Scholarlink

<p>In cloud computing, cost optimization remains a critical challenge for businesses using cloud services. Spot pricing, a dynamic pricing model where costs fluctuate based on supply and demand, offers an opportunity for cost savings but introduces complexity in budget planning due to its unpredictable nature. This study analyses neural network and deep learning techniques to forecast Amazon EC2 (Amazon Elastic Compute Cloud) spot prices, aiming to empower users with predictive insights for strategic decision-making. We compare the predictive performance of Long-Short Term Memory (LSTM) networks, renowned for their strengths in handling time series and pattern recognition tasks, with Multilayer Perceptron (MLP), a classical model and Convolutional Neural Network (CNN), a newer approach. We later highlight the impact of auto-correlation in residuals on prediction accuracy of the models. The datasets used were collected from the Amazon EC2 platform encompassing spot prices over a 90-day period. Initial results reaffirm the LSTM’s effectiveness, aligning with the literature, while the performances of MLP and CNN are promising, offering comparable accuracy metrics based on residuals. All models achieved low residual errors, with some CNN configurations reaching a Mean Absolute Percentage Error (MAPE) below 1%. A Durbin-Watson test revealed severe auto-correlation in the residuals of all models (test statistics consistently near 0). This indicates that the models’ apparent accuracy may be an artifact of the data’s statistical properties rather than a true fit to its underlying dynamics. This study underscores the importance of considering auto-correlation in the evaluation process, preventing overly optimistic conclusions about model accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Forecasting AWS Spot Prices: Comparative Analysis of Deep Learning Architectures and Residual Auto-Correlation

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Savio Berdine,
  • Valmir Macario,
  • Robson Wagner A. de Medeiros

摘要

In cloud computing, cost optimization remains a critical challenge for businesses using cloud services. Spot pricing, a dynamic pricing model where costs fluctuate based on supply and demand, offers an opportunity for cost savings but introduces complexity in budget planning due to its unpredictable nature. This study analyses neural network and deep learning techniques to forecast Amazon EC2 (Amazon Elastic Compute Cloud) spot prices, aiming to empower users with predictive insights for strategic decision-making. We compare the predictive performance of Long-Short Term Memory (LSTM) networks, renowned for their strengths in handling time series and pattern recognition tasks, with Multilayer Perceptron (MLP), a classical model and Convolutional Neural Network (CNN), a newer approach. We later highlight the impact of auto-correlation in residuals on prediction accuracy of the models. The datasets used were collected from the Amazon EC2 platform encompassing spot prices over a 90-day period. Initial results reaffirm the LSTM’s effectiveness, aligning with the literature, while the performances of MLP and CNN are promising, offering comparable accuracy metrics based on residuals. All models achieved low residual errors, with some CNN configurations reaching a Mean Absolute Percentage Error (MAPE) below 1%. A Durbin-Watson test revealed severe auto-correlation in the residuals of all models (test statistics consistently near 0). This indicates that the models’ apparent accuracy may be an artifact of the data’s statistical properties rather than a true fit to its underlying dynamics. This study underscores the importance of considering auto-correlation in the evaluation process, preventing overly optimistic conclusions about model accuracy.