Current Conclusion

The Wavelet-RNN program did not produce a retained production forecasting model.

It produced something more durable:

• A documented example of lookahead leakage.
• A documented example of target misalignment.
• A multiscale feature-engineering foundation.
• A stronger walk-forward discipline.
• A reason to prefer causal state estimation over visually compelling direct prediction.

The core lesson is simple:

A sophisticated architecture cannot compensate for an invalid information set, an unstable target, or a weak validation design.