A Hybrid Autoregressive LSTM Model-based Gasoline Price Pre-dicting Method Using Optimal Time Window

Siyi Wei, Xin Zhang, Chaoran Zhou, Wenwei JiangICCEIC 2022 · OpenReview

TL;DR

This paper introduces a novel HARLSTM model for predicting gasoline prices by incorporating multiple influencing factors and optimizing the time window for feature extraction. The model demonstrates superior performance over baseline methods, offering more accurate and robust gasoline price predictions.

What Problem Does This Paper Address?

Accurately predicting gasoline prices is crucial for urban economic planning, energy security, and consumer behavior. However, gasoline prices are influenced by a complex interplay of factors, exhibiting non-stationarity, non-linearity, and time variability. Existing forecasting methods often rely on single factors or lack a systematic approach to incorporate diverse influences and temporal dynamics, leading to suboptimal prediction accuracy. This paper aims to bridge this gap by developing a comprehensive model that accounts for multiple factors and their temporal dependencies.

Methods

The core of this research is the Hybrid Autoregressive LSTM (HARLSTM) model. The authors propose a feature building method based on a data window to extract comprehensive feature data by considering multiple influencing factors. This approach aims to capture the non-stationarity, non-linearity, and time variability inherent in gasoline price data.

Key methodological steps include:

  • Feature Analysis: An impulse response function analysis using a vector autoregressive model is employed to identify key factors affecting gasoline prices and their impact duration.
  • Feature Construction: A multi-step time window approach is used to construct features, extending beyond single time steps to capture temporal dependencies.
  • HARLSTM Model: This model combines autoregressive principles with Long Short-Term Memory (LSTM) networks. The model architecture involves an initial preparation structure (RNN and Dense layers) to extract time series features, followed by a pseudo-cyclic neural network structure (LSTM cells) that iteratively predicts future values, using previous predictions as feedback.
  • Optimal Time Window Selection: The study investigates the impact of different time window lengths on prediction accuracy, aiming to identify the optimal window for capturing relevant temporal dynamics.

Datasets

The petrol price data used in this study is sourced from the official website of the U.S. Energy Information Administration (EIA). The dataset includes:

  • Retail petrol price data for cities in nine states, including Florida, from January 3, 2000, to June 28, 2021.
  • Crude Oil Price Data from Cushing, Oklahoma.
  • Fuel Oil Price from New York Port No. 2 Fuel Oil.
  • National Car Volume data from the Trade Economics Network.
  • Additional variables such as crude oil stocks, crude oil acquisition costs of refineries, gasoline imports, gasoline stocks, gasoline production, and gasoline demand from the Petroleum Authority area of the Defence Zone.

The paper does not explicitly mention the creation of new benchmark datasets; rather, it utilizes existing publicly available data.

Key Insights and Ideas

  • Multifactorial Influence is Crucial: Gasoline prices are not driven by a single factor. The paper highlights how demand, inventory, production, and imports all play significant, albeit sometimes complex, roles. For instance, increased gasoline inventory acts as a buffer, suppressing prices, while falling inventories can lead to price hikes.
  • Temporal Dynamics Matter: The "time window" is not just a passive input but an active component in prediction. The optimal window length (found to be 12 weeks) is critical for capturing the cyclical nature and influence periods of various factors on gasoline prices.
  • LSTM's Strength in Time Series: The paper reinforces the effectiveness of LSTM networks in handling the non-linear and time-varying nature of financial time series data, outperforming traditional models like MLP and SVR.
  • Autoregression Enhances LSTM: Combining autoregressive principles with LSTM (HARLSTM) allows the model to leverage past predictions as feedback, creating a more robust iterative prediction process.

Key Results

The HARLSTM model demonstrated superior performance compared to baseline models (MLP, Conv_dense, residual_LSTM, LGBM, RFR, SVR) across multiple evaluation metrics:

  • MSE (Mean Squared Error): The HARLSTM model achieved the lowest MSE of 0.008 for single-step predictions and 0.09 for multi-step predictions.
  • MAE (Mean Absolute Error): The HARLSTM model achieved the lowest MAE of 0.06 for single-step predictions.
  • MAPE (Mean Absolute Percentage Error): The HARLSTM model achieved the lowest MAPE of 0.19 for single-step predictions.
  • Optimal Window Length: The study found that a window length of 12 weeks yielded the best performance, with MSE, MAE, and MAPE reduced by at least 20%, 14%, and 3% respectively, compared to other window lengths.
  • Qualitative Results: Figure 4 shows that the predicted gasoline price trend closely matches the actual values, indicating good fitting performance.

Limitations and Future Work

  • Geographical Specificity: The current model's geographical scope is limited to the U.S. (specifically Florida for detailed analysis). The authors acknowledge that gasoline prices vary geographically and suggest future work to incorporate geographical feature information for more accurate and objective predictions.
  • Data Granularity: While the paper discusses time stamp consistency, the exact granularity of all input features (daily, weekly, etc.) and how they are aggregated or aligned for the 12-week prediction window could be further elaborated.
  • Model Complexity: While HARLSTM shows improved accuracy, the computational cost and interpretability of such deep learning models can be a consideration.

Future work proposed by the authors includes integrating geographical features into the HARLSTM model to enhance prediction accuracy and objectivity.

Summary updated

BibTeX
@inproceedings{wei2022hybrid,
  author = {Siyi Wei and Xin Zhang and Chaoran Zhou and Wenwei Jiang},
  title = {{A Hybrid Autoregressive LSTM Model-based Gasoline Price Pre-dicting Method Using Optimal Time Window}},
  booktitle = {ICCEIC},
  year = {2022},
  url = {https://openreview.net/forum?id=0wfMFZ2fmj},
}
Figure