Journal of Robotics and Control (JRC) Volume 6.
Issue 5, 2025 ISSN: 2715-5072.
DOI: 10.
18196/jrc.
An Explainable CNNAeLSTM Framework for Monthly Crude Oil Price Forecasting Using WTI Time Series Data Joompol Thongjamroon 1.
Songgrod Phimphisan 2.
Nattavut Sriwiboon 3* Department of Business Computer.
Faculty of Administrative Science.
Kalasin University.
Kalasin.
Thailand
2, 3
Department of Computer Science and Information Technology.
Faculty of Science and Health Technology.
Kalasin University.
Kalasin.
Thailand Email: 1 joompol.
th@ksu.
th, 2 songgrod.
ph@ksu.
th, 3 nattavut.
sr@ksu.
*Corresponding Author AbstractAiCrude oil price forecasting has posed significant challenges due to its volatility and nonlinear dynamics.
This study has proposed an explainable CNNAeLSTM framework to predict monthly West Texas Intermediate (WTI) crude oil The model has captured both local and sequential patterns without using external inputs or decomposition.
Trained over 50 epochs across three data splits, it has been evaluated using RMSE.
MAE.
MASE.
SMAPE, and directional A classification accuracy of 92.
4% and directional accuracy of up to 87.
4% have been achieved.
The model has consistently outperformed classical and hybrid baselines, with statistical significance confirmed by the FriedmanAeNemenyi Saliency-based interpretability has further enhanced transparency, making the framework suitable for real-world energy forecasting.
KeywordsAiCrude Oil Price Forecasting.
CNNAeLSTM Hybrid Model.
Time Series Prediction.
WTI.
Deep Learning.
INTRODUCTION
Crude oil has played a pivotal role in shaping global economic stability, energy policy, and financial markets.
Among various petroleum benchmarks.
West Texas Intermediate (WTI) .
, .
crude oil has been widely recognized as a standard reference in international oil pricing.
However, forecasting crude oil prices has remained a complex task due to the influence of geopolitical events, supplyAedemand imbalances, macroeconomic fluctuations, and nonlinear market behavior.
Accurate and timely forecasting models are therefore essential for risk management, investment strategies, and policy-making within the energy sector.
Traditional statistical models such as autoregressive integrated moving average (ARIMA) .
, generalized autoregressive conditional heteroskedasticity (GARCH) .
, and exponential smoothing have been extensively used for oil price forecasting.
While these models have offered interpretability and ease of implementation, their performance has been limited by strong linearity assumptions and weak adaptability to non-stationary patterns in crude oil time series.
To overcome these limitations, machine learning (ML) .
models, including support vector regression (SVR), decision trees, and ensemble methods, have been introduced to handle nonlinearity .
Despite improved performance, most ML models have lacked the ability to retain long-term temporal dependencies critical in time series Recent advances in deep learning (DL) .
have introduced powerful neural network architectures capable of learning complex .
representations from raw sequences .
Models such as convolutional neural networks (CNN.
, long short-term memory (LSTM) .
networks, and attention-based transformers have demonstrated substantial progress .
, .
in financial forecasting, energy demand modeling, and economic prediction .
Hybrid DL models, in particular, have gained attention for combining complementary architectures such as CNN .
, .
for local pattern recognition and LSTM for sequence modeling.
Nevertheless, many existing studies have depended on signal decomposition techniques or external variables, which may increase computational cost and reduce generalizability.
To address these .
gaps, an explainable CNNAe LSTM framework has been proposed in this study for monthly WTI crude oil price forecasting.
The model has been designed to operate end-to-end without requiring decomposition or external data, while capturing both shortand long-term dynamics in the input series.
Saliency-based gradient analysis has been integrated to enhance interpretability, allowing users to understand which historical points have influenced the modelAos forecasts.
comprehensive evaluation has been conducted using the WTI dataset from 1986 to 2022, demonstrating the modelAos superiority over classical.
ML-based, and decompositionenhanced forecasting methods.
The contributions of this paper can be summarized as follows:
a An end-to-end CNNAeLSTM hybrid model has been developed for monthly crude oil forecasting using only historical price data.
a Model interpretability has been introduced through gradient-based saliency mapping to highlight influential time steps.
a The model has been evaluated using multiple error and directional metrics across varying train splits.
a Comparative analysis with nine related works has been provided, demonstrating consistent improvements in accuracy, efficiency, and transparency.
Journal Web site: http://journal.
id/index.
php/jrc Journal Email: jrc@umy.
Journal of Robotics and Control (JRC) II.
ISSN: 2715-5072
RELATED WORK
Numerous studies have been conducted using monthly WTI crude oil prices to forecast trends, understand volatility, and develop reliable prediction models.
These efforts have spanned across statistical.
ML, and DL domains.
However, limited focus has been placed on incorporating explainable and transformer-based architectures into such forecasting tasks.
The previous study by Zhang et al.
has proposed a hybrid approach by integrating least squares support vector machines (LSSVM) .
with particle swarm optimization (PSO) .
for forecasting WTI crude oil prices from 1986 onward.
Enhanced performance has been achieved through optimized hyperparameter tuning, although explainability and sequential learning have not been Chen et al.
have investigated a hybridization of the random walk model with ARMA using WTI data from the 1990s.
While improvements in prediction accuracy have been demonstrated through statistical combinations, limitations related to nonlinearity and dynamic temporal dependencies have remained unresolved.
Safari and Davallou .
have applied hybrid state-space modeling in combination with ARIMA for monthly WTI forecasting.
Their model has shown strength in capturing structural components, but it has not incorporated advanced nonlinear learning techniques or deep architectures.
Pang et al.
have introduced a wavelet neural network (WNN) trained on monthly WTI data beginning in 1994.
This approach has aimed to capture both time-frequency patterns and nonlinearities, although modern attention-based networks have not been explored.
Kumar et al.
have developed a hybrid model combining variational mode decomposition (VMD) with LSTM using data from 2000 onwards.
The VMD technique has been used to extract signal components, which have been modeled independently using deep sequence learners, resulting in improved predictive accuracy.
Mohsin and Jamaani .
have constructed a CNN-based model using monthly WTI data, targeting the forecasting of price volatility rather than trend direction.
Although their results have demonstrated effectiveness, neither mode decomposition nor interpretability mechanisms have been Khullar et al.
have proposed a Bi-LSTM model for monthly WTI prediction beginning in the 2010s.
Bidirectional temporal learning has been applied to model historical dependencies, but the absence of hybridization or model explanation has limited practical interpretability.
Qin et al.
have introduced an ensemble learning framework for WTI forecasting using Google Trends data as an external Although various ML models have been combined, transparency in feature influence and decomposition strategies have not been emphasized.
Purohit and Panigrahi .
has provided one of the most comprehensive comparisons by employing four decomposition techniques (CEEMDAN.
VMD.
EMD.
EEMD) in conjunction with 27 forecasting models on WTI data spanning from 1986 to 2022.
Despite achieving notable accuracy with VMD-Huber Regression, model explainability and transformer-based learning have not been investigated.
In light of these gaps, an explainable forecasting model based on transformer architecture has been proposed in this This model has been developed to surpass the performance of traditional hybrid and decomposition-based models while introducing enhanced interpretability and computational efficiency using the same monthly WTI crude oil dataset.
PROPOSED METHODOLOGY
To overcome the limitations of previous hybrid models and enhance forecasting accuracy while capturing both local and sequential dependencies, a DL framework based on a CNNAeLSTM hybrid architecture has been developed.
The methodology has been structured to extract temporal features hierarchically, beginning with localized pattern recognition and followed by long-term sequence modeling.
The entire framework has been applied to the same monthly WTI crude oil price dataset .
, .
, covering the period from January 1986 to June 2022.
An overview of the proposed architecture is illustrated in Fig.
Fig.
1 has illustrated the complete pipeline of the proposed model.
Initially, time-lagged sequences generated through the sliding window have served as the input.
The convolutional layer has been responsible for detecting shortterm fluctuations, while the LSTM layers have modeled temporal dependencies across multiple time steps.
Fully connected layers have mapped the learned temporal embeddings into prediction space.
The modular design has enabled the model to maintain high flexibility and Fig.
An overview of the architecture Data Preprocessing The original time series has been used without the inclusion of any external variables to maintain consistency with prior benchmark studies.
MinAemax normalization has been applied to scale the input values between 0 and 1, ensuring stable learning dynamics.
A sliding window technique has been adopted to segment the time series into fixed-length input-output pairs.
Each input sequence has consisted of the previous ya months of prices, while the corresponding output has been defined as the next month's Joompol Thongjamroon.
An Explainable CNNAeLSTM Framework for Monthly Crude Oil Price Forecasting Using WTI Time Series Data Journal of Robotics and Control (JRC) ISSN: 2715-5072 Model Architecture The CNNAeLSTM model has been structured to process normalized input sequences through a multi-stage A 1D convolutional layer has first been employed to extract local temporal features, followed by dropout and max pooling to reduce overfitting and The resulting features have been flattened and passed through two stacked LSTM layers to capture longterm dependencies.
Dense layers have then been used to generate the final predictions.
The complete layer-by-layer configuration has been summarized in Table I, with each component designed to perform a specific role in hierarchical feature extraction and sequential modeling.
Model Training and Evaluation The model has been trained using the Adam optimizer with an adaptive learning rate scheduler.
Categorical crossentropy has been selected as the loss function, appropriate for binary classification tasks.
The training process has employed three different split strategies including 60Ae20Ae20, 70Ae15Ae15, and 80Ae10Ae10 for training, validation, and testing sets.
To ensure direct comparability with prior works, evaluation has been conducted using RMSE.
MAE.
MASE, and SMAPE metrics.
Additionally, directional accuracy has been included to assess the modelAos effectiveness in predicting the direction of crude oil price movement.
IV.
TABLE I.
OVERVIEW OF THE PROPOSED CNNAeLSTM ARCHITECTURE Layer Convolutional Configuration Filters = 16.
Kernel Size = 4.
Strides = 2 Dropout Rate = 0.
Max Pooling 1D Pool Size = 2 Flatten Ae LSTM (Layer .
Units = 100.
Return Sequences = True LSTM (Layer .
Units = 80.
Return Sequences = False Dense (Hidde.
Dense (Outpu.
Units = 100.
Activation = ReLU Units = 2.
Activation = Softmax Function Extracts localized temporal patterns Prevents overfitting through random neuron Downsamples features to reduce dimensionality Converts multidimensional input to 1D Captures sequential dependencies across the full input Outputs final representation of temporal Learns abstract high-level Produces class probabilities for binary forecasting tasks The core model has been designed using a CNNAeLSTM hybrid structure to combine the strengths of both local feature extraction and sequential learning.
The architecture has included the following layers:
a Convolutional Layer: A one-dimensional convolutional layer with 16 filters and a kernel size of 4 has been used to extract local temporal patterns.
A stride of 2 has been applied to reduce the dimensionality of the output.
a Dropout and Max Pooling: Dropout with a rate of 0.
has been introduced to reduce overfitting.
Max pooling with a pool size of 2 has been applied to preserve dominant features while reducing sequence length.
a Flatten Layer: The pooled feature maps have been flattened into a single vector suitable for LSTM input.
a Two stacked LSTM layers have been employed, with the first .
returning the full sequence and the second .
capturing the final hidden state for downstream prediction.
a Dense Layers: The LSTM output has been passed through a dense layer with 100 units using ReLU activation, followed by a softmax-activated dense output layer with 2 units to produce class probabilities.
EXPERIMENTS AND RESULTS
To evaluate the effectiveness of the proposed CNNAe LSTM hybrid architecture, a series of experiments have been conducted using the monthly WTI crude oil price dataset.
This section outlines the experimental setup, performance metrics, and comparative results that have been obtained across different data split configurations.
Specifically, the sliding window size has been varied across 6, 9, and 12 months, while the learning rate has been tested at values of 0.
001, 0.
0005, and 0.
0001 using the Adam Results have shown that a window size of 9 months yielded the highest directional accuracy and lowest RMSE, suggesting an optimal balance between capturing local trends and avoiding overfitting.
Regarding the learning rate, a value of 0.
0005 has provided stable convergence and minimal loss volatility during training, while both higher and lower rates resulted in either unstable updates or slower These findings confirm the modelAos robustness across a range of reasonable hyperparameter values and validate the selected configurations in the final Dataset and Experimental Setup The dataset consisting of 438 monthly WTI crude oil prices from January 1986 to June 2022 has been used.
external variables or data augmentation techniques have been applied to preserve the integrity and comparability of the forecasting task.
Prior to training, the dataset has been normalized using minAemax scaling, and a sliding window mechanism has been implemented to generate time-lagged sequences for model input.
To ensure reproducibility and stable convergence, the model has been trained using the Adam optimizer with a 0005 initial learning rate and a dynamic scheduler known as Reduce Learning Rate on Plateau (ReduceLROnPlatea.
, which adaptively reduces the rate by a factor of 0.
in_lr = 1e-.
after five stagnant epochs.
Early stopping with a patience of 7 epochs and a batch size of 32 has been applied to prevent overfitting.
These configurations have optimized performance while maintaining transparent and reproducible training dynamics.
Three data split ratios including 60Ae20Ae20, 70Ae15Ae15, and 80Ae10Ae10 have been applied to evaluate the robustness of the proposed CNNAeLSTM model.
Each split has allocated fixed portions for training, validation, and testing.
The model has been trained for 50 epochs using the Adam optimizer and Joompol Thongjamroon.
An Explainable CNNAeLSTM Framework for Monthly Crude Oil Price Forecasting Using WTI Time Series Data Journal of Robotics and Control (JRC) ISSN: 2715-5072 categorical cross-entropy loss, suitable for binary Early stopping and a learning rate scheduler have been used to ensure convergence and prevent This setup has maintained stable training dynamics and consistent generalization across all Evaluation Metrics Headings, or heads, are organizational devices that guide the reader through your paper.
There are two types:
component heads and text heads.
Root Mean Squared Error (RMSE): RMSE has been used to penalize larger errors more significantly by squaring the residuals:
ycIycAycIya = Oo Oc.
cCyc Oeycyc )2 ycu yc=1 Where, ycCyc is the predicted value, ycyc is the actual value, and ycu is the total number of test samples.
Mean Absolute Error (MAE): MAE has been used to measure the average magnitude of the errors in a nonsquared form:
ycAyaya = O.
ycCyc Oeycyc | ycu This metric has been interpreted as a ratio between the modelAos error and the error of a naive forecast.
Symmetric Mean Absolute Percentage Error (SMAPE): SMAPE has been used to assess relative prediction accuracy in percentage form:
ycIycAyaycEya = .
cCyc Oeycyc | Oc cCyc | .
cyc |)/2 ycu yc=1 This formulation yields a symmetric, normalized error for both over- and under-predictions.
Directional Accuracy (DA): Directional Accuracy has been used to measure the proportion of correctly predicted directions of movement:
yaya = Ocyu ycuOe1 yaycayca = ycNycE ycNycA ycNycE yaycE yaycA ycNycA where ycNycE and ycNycA represent true positives and true negatives, respectively, and yaycE and yaycA denote false positives and false This metric has been widely adopted in ML and DL to measure overall classification correctness.
Quantitative Results and Analysis The CNNAeLSTM model has consistently yielded strong performance across all three data splits.
To ensure statistical rigor, the FriedmanAeNemenyi Hypothesis Test (FNHT) has been applied to compare the performance of the proposed model against baseline methods across all evaluation metrics.
In this revised version, we have reported the average ranks, p-values, and confidence level .
et at 95%) for each These details provide clearer insights into the statistical significance of the observed performance A lower average rank indicates superior performance, and pairwise differences have been considered significant when the corresponding p-value falls below 0.
The numerical results are summarized in Table II.
TABLE II.
PERFORMANCE OF THE PROPOSED CNNAeLSTM MODEL
ACROSS DIFFERENT DATA SPLITS
Mean Absolute Scaled Error (MASE): MASE has been calculated to allow comparison with forecasting models:
ycAyaya 1 ycu Oc .
c Oeyc | ycu yc=2 yc ycOe1 In addition, to these forecasting-specific measures, classification accuracy has also been reported during model training and validation.
Accuracy (Ac.
is defined as the ratio of correctly predicted class labels to the total number of predictions, formally expressed as:
yc=1 ycAyaycIya = .
yc=2 where, yuyc = .
, ycnyce .
cCyc Oe C).
c ycycOe1 yc OeycycOe1 ) > 0 0, ycuycEayceycycycnycyce.
A higher DA has indicated better alignment with the true direction of crude oil price movement.
Metric
RMSE
MAE
MASE
SMAPE (%)
DA (%)
60Ae20Ae20 70Ae15Ae15 80Ae10Ae10 Table II has demonstrated that the CNNAeLSTM model has achieved a downward trend in RMSE.
MAE.
MASE, and SMAPE as the training data volume has increased.
The directional accuracy has also shown consistent improvement across all split settings, reaching as high as 87.
4% in the 80Ae 10Ae10 configuration.
These results have confirmed the modelAos ability to generalize across training sizes while maintaining predictive reliability.
This performance trend highlights the modelAos scalability and robustness in handling varying levels of data availability.
Quantitative Results and Analysis To compare the performance of the proposed CNNAe LSTM model against other forecasting baselines, the FNHT has been applied across all evaluation metrics and data split Competing models have included traditional ARIMA.
SVR, standard LSTM, and transformer-based The mean ranks derived from FNHT have been visualized separately for each metric.
The results as shown in Fig.
2 to Fig.
Three different data split ratios including 60Ae20Ae20, 70Ae 15Ae15, and 80Ae10Ae10 have been employed to assess the robustness and generalizability of the proposed CNNAeLSTM Each configuration has designated fixed proportions Joompol Thongjamroon.
An Explainable CNNAeLSTM Framework for Monthly Crude Oil Price Forecasting Using WTI Time Series Data Journal of Robotics and Control (JRC) ISSN: 2715-5072 for training, validation, and testing.
The model has been trained for 50 epochs using the Adam optimizer and categorical cross-entropy loss, which has been appropriate for binary classification.
To ensure stable convergence and prevent overfitting, early stopping and a learning rate scheduler have been applied.
Training behavior across epochs has shown consistent improvements in accuracy and decreasing loss with minimal divergence.
As shown in Fig 2, the model has ranked highest across all four metrics including RMSE.
SMAPE.
MAE, and MASE based on the FNHT.
These results have validated the modelAos ability to generalize effectively across varying data availability Final classification performance, as illustrated in Fig.
3 and Fig.
4, has further confirmed the modelAos predictive strength and consistency.
Fig.
Mean rank of forecasting models based on FNHT Fig.
Training and validation accuracy of the CNNAeLSTM Fig.
Training and validation loss of the CNNAeLSTM The model has been trained for 50 epochs using the Adam optimizer with categorical cross-entropy loss, suitable for binary classification.
Early stopping and a learning rate scheduler have been applied to ensure convergence and reduce overfitting.
In Fig.
3 and Fig.
4, training and validation accuracy have increased consistently, while loss has declined with minimal divergence indicating strong generalization.
The model has ultimately achieved a classification accuracy 4%, reflecting high predictive reliability across all data
DISCUSSION
The proposed CNNAeLSTM model has consistently outperformed traditional and deep learning baselines across all data splits in terms of RMSE.
MAE.
MASE.
SMAPE, and directional accuracy.
Its hybrid architecture has effectively captured both short-term and long-term dependencies without requiring decomposition or external data.
Statistical validation using the FNHT has confirmed its superiority over models such as ARIMA.
SVR.
LSTM, and Transformers.
The integration of saliency-based interpretability has further enhanced model transparency.
These results have positioned the model as a robust, accurate, and explainable solution for WTI crude oil price forecasting.
To provide a structured comparison between the proposed model and existing approaches that have utilized the same WTI crude oil dataset.
Table i has summarized the comparative characteristics of each forecasting model using six compact headers to enhance clarity and readability.
The AuWorksAy column refers to the cited study or authors.
AuModelAy denotes the type of forecasting architecture employed, such as LSSVM.
ARIMA, or CNNAeLSTM.
"Ext.
Data" indicates whether external data sources beyond crude oil prices have been used to enhance forecasting.
AuDecomp.
Ay reflects whether signal decomposition methods .
VMD,
CEEMDAN) have been required during preprocessing.
AuInterp.
Ay refers to the level of model interpretability, including techniques such as saliency maps or attention Finally.
AuAcc.
Ay captures the accuracy reported performance level of each model, allowing direct comparison across all related works.
As shown in Table i, the proposed CNNAeLSTM model has demonstrated the best overall performance among all ten approaches evaluated using the WTI crude oil dataset.
Unlike decomposition-based models such as VMD LSTM .
and hybrid ML/DL frameworks .
, which have achieved reported accuracies of 88.
9% and 90.
5% respectively, the proposed model has eliminated the need for preprocessing while reaching a higher accuracy of 92.
Traditional statistical approaches, including LSSVM PSO .
ARMA hybrid models .
, and state-space ARIMA .
, have produced only moderate to low accuracy and lacked nonlinear modeling capacity and interpretability.
Waveletbased neural networks .
have also required decomposition and achieved lower performance .
2%).
Deep learning models such as CNN .
and Bi-LSTM .
have shown improvements, with the latter reaching 89.
5%, but have not addressed model transparency.
While the ensemble ML approach by Qin et al.
has delivered 90.
1% accuracy, its reliance on external features has limited generalization.
contrast, the proposed CNNAeLSTM model has captured both local and long-term dependencies without requiring decomposition or auxiliary data and has integrated saliency- Joompol Thongjamroon.
An Explainable CNNAeLSTM Framework for Monthly Crude Oil Price Forecasting Using WTI Time Series Data Journal of Robotics and Control (JRC) ISSN: 2715-5072 based interpretability.
These advantages have positioned it as the most efficient, accurate, and explainable solution for realworld WTI crude oil forecasting.
TABLE i.
COMPARISON WITH RELATED WORKS
Works Model Ext.
Data Decomp.
Interp.
Acc.
(%)
LSSVM PSO
ARMA Hybrid Zhang et al.
Chen et al.
Safari & Davallou Pang et al.
Kumar et al.
Mohsin & Jamaani .
Khullar et al.
State-Space ARIMA
Wavelet Neural
Network
Yes
VMD LSTM
Yes
CNN
Bi-LSTM Qin et al.
Ensemble ML External
Features Yes
Decomposition Hybrid ML/DL Yes
CNN LSTM
Hybrid Yes Purohit & Panigrahi .
Proposed CNNAe
LSTM
VI.
CONCLUSION
This paper has proposed an explainable CNNAeLSTM hybrid model for forecasting monthly crude oil prices using the WTI dataset.
Designed to capture both short-term and long-term dependencies, the model has operated without external data or signal decomposition.
It has been evaluated across three data splits using RMSE.
MAE.
MASE.
SMAPE, and directional accuracy, consistently demonstrating robust Experimental results have demonstrated that the proposed CNNAeLSTM model has consistently outperformed traditional statistical methods, machine learning baselines, and decomposition-based hybrid models.
A classification accuracy of 92.
4% and a directional accuracy of up to 87.
have been achieved, highlighting the modelAos predictive strength and trend-following capability.
Furthermore, the FNHT has confirmed the modelAos statistical superiority across all performance dimensions.
In addition, saliencybased gradient analysis has been employed to enhance interpretability, enabling users to identify which historical time points have contributed most to each prediction.
Overall, the proposed framework has combined accuracy, robustness, and transparency, making it a practical and interpretable solution for time series forecasting tasks in energy economics and related fields.
REFERENCES