Scholarlink

In this work, we compare the performance of a non-sequentially trained Multilayer Perceptron (MLP) for financial time series forecasting with more complex models like Time-series Neural Networks (TNN) and Long Short-Term Memory (LSTM) networks. The results show that a simple MLP architecture combined with feature engineering had performance comparable to more complex neural network models. Our best performing MLP achieved a mean absolute percentage error (MAPE) of 1.9% on S&P 500 data, matching LSTM’s performance and closely trailing the MAPE of TNN of 1.3%. Moreover, when tested on Bitcoin’s volatile price data, the trained MLP showed only a slight decrease in performance, showing strong generalization abilities across different market conditions. This suggests that our approach, which focuses on feature engineering rather than temporal patterns, can be successful even in highly volatile markets.

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

A Non-sequentially Trained MLP Model for Financial Time Series Forecasting

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Mohsen Larni,
  • Kazem Taghva

摘要

In this work, we compare the performance of a non-sequentially trained Multilayer Perceptron (MLP) for financial time series forecasting with more complex models like Time-series Neural Networks (TNN) and Long Short-Term Memory (LSTM) networks. The results show that a simple MLP architecture combined with feature engineering had performance comparable to more complex neural network models. Our best performing MLP achieved a mean absolute percentage error (MAPE) of 1.9% on S&P 500 data, matching LSTM’s performance and closely trailing the MAPE of TNN of 1.3%. Moreover, when tested on Bitcoin’s volatile price data, the trained MLP showed only a slight decrease in performance, showing strong generalization abilities across different market conditions. This suggests that our approach, which focuses on feature engineering rather than temporal patterns, can be successful even in highly volatile markets.