On the recurrent neural network model with robust expectile-based loss function in economic data forecasting.

Area of Science:

• Machine Learning
• Econometrics
• Time Series Analysis

Background:

• Recurrent Neural Networks (RNNs), including LSTM and GRU, are standard for sequential data but struggle with non-stationary and heterogeneous time series.
• Their limitation stems from symmetric loss functions (e.g., MSE) assuming data homogeneity.
• This hinders accurate forecasting across diverse data patterns and conditions.

Purpose of the Study:

• To propose a novel Expectile-based Recurrent Neural Network (E-RNN) framework integrating expectile regression into RNNs.
• To develop and compare E-LSTM and E-GRU variants for superior time series forecasting.
• To enable scenario-based forecasting (pessimistic to optimistic) by adjusting an asymmetric parameter (τ).

Main Methods:

• Developed Expectile-based Recurrent Neural Network (E-RNN) variants: E-LSTM and E-GRU.
• Utilized an asymmetric least squares loss function to model conditional data distributions beyond central tendency.
• Implemented Expectile-based Generalized Approximate Cross Validation (EGACV) for robust model selection.

Main Results:

• E-RNN models demonstrated superior performance in forecasting Indonesia's quarterly economic growth.
• Achieved lower EGACV scores and higher forecast accuracy compared to standard RNNs.
• Showed significant improvements on volatile quarter-to-quarter (qtq) data, enhancing forecast reliability.

Conclusions:

• E-RNNs provide adaptive forecasting models resilient to changes in data distribution, overcoming homogeneity assumptions.
• The EGACV criterion offers a robust method for balancing model fit and complexity.
• The framework allows for generating diverse forecast scenarios by adjusting the asymmetry parameter (τ).