Trend time-series modeling and forecasting with neural networks.

Area of Science:

• Time-series analysis
• Machine learning
• Econometrics

Background:

• Modeling trends in time-series data lacks a general consensus.
• Neural networks (NNs) offer potential advantages over traditional forecasting methods.
• Understanding NN performance across diverse trend types is crucial.

Purpose of the Study:

• To investigate optimal strategies for modeling trend time series using neural networks.
• To compare the effectiveness of different data preprocessing techniques with NNs.
• To evaluate NN performance on various trend patterns including linear, nonlinear, deterministic, stochastic, and breaking trends.

Main Methods:

• Utilized neural networks (NNs) for time-series forecasting.
• Applied four distinct data strategies: raw data, raw data with time index, detrending, and differencing.
• Tested models on simulated data representing diverse trend patterns and a real-world Gross National Product (GNP) series.

Main Results:

• Differencing the time-series data consistently produced meritorious results across various underlying data generating processes (DGPs) when using NNs.
• The effectiveness of differencing was robust, irrespective of the trend's nature (linear, nonlinear, deterministic, stochastic, breaking).
• Real-world Gross National Product (GNP) data corroborated the findings on differencing.

Conclusions:

• Differencing is a highly effective preprocessing strategy for neural network-based time-series trend modeling.
• The proposed NN approach with differencing provides a reliable method for forecasting complex time-series trends.
• This study offers practical guidance for applying NNs to trend analysis in economics and other fields.