Source-linked AI summary
A convolutional framework for detecting event-driven dynamics in energy price series
Caixia Xu, Piotr Fryzlewicz
TL;DR
The paper addresses heterogeneous abnormal dynamics in price series, where existing methods may not perform equally across distinct changes. It develops a common 1D CNN framework and finds that it can retain classifier performance while distinguishing geopolitical and weather-associated energy-market events.
Problem
Price series can exhibit distinct forms of abnormal change, while existing methods may not perform equally across explosiveness, level shifts, variance breaks, and slope changes.
Method
The paper develops a common 1D convolutional framework for detecting heterogeneous event-driven dynamics in univariate time-series windows.
Results
The learned common head retains classifier performance in simulations and, in an independent 2026 case study, distinguishes weather-associated natural-gas dynamics from predominantly geopolitical oil and refined-product patterns.
Takeaways & Limitations
The framework supports classification of heterogeneous energy-price event patterns across several markets and event families.
Takeaways & Limitations
The outputs are classifications rather than causal estimates, and the binary results evaluate complete prediction procedures rather than architecture alone.
Abstract
from arXiv · showhide
This paper develops a general convolutional neural network (CNN) framework for detecting heterogeneous event-driven dynamics in univariate time series windows. We show that the induced CNN class exactly represents classifiers based on range, maximum drawup, maximum drawdown and slope change, and uniformly approximates realised volatility and autoregressive explosiveness on compact domains. We further establish error bounds for representative rules in finite samples and an oracle inequality for learning across them. Simulations show that the proposed model can match or outperform classifiers based on individual statistics as the training sample grows. In an application to six daily energy price series, a hierarchical CNN distinguishes event windows and event families. Applied without retraining to observations withheld after 20 February 2026, the fitted model identifies predominantly geopolitical dynamics in several oil and refined product series around the outbreak of the 2026 Iran war, while distinguishing a contemporaneous natural gas spike associated with weather.
1. Introduction
Existing detectors target particular forms of abnormal price dynamics, while a single price series may exhibit several distinct changes. The paper therefore develops a common CNN framework intended to accommodate multiple statistical detectors and event families.
- Motivation: Abnormal price dynamics can appear as unusually large ranges, rapid movements, volatility bursts, trend changes, or persistent departures from an earlier regime.Related approaches include range, drawup and drawdown statistics, realised volatility, autoregressive explosiveness, and change-point procedures.
- Motivation: Existing change-point frameworks can accommodate several departures, but their application remains conditional on a previously specified statistical characterisation of the change of interest.Neural approaches likewise remain tied to particular stochastic definitions, discrepancy measures, or sequential testing constructions.
- Motivation: A detector designed for autoregressive explosiveness, level shifts, variance breaks, or slope changes may not perform equally well for other forms of abnormal behaviour.This motivates seeking a method that can represent or approximate several detectors within one architecture.
- Contributions: The paper develops a common convolutional framework whose induced CNN class contains classifiers based on range, maximum drawup, maximum drawdown, and slope change.It also uniformly approximates classifiers based on realised volatility and autoregressive explosiveness on compact domains.
- Contributions: The framework is complemented by finite-sample bounds for representative rules and an oracle inequality comparing learning in the common class with fixed statistical comparators.These results connect the shared architecture to both individual detectors and learning guarantees.
- Contributions: A hierarchical implementation first detects event windows and then distinguishes weather, geopolitical, and supply-financial events across six energy price series.An independent holdout study applies the fitted hierarchy without retraining to the onset of the 2026 Iran war.
2. Methodology
The methodology formulates event detection as binary classification of energy-price windows and develops interpretable classical-statistic rules alongside a flexible 1D CNN architecture. The CNN uses parallel temporal branches, pooling, and local-to-global aggregation to represent or approximate heterogeneous event patterns.
- Event-window classification: Event detection assigns positive labels to windows overlapping events and negative labels to non-event windows.
- Classical statistics: The range measures overall price amplitude without regard to movement direction or timing.
- Classical statistics: The classical statistics capture overall movement amplitude, directional cumulative changes, volatility bursts, trend changes, and autoregressive explosiveness.
- Classical statistics: Maximum drawup and drawdown capture strongest cumulative upward and downward movements, respectively, helping distinguish upward pressure from downward market stress.
- 1D CNN event classifier: The CNN uses parallel convolutional branches with different temporal spans, enabling examination of price movements at multiple scales.
- 1D CNN event classifier: Convolutional responses are transformed into local scores, globally pooled, and aggregated into window-level features that are less sensitive to response position.
3. Representation, approximation and risk bounds
The CNN class exactly represents several classical event-statistic classifiers and uniformly approximates realised volatility and autoregressive statistics on compact domains. The paper also derives finite-sample risk controls and an oracle inequality for learning across these rules.
- Representation and approximation: The CNN class is rich enough to contain or uniformly approximate classifiers based on the paper’s classical statistics.The results cover exact representation for range, directional, and slope-change rules, plus uniform approximation for realised volatility and autoregressive statistics.
- Exact representations: The range classifier is represented exactly, with a CNN score equal to R(X) − λ for every X ∈ R^T.The construction holds on all of R^T and requires no input bound.
- Exact representations: The slope-change classifier is represented exactly by branches spanning the full input window for each candidate split point and sign.The affine head combines the resulting nonnegative branch scores.
- Representation and approximation: Outside an ϵ-neighbourhood of λ, the approximating CNN and realised-volatility rule make the same decision on X_M.The realised-volatility construction uses a finite ReLU square approximant on the clipped compact domain.
- Representation and approximation: The autoregressive construction gives analogous score and decision guarantees by uniformly approximating local products on X_M.Clipping and the approximation construction support agreement with the realised-volatility and autoregressive rules on the relevant domain.
- Risk and generalization bounds: The finite-sample theory controls false positives and gives exponentially decreasing false negatives under separated slope-change signals, while an oracle bound compares learned CNN risk with three fixed comparators.The oracle inequality adds a complexity penalty to the comparator risk.
4. Simulation
The simulation compares a joint CNN with fixed statistical comparators and a single-comparator selector across increasing training samples. The joint affine head improves on fixed comparators in slope, volatility, and mixed scenarios, with its advantage increasing with sample size.
- Design: The experiment compares the joint CNN with the best fixed statistical rule and studies how the comparison changes with training sample size.The joint classifier combines slope-change, realised-volatility, and autoregressive evidence through a learned affine head.
- Results: The joint affine head is slightly worse than the oracle at N = 200, approximately matches it at N = 500, and improves on it at N = 1000.The absolute training–test error gap decreases with N while AUC remains high.
- Results: The joint affine head has lower test error than the oracle comparator in the slope, volatility, and mixed scenarios.Its paired test-error difference is below zero for every training size in those scenarios, with the advantage increasing with N.
- Results: The single-comparator ERM selector remains close to the zero-oracle line, while in volatility and mixed scenarios it selects the same fixed rule as the test-set oracle.Realised volatility is the oracle comparator in the mixed scenario.
5. Empirical results
The empirical analysis applies a hierarchical CNN to six energy price series, first detecting event windows and then classifying weather, geopolitical, and supply-financial events. The model distinguishes explosive event-driven patterns from stable regimes, performs strongly in binary classification, but faces its main multiclass difficulty separating geopolitical from supply-financial events.
- Data and labelling: The study uses six daily energy commodity spot-price series and manually curated event records with categories, dates, sources, and shock descriptions.The series include crude oil, natural gas, and refined petroleum products.
- Hierarchical CNN: The hierarchical procedure detects event windows first, then routes detected events through classifiers separating weather, geopolitical, and supply-financial families.The final output contains four classes: no event and the three event families.
- Binary classification: Henry Hub natural gas achieves the highest accuracy, F1 and AUC, while Brent crude oil has the lowest accuracy and AUC; diesel has near-complete recall but lower precision.Diesel’s main errors are false event detections.
- Binary classification: The CNN distinguishes explosive event-driven price patterns from relatively stable non-event regimes.Representative errors suggest that local fluctuations can resemble labelled events, while complex multi-spike events may reduce classification confidence; one false negative receives event probability 0.406.
- Binary classification: The proposed CNN attains the highest accuracy, precision, recall, F1 and AUC among the compared prediction procedures.The comparison covers complete prediction procedures rather than isolating architecture alone, because loss and optimisation choices remain method specific.
- Multiclass classification: The multiclass model identifies no-event windows most reliably, while weather and geopolitical events remain distinguishable and supply-financial events are often reassigned.The reported multiclass accuracy is 0.7639 and macro F1 is 0.6067.
- Holdout application: Geopolitical classifications dominate Brent crude oil and two Gulf Coast products after the marked date, whereas natural gas windows around the weather-related spike are assigned mainly to the weather family.For natural gas, classifications shift mainly to no event once the spike leaves the 80-observation input window.
6. Conclusion
The paper develops a common CNN framework for heterogeneous event-driven dynamics, with theoretical guarantees, simulation evidence, and an energy-price application. The holdout study classified predominantly geopolitical patterns in oil and refined products while distinguishing a weather-associated natural-gas spike; these outputs are classifications rather than causal estimates.
- The induced CNN class exactly represents classifiers based on range, maximum drawup, maximum drawdown, and slope change.
- It uniformly approximates classifiers based on realised volatility and autoregressive explosiveness on compact domains.
- Finite-sample results connect these constructions to error control for individual rules and an oracle comparison over representative statistical classifiers.
- Simulations show that the learned common head can retain a strong fixed rule's performance and exploit complementary information across branches.
- The hierarchical model classified events in six energy price series, while the independent 2026 case study distinguished a weather-associated natural-gas spike from predominantly geopolitical patterns in oil and refined products.
- These outputs are classifications rather than causal estimates, and extensions to forecasting, multivariate series, network dependence, and uncertainty quantification remain future work.
Data availability statement
The energy price series are publicly available from the Federal Reserve Economic Data database, while curated event data and supporting code are available from the corresponding author upon reasonable request.
- The energy price series used in the study are publicly available from the Federal Reserve Economic Data database.
- Curated event data and code supporting the findings are available from the corresponding author upon reasonable request.