Integrating LSTM Networks with Neural Levy Processes for Financial Forecasting

A new study introduces a Long Short-Term Memory (LSTM) model designed to predict NBA game outcomes using a comprehensive dataset spanning from the 2004-05 to 2024-25 seasons. This model utilizes an extensive sequence of 9,840 games to effectively capture evolving team dynamics and dependencies across seasons, addressing challenges faced by traditional prediction models.

A recent study highlights the issue of data leakage in Long Short-Term Memory (LSTM) networks used for time series forecasting, revealing that improper sequence construction before dataset partitioning can lead to misleading evaluation results. The research evaluates three validation techniques under both leaky and clean conditions, demonstrating how validation design can influence leakage sensitivity and performance metrics such as RMSE Gain.

A new real-time air pollution prediction model has been developed in Daegu, Korea, utilizing spatiotemporal big data collected from sensors installed on taxis. This model employs a Convolutional Neural Network (CNN) for spatial distribution analysis and integrates Long Short-Term Memory (LSTM) units to account for temporal data and other influencing factors like weather conditions.

The Neural Physics approach has been utilized to estimate parameters for a simple land-surface model, employing PyTorch's backpropagation engine for optimization. A synthetic dataset was generated to test the inverse model, revealing that reliable parameter estimation requires soil temperature measurements at two depths rather than one. The model was applied to urban flux tower data in Phoenix, allowing for the estimation of key thermal properties.