A Study on Time Series Data Forecasting Methods Using ARIMA and BP Neural Network Hybrid Modelling

Abstract

This study constructs a multi-model prediction framework integrating ARIMA and BP neural networks, aiming to reveal the dynamic relationships between variables and achieve accurate trend forecasting. First, multiple linear regression is used to preliminarily analyse the association mechanism between influencing factors and target variables. EWM-TOPSIS is then employed to allocate indicator weights and evaluate performance, providing a basis for feature selection in subsequent model construction. Furthermore, the ARIMA model is used to pre-process time series data and construct the model, enabling trend prediction and verification of model residual characteristics. Additionally, a three-layer neural network structure is constructed using the BP neural network model, with neuron configurations and activation functions set for the input layer, hidden layer, and output layer to enable complex system prediction driven by non-linear features. This framework achieves end-to-end modelling from variable association analysis to dynamic trend forecasting through multi-level data processing and model optimisation, providing a technical paradigm that combines interpretability and predictive accuracy for relevant fields.