Scholarlink

<p>This study proposes a novel hybrid framework for stock price prediction on the NASDAQ index by integrating Autoencoder (AE) for dimensionality reduction, Long Short-Term Memory (LSTM) networks for temporal modeling, and the Multi-Objective Seahorse Optimization (MOSHO) algorithm for hyperparameter tuning. Financial time series are inherently high-dimensional, noisy, and non-linear, making traditional models such as Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Networks (ANN) inadequate. The AE component in our pipeline reduces redundant features and denoised the input, the LSTM learns temporal dependencies, and MOSHO simultaneously optimizes prediction error, model complexity (total trainable parameters), and directional accuracy (correct trend prediction).evaluate the model across ten major NASDAQ stocks (e.g., AAPL, MSFT, GOOGL) over multiple time horizons—yearly, daily, and hourly—from 2018 to 2024. The proposed AE-LSTM-MOSHO model outperformed LSTM-GA (Genetic Algorithm), LSTM-1D, LSTM-2D, LSTM-3D, and baseline ANN models. It achieved the lowest Mean Absolute Percentage Error (MAPE) of 1.94% to 2.98% and highest R² scores from 0.889 to 0.938 across stocks and timeframes. Hourly predictions particularly benefited from this integration, confirming superior responsiveness to short-term market fluctuations. This multi-objective strategy presents a deployable and scalable solution for modern financial forecasting challenges.</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

Bio-Inspired deep learning framework for stock prices prediction using autoencoder-compressed LSTM and MOSHO

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Sukhbir Singh,
  • Vijay Dhir

摘要

This study proposes a novel hybrid framework for stock price prediction on the NASDAQ index by integrating Autoencoder (AE) for dimensionality reduction, Long Short-Term Memory (LSTM) networks for temporal modeling, and the Multi-Objective Seahorse Optimization (MOSHO) algorithm for hyperparameter tuning. Financial time series are inherently high-dimensional, noisy, and non-linear, making traditional models such as Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Networks (ANN) inadequate. The AE component in our pipeline reduces redundant features and denoised the input, the LSTM learns temporal dependencies, and MOSHO simultaneously optimizes prediction error, model complexity (total trainable parameters), and directional accuracy (correct trend prediction).evaluate the model across ten major NASDAQ stocks (e.g., AAPL, MSFT, GOOGL) over multiple time horizons—yearly, daily, and hourly—from 2018 to 2024. The proposed AE-LSTM-MOSHO model outperformed LSTM-GA (Genetic Algorithm), LSTM-1D, LSTM-2D, LSTM-3D, and baseline ANN models. It achieved the lowest Mean Absolute Percentage Error (MAPE) of 1.94% to 2.98% and highest R² scores from 0.889 to 0.938 across stocks and timeframes. Hourly predictions particularly benefited from this integration, confirming superior responsiveness to short-term market fluctuations. This multi-objective strategy presents a deployable and scalable solution for modern financial forecasting challenges.