A causal graph-informed temporal convolution architecture for interpretable retail electricity price forecasting

• CG-TCN addresses the gap in unified retail price formation analysis by combining wholesale market conditions with contract-level attributes in a causal forecasting framework.Existing studies largely examine wholesale and retail drivers in isolation, while few studies analyze retail price formation and none evaluate their combined influence.Source: §1 · pp. 2–3
• Interpretable causal forecasting is motivated by the limits of black-box and correlation-based models, whose weak mechanism identification reduces robustness under regime shifts and policy shocks.Reliable forecasts support consumer protection, procurement planning, risk management, regulation, market design, and investment decisions.Source: §1 · pp. 2–3
• The model learns a directed causal graph and injects it through a graph-neural embedding that conditions the TCN, aligning representation learning with hypothesized causal pathways.The causal graph serves as an inductive bias for temporal representation learning and is intended to expose mechanisms that correlation-based or black-box forecasts do not establish.Source: §1 · pp. 3–4
• Economically motivated constraints restrict the retail price target from being a parent and preserve wholesale forward prices and time features as exogenous variables.These restrictions prevent reverse-causation feedback loops, encode basic economic structure, and filter spurious correlations from shared trends or temporal persistence.Source: §1 · p. 4
• The study uses causal, one-sided multi-resolution decomposition through STL and the deterministic Temporal Smoothing Network to extract horizon-specific price components in real time.Past-only components are fed into causal graph discovery to disentangle formation mechanisms across horizons and stabilize estimation under nonstationarity.Source: §1 · pp. 4–5
• CG-TCN unifies multi-resolution decomposition, causal discovery, and neural forecasting so temporal dynamics and causal pathways inform accurate, interpretable price predictions.The framework decomposes prices into interpretable components, learns directional dependencies, and uses a causal embedding to condition the TCN.Source: §1 · p. 5
• The one-sided STL-and-TSN decomposition extracts interpretable semiannual, quarterly, monthly, and short-term price components in real time.These components are intended to capture dominant temporal dynamics driving forecasting performance.Source: §1 · p. 5
• A causal discovery stage identifies directional relationships among contract attributes, market fundamentals, and decomposed price components, using those links to enhance prediction.NOTEARS discovers the graph, which is encoded as an adjacency embedding that guides the TCN.Source: §1 · p. 5
• The proposed CG-TCN injects a learned causal DAG into a TCN through a graph-neural embedding, aligning convolutional representations with causal pathways.The approach combines neural-network predictive strength with causal-inference interpretability for retail electricity price forecasting.Source: §1 · p. 5
• The framework adds spike-aware training that emphasizes spike windows and penalizes errors there more heavily to capture abrupt price surges.It is demonstrated on every daily retail choice offer filed by every supplier in every Ohio service territory over a decade, with state-of-the-art accuracy and interpretable attributions.Source: §1 · p. 5

2. The proposed CG-TCN forecasting framework

• CG-TCN jointly encodes cross-variable and temporal causality through a domain-constrained directed graph whose parent-linked neural features condition causal convolutions and channel-wise gating.The graph excludes target parents and preserves exogeneity, while graph-neural features are fused with dilated temporal features to form a unified predictive representation.Source: §2 · p. 7
• The framework decomposes retail prices into semiannual, quarterly, monthly, and high-frequency components that enter both causal graph discovery and forecasting.Dilated causal convolutions capture long- and short-range dependencies while preserving chronological order, and spike-aware training emphasizes rare stress periods.Source: §2 · p. 7
• The model produces interpretability outputs spanning the learned causal graph, effective information-routing adjacency, and temporal saliency by target component and lag.Together, these outputs identify drivers of long-horizon trends versus short-term fluctuations and indicate when those influences matter for procurement, hedging, and policy evaluation.Source: §2 · p. 8

2.1. Causal multi-scale time series decomposition

• The decomposition stage provides causal, horizon-specific inputs by applying either trailing-window STL or causal Butterworth low-pass filters to daily retail prices.Both alternatives produce semiannual, quarterly, monthly, and high-frequency components while using only current and past observations for real-time implementation.Source: §2.1 · pp. 8–9, §2.1.2 · p. 11

2.1.1. Causal STL-based multi-scale decomposition

• Causal STL estimates low-frequency trends from trailing observations at nested semiannual, quarterly, and monthly horizons, then forms band-limited components by differencing.Evaluating STL at sparse anchor times and interpolating between them reduces computational burden while retaining one-sided, real-time trend estimates.Source: §2.1.1 · p. 9

2.1.2. Causal TSN decomposition

• Causal TSN uses a bank of forward Butterworth low-pass filters whose recursions depend only on current and past observations, producing strictly one-sided components.Successive low-pass outputs are differenced at the same nested horizons, yielding a linear, time-invariant and explicitly frequency-domain alternative to STL.Source: §2.1.2 · pp. 10–11
• The decomposition experiments compare one-step-ahead forecasting across window-length settings, after which the best-performing horizon settings are used in subsequent experiments.The selected windows affect component smoothness, phase lag, and predictive usefulness.Source: §2.1.2 · p. 11

2.2. Causal graph discovery

• The causal graph discovery stage learns sparse, directed dependencies among decomposed price components and covariates, using NOTEARS with economic constraints to support interpretable causal pathways.Forecasting targets cannot be parents, exogenous inputs remain sources, and prohibited or redundant relationships are masked during optimization.Source: §2.2 · pp. 11–12

2.3. Graph neural network with causal embedding

• The causal GNN propagates information along the learned adjacency and produces a direction-aware embedding that conditions temporal convolutions, coupling cross-variable structure with sequential learning.The embedding respects domain-relevant directionality and helps reveal how wholesale prices and retail contract information propagate through price formation.Source: §2.3 · pp. 13–14

2.4. Temporal convolutional network and feature fusion

• The TCN captures sequential dependencies with dilated causal convolutions and exponentially increasing receptive fields, ensuring predictions use only information available before time t.Residual convolutional blocks span the longest semiannual decomposition horizon while maintaining sample efficiency.Source: §2.4 · p. 14
• The model concatenates TCN temporal features with GNN causal embeddings, forecasts four decomposed components, and reconstructs overall prices by summing them across horizons.The causal embedding captures dependencies among decomposed price components and exogenous drivers, enabling attributions such as wholesale signals dominating long horizons and contract attributes shaping high-frequency behavior.Source: §2.4 · p. 15

2.5. Weighed loss function

• The weighted loss combines component and sample weighting with Smooth L1 discrepancy to emphasize spike windows, improve robustness, and stabilize training under volatile prices.Component weights support interpretability, while batch-wise normalization preserves scale-invariance and higher spike-window weights address rare but consequential errors.Source: §2.5 · pp. 16–17

Data construction

• The database covers the full universe of residential retail offers filed with PUCO from January 2014 through December 2023, retaining real-time consumer prices and contractual terms.Unlike aggregated utility-level or EIA datasets, the contract-level database captures detailed retail price formation and forecasting inputs.Source: §3 · p. 17

Study setting and outcomes

• Using Duke Energy Ohio data, the analysis constructs daily median and minimum price series for 12-month fixed residential contracts from January 21, 2014, through December 31, 2023.Source: §3 · p. 18

Evaluation design

• Both price series are independently split into 85% training and 15% testing subsets, with time-series cross-validation used for model selection within training.The median series is the primary benchmark, while the minimum series assesses robustness under more volatile pricing conditions.Source: §3 · p. 18

Covariates

• Covariates combine contract attributes, PJM West Hub 12-month wholesale forwards aligned to origination, and daily calendar indicators for weekly, seasonal, and holiday effects.Contract features include fixed monthly and early termination fees; market conditions also include default utility service rates.Source: §3 · p. 19

4.1.1. Evaluation metrics

• Forecasting is evaluated with MAE, MSE, RMSE, MAPE, NRMSE, and R2, where lower error values and higher R2 indicate better performance.MAPE is unavailable for minimum-price series containing zeros because it becomes undefined or unstable.Source: §4.1.1 · pp. 20–21, §4.3 · p. 25

4.1.2. Baseline methods

• The comparison includes RF, XGB, AR, LSTM, TCN, MTGNN, Autoformer, and CTCN, with TCN, MTGNN, and CTCN configured without the proposed pre-learned domain-constrained DAG.These baselines respectively represent tree ensembles, autoregression, recurrent and convolutional networks, graph-temporal learning, decomposition-based Transformers, and causal temporal convolution.Source: §4.1.2 · pp. 21–22

4.2. Time series decomposition results

• The decomposition includes quarterly and monthly components, while its fluctuation component captures sharp, short-lived price drops and spikes that are more pronounced for minimum prices.Source: §4.2 · p. 22

4.3. Forecast accuracy and stability

• CG-TCN achieves strong median-price performance with MAE 0.2685, MSE 0.1302, RMSE 0.3609, NRMSE 6.49%, and R2 = 0.9648, outperforming deep-learning and conventional baselines overall.Autoformer has slightly lower MAE and MAPE but higher RMSE and lower stability; RF, XGB, AR, LSTM, TCN, and CTCN are less accurate overall.Source: Table 3 · p. 24, §4.3 · p. 24
• CG-TCN also achieves the best minimum-price performance, with MAE 0.2772, MSE 0.1534, RMSE 0.3917, NRMSE 4.04%, and R2 = 0.9693.MAPE is not reported because zero values make it undefined or unstable; competing deep-learning and traditional models perform worse, especially on RMSE and explained variance.Source: Table 4 · p. 25, §4.3 · p. 25
• CG-TCN represents price behavior across multiple temporal scales, tracking long-run trends, medium-term adjustments, short-horizon oscillations, and high-frequency variability with modest smoothing of extreme spikes.The model captures the long-run downward trend, aligns monthly oscillations with observations, and reproduces high-frequency variability while slightly attenuating extreme spikes.Source: §4.3 · pp. 25–26

4.4. Multi-step-ahead forecasting results

• CG-TCN consistently outperforms benchmark models across 5-, 10-, and 15-day horizons, retaining the lowest errors and highest R2 as forecast distance increases.At 5 days, CG-TCN achieves MAE 0.404 and R2 0.935 versus Autoformer’s MAE 0.546 and R2 0.840; at 10 days, it achieves RMSE 0.499 and R2 = 0.934.Source: §4.4 · pp. 26, 28, Table 5 · p. 28

4.5.1. Causal DAG

• The constrained causal graph is sparser and more interpretable than the unconstrained graph, aligning price-component dependencies with exogenous market inputs and theoretical causal ordering.The unconstrained graph produces dense, feedback-like relationships, whereas constraints prevent forecasting targets from being parents and keep time features and wholesale forward prices from being causal children.Source: §4.5.1 · p. 29
• Nonlinear NOTEARS learns a more selective graph emphasizing strong links among wholesale forward prices, contract fees, and price trends, while both variants achieve similar predictive accuracy.The nonlinear graph marginally improves R2 and RMSE/MSE, whereas the linear graph achieves slightly lower MAE and MAPE, reflecting a flexibility–data-availability trade-off.Source: §4.5.1 · p. 31
• The GNN embedding extends sparse NOTEARS graphs by aggregating higher-order neighbors, preserving interpretability while capturing distributed nonlinear interactions for temporal forecasting.NOTEARS retains statistically dominant direct edges but may omit weaker or distributed pathways.Source: §4.5.1 · pp. 32–33
• Wholesale forward prices strongly influence long-horizon trends, whereas monthly contract fees primarily drive high-frequency fluctuations and contract attributes modulate short-term dynamics.The GNN adjacency matrix also identifies the SSO as an important driver, while calendar and holiday effects are relatively weak.Source: §4.5.1 · pp. 33–34

4.5.2. Temporal causality

• Recent observations at lags T–1 to T–3 contain the strongest predictive influence across targets, with saliency rapidly attenuating at longer horizons.Some features retain moderate longer-horizon influence, reflecting persistent structural effects.Source: §4.5.2 · p. 38
• Temporal saliency varies by component: wholesale forward prices and Trend 180 dominate semiannual trends, while monthly fees and IMF sum influence fluctuations.Quarterly and monthly trends show more diffuse dependence on contract attributes, seasonal indicators, and intermediate-term price dynamics.Source: §4.5.2 · p. 38
• Temporal causality analysis shows that retail electricity price dynamics reflect both persistent causal drivers and time-localized responses.Static causal structure, learned graph embeddings, and temporal saliency jointly reveal when causal relationships exert the greatest forecast influence, enhancing interpretability and credibility.Source: §4.5.2 · p. 38

4.5.3. Causality Examination in Autoformer

• Under Autoformer’s within-week calendar intervention, MAPE increases 31.5%, MSE increases 68.9%, and R2 declines from 0.961 to 0.934, indicating substantial predictive-accuracy degradation.The intervention preserves economic covariates, seasonal patterns, and long-run temporal structure; the losses indicate that a nontrivial portion of predictive accuracy is driven by calendar-based effects.Source: Table 7 · p. 39, §4.5.3 · p. 40

4.6. Ablation study

• The ablation study finds that each CG-TCN module contributes distinct, complementary benefits, and removing any single component measurably degrades forecasting accuracy or generalization.Removing the causal graph increases test MAE to 0.5594 and decreases R2 to 0.8391; removing the temporal convolution module causes a more pronounced deterioration.Source: §4.6 · p. 41

4.7. Sensitivity analysis

• The sensitivity analysis is motivated by the need to understand how architectural choices affect forecasting performance during deployment.No sensitivity-analysis findings are supplied in this section part.Source: §4.7 · p. 42
• MAE = 0.5563 and R2 = 0.8447 for the 10step-2layerTCN show that short input windows substantially worsen test performance.Extending the window to 25 or 40 steps improves forecasts by capturing longer-term temporal dependencies alongside short-term fluctuations.Source: §4.7 · p. 42
• Increasing network depth generally improves performance when sufficient historical context is available, but complexity gains are not monotonic because deeper models risk overfitting.The 25step-2layerTCN reaches R2 = 0.9485 with substantially lower architectural complexity than deeper alternatives, indicating a favorable accuracy–efficiency trade-off.Source: §4.7 · pp. 42–43
• The 40step-2layerTCN achieves the best overall test performance, with the lowest MAE (0.3007), RMSE (0.4183), and MAPE (3.22%), and the highest R2.Table 9 reports sensitivity analysis of CG-TCN base models under different input window lengths and TCN depths, without hyperparameter optimization.Source: §4.7 · pp. 42–43, Table 9 · p. 43
• Longer input sequences provide richer historical context, while appropriate depth enhances temporal feature extraction, motivating problem-specific calibration in the full CG-TCN.The results associate the architecture’s effectiveness with learning multi-scale temporal representations suited to retail electricity price volatility and structural persistence.Source: §4.7 · pp. 42–43

• Integrating causal graphs with temporal convolution improves predictive performance and interpretability by distinguishing structural price movements from short-term noise, especially in volatile retail markets.The discussion contrasts this approach with correlation-driven forecasting and emphasizes improved generalization in markets characterized by price variability or volatility.Source: §5 · pp. 43–44
• Multi-scale decomposition is critical because longer windows and deeper temporal representations capture forward-price and contract-term trends, while spike-aware mechanisms respond to rare, impactful fluctuations.These findings support aligning model design with the heterogeneous temporal dynamics of retail electricity price formation.Source: §5 · p. 44
• Learned causal graphs indicate that wholesale forward prices primarily shape long-horizon trends, whereas monthly fees and early termination fees more strongly influence short-term fluctuations and discrete price adjustments.Fee-based features may let suppliers manage costs and risk without uniformly increasing headline per-kWh prices, creating implications for regulators and consumer advocates.Source: §5 · p. 44
• The framework can help anticipate elevated price risk and assess how default service rates or contract fee structures affect retail price magnitude and volatility.Its potential applications include consumer protection, market monitoring, competition enforcement, welfare analysis, affordability assessment, and private-sector market operations.Source: §5 · p. 44
• Using a decade of daily retail offers from every supplier in one utility territory provides richer behavioral and product-heterogeneity information than mechanistic wholesale data and better reflects prices paid by households and businesses.The study focuses on depth rather than breadth in a single territory, while the methodology is presented as generalizable to other deregulated and retail-choice markets.Source: §5 · pp. 44–45
• Future extensions include probabilistic forecasting for full price distributions, uncertainty and tail-risk quantification, and application across multiple territories and contract types for cross-market comparison.These extensions could support risk-aware procurement, budget planning, consumer protection, and identification of high-price or risky offers, including during extreme market conditions.Source: §5 · p. 45

• Learned causal structures provide economically meaningful insights into interactions between wholesale markets and retail contract design, improving transparency in behaviorally complex markets.This addresses the need to move beyond short-term pattern recognition toward structural explanations of price formation.Source: §6 · p. 46
• CG-TCN integrates multi-resolution signal decomposition, causal graph discovery, and hybrid temporal deep learning to model structural mechanisms underlying retail electricity price formation.The framework disentangles semiannual, quarterly, monthly, and high-frequency dynamics while modeling causal links with wholesale forward prices, contract attributes, and calendar effects.Source: §6 · p. 46
• CG-TCN achieves superior forecasting accuracy relative to transformer-based and statistical benchmarks while revealing how market drivers propagate through distinct temporal layers.The conclusion attributes these gains to explicit modeling of decomposed price dynamics and causal links.Source: §6 · p. 46
• The research delivers predictive and transparent models that are relevant to regulators, policymakers, and market participants concerned with price stability, affordability, and consumer protection.Understanding how wholesale and retail characteristics contribute to price formation informs economic and regulatory questions in open retail electricity markets.Source: §6 · pp. 46–47

• The potential factors are organized into Tables A.1, A.2, and A.3 covering contract, energy market, and time-feature groups.Source: Appendix A · p. 53, Table A.1 · p. 53, Table A.2 · p. 53
• Time features include contract publish day, with values from day 0 to day 6 of each week.Source: Appendix A · p. 53
• Time features include contract publish month, spanning months 1 through 12 from January to December.Source: Appendix A · p. 53
• Time features include whether the contract publish date is a US holiday.Source: Appendix A · p. 53

Construction

• The forward wholesale price aligns wholesale-contract maturities with each retail offer’s duration and averages monthly prices over the complete delivery period.For a contract originating on day t with n months of length, calendar-month cycling and delivery-year adjustment handle horizons spanning multiple years.Source: Appendix A.1 · p. 53

Interpretation and data source

• The resulting measure approximates the wholesale procurement cost a supplier would incur if hedging the retail contract at origination.Wholesale forward prices come from S&P Market Intelligence for PJM’s West Hub and are expressed in cents per kilowatt-hour.Source: Appendix A.1 · p. 54

Evaluation rationale

• Traditional random k-fold cross-validation can leak future information, violating the requirement that only past observations determine future outcomes.This motivates using TSCV for all experiments.Source: Appendix B · p. 55

TSCV procedure

• Each TSCV fold trains on observations through t_i and validates on the subsequent interval (t_i, t_i+1], with rolling or expanding windows preserving chronological order.The training window advances sequentially so forecasts use historical data only.Source: Appendix B · p. 55

Robustness assessment

• TSCV assesses robustness across multiple historical windows, testing generalization under changing market regimes and volatility conditions.This is important for electricity price forecasting, where market conditions can change over time.Source: Appendix B · p. 55

TSCV procedure

• Figure B.1 illustrates the procedure: each fold trains on historical observations and validates on the next temporal segment while rolling forward.Dark blue denotes training observations and light blue denotes the validation segment.Source: Figure B.1 · p. 55, Appendix B · p. 55

Evaluation rationale

• Time-series cross-validation provides a more realistic and causally consistent evaluation than random-split cross-validation for retail electricity pricing.The approach is particularly suited to nonstationary domains with regime shifts.Source: Appendix B · pp. 55–56

Spike Detection and Training Balance

• Spikes are defined as large positive or negative deviations from a smoothed baseline exceeding a specified z-score threshold, capturing short-lived episodes of market stress and heightened volatility.Detection is applied to the final reconstructed retail price series rather than individual decomposed components, so extreme movements are treated at the consumer-facing price level.Source: Appendix C · p. 57

Spike-Aware Training Mechanism

• The spike-aware sampler oversamples spike windows with replacement so a fixed proportion p_spike of each mini-batch contains spike observations, while higher loss weights amplify their training influence.For p_spike = 0.5, approximately half of each training batch is expected to contain spike windows, and the weighting improves responsiveness to abrupt price surges.Source: Appendix C · p. 57

Spike Detection and Training Balance

• 667 of 3,583 windows (18.62%) are identified as spike windows, enabling balanced training under extreme volatility without artificially altering the underlying data distribution.The quantified spike frequency also verifies that detection is neither overly aggressive nor overly conservative.Source: Appendix C · pp. 57–58

Selection for forecasting

• The best-performing median- and minimum-price decomposition settings are adopted in subsequent forecasting experiments.Configurations are evaluated under one-sided STL and TSN decompositions using multiple long-, medium-, and short-horizon windows.Source: Appendix D · p. 59

Median prices

• For median prices, STL (180, 60, 20) achieves the lowest MAE (0.2685), lowest RMSE (0.3609), and highest R2 (0.9648), while TSN (180, 60, 10) achieves R2 = 0.9290.The comparison uses identical architectures and training procedures for one-step-ahead out-of-sample forecasts; Table D.1 reports the configurations.Source: Appendix D · p. 59

Interpretation

• The results indicate that moderate medium-horizon smoothing combined with relatively short short-horizon components best balances stability and responsiveness.Shorter short-horizon windows help capture abrupt minimum-price movements, although the optimal medium-horizon smoothing differs by target.Source: Appendix D · p. 59

Minimum prices

• For minimum prices, STL performs best with (180, 90, 15), whereas TSN (180, 60, 10) achieves the lowest MAE (0.2772) and highest R2 (0.9693).The longer STL medium-term horizon helps stabilize extreme lower-bound movements, while TSN retains the shorter short-horizon setting.Source: Appendix D · p. 59, Table D.2 · p. 60

• Hyperparameter optimization was applied to the proposed CG-TCN model and the deep-learning baselines MTGNN and Autoformer.Source: Appendix E · p. 61
• The procedure was designed to ensure a fair and systematic comparison across models.Source: Appendix E · p. 61
• Each model was tuned within its standard architectural design while avoiding ad hoc manual calibration.Source: Appendix E · p. 61

Appendix E.1. Optuna-based optimization framework

• Optuna searches CG-TCN, MTGNN, and Autoformer parameter spaces using model-specific architectural and training configurations.The define-by-run framework supports flexible search spaces spanning network depth, channel dimensions, learning rate, dropout, and weight decay.Source: Appendix E.1 · p. 61
• The TPE algorithm [46] adaptively balances exploration and exploitation when sampling candidate hyperparameter configurations.Optuna pruning terminates unpromising trials early using intermediate validation performance, reducing training cost and mitigating overfitting to specific hyperparameter choices.Source: Appendix E.1 · p. 61
• Each model minimizes held-out validation loss under time-series cross-validation, then retrains the selected configuration on the full training set for test evaluation.The resulting optimal hyperparameters are reported in Tables E.1–E.3.Source: Appendix E.1 · p. 61
• CG-TCN jointly optimizes general architecture, training, and spike-aware parameters, including detection threshold, spike loss weight, and spike-window sampling probability.The search includes TCN, GNN, and fusion dropout rates; kernel size; TCN depth; channel widths; and GNN output dimensions.Source: Appendix E.1 · p. 61
• MTGNN and Autoformer use separate search spaces covering graph-temporal and transformer-specific architectures alongside standard training parameters.MTGNN optimization includes graph convolution depth, subgraph size, channel dimensions, and dilation factors; Autoformer optimization includes attention heads, encoder-decoder depths, feed-forward width, and moving-average window length.Source: Appendix E.1 · pp. 61, 64

Appendix E.2. The other baseline models

• LSTM, TCN, and CTCN are not automatically optimized; instead, they use commonly adopted literature configurations as reference benchmarks rather than fully optimized competitors.All baseline models use Adam [47], with weight decay for learning-rate regularization and default settings otherwise.Source: Appendix E.2 · p. 64

• Figure F.1 shows that the baseline CTCN places dominant emphasis on the long-horizon trend component, Trend 180, particularly at distant lags, in its saliency-based attribution.The results indicate emphasis on long-term trend components while underutilizing exogenous or calendar-based features.Source: Figure F.1 · p. 65, Appendix F · p. 65
• CTCN saliency captures predictive influence but cannot distinguish true causal drivers from features merely correlated with retail prices.Its flat attribution structure treats covariates as direct contributors, risking correlation–causation conflation when inputs share temporal structure.Source: Appendix F · p. 65
• CG–TCN embeds an explicit causal graph that routes exogenous variables through decomposed price trends before influencing the final retail price.This hierarchical routing redistributes attribution across interpretable causal pathways rather than concentrating weight on a small set of dominant predictors.Source: Appendix F · p. 66
• Figure F.2 contrasts CTCN’s independent-predictor structure with CG–TCN’s layered hierarchy, which yields a more interpretable representation grounded in structured causal reasoning.In CG–TCN, exogenous features influence intermediate price trends through learned causal links, which jointly determine the final retail price.Source: Figure F.2 · p. 66, Appendix F · p. 66