Data preprocessing techniques and neural networks for trended time series forecasting

Research on time series forecasting continues to attract significant attention, particularly in the use of Artificial Neural Networks (ANN) due to their ability to model nonlinear behaviors. However, forecasting economic time series with steep upward trends presents challenges, often leading to poorly fitting predictions. This study addresses the issue by applying differentiation as a preprocessing step. Three real-world time series exhibiting this behavior were analyzed and forecasted using two neural network models—Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP)—with and without preprocessing. The differentiated series were further processed using techniques such as Empirical Mode Decomposition (EMD) and trend-fluctuation decomposition via Moving Average of Wavelet Transform. The results demonstrate that differentiation significantly enhances forecasting accuracy across all tested models, reducing errors by up to 30 % compared to models without preprocessing. This approach effectively mitigates trend-related distortions, leading to more reliable predictions in complex economic time series. • Differentiating time series before forecasting enhances prediction accuracy in complex, nonlinear economic data. • Techniques like Empirical Mode Decomposition and Wavelet Transform improve ANN performance in time series forecasting. • LSTM and MLP networks achieve higher accuracy with differentiated time series, surpassing models without preprocessing.