Computational Analysis for Enhanced Forecasting of India’s GDP Growth using a Modified LSTM Approach

Abstract

Forecasting GDP growth accurately is critical for effective legislation, investment strategies, and corporate planning decision-making. This study focuses on the creation of a novel modified Long Short-Term Memory (LSTM) model for forecasting India's GDP growth. Outlier identification employs strong statistical techniques, such as the modified Z-score method, to identify and deal with extreme observations that may significantly impact the forecasting model's performance. Missing data is a typical issue in economic datasets. Imputation approaches, such as the Expectation-Maximization (EM) or MissForest algorithms are used to impute missing values by leveraging the correlations between variables and observed data. These advanced imputation techniques consider the data's complexity and produce more reliable and accurate imputed values, reducing biases in the subsequent forecasting process. Lasso regression with lagged variables is applied to select relevant features for forecasting GDP growth. Furthermore, Time Series SMOTE is used to address the class imbalance challenge in GDP growth datasets. Then, the improved LSTM model is then trained on the pre-processed data using robust optimization methods. Cross-validation procedures are used to validate the model's capacity to generalize and minimize overfitting. Performance metrics such as MAE, RMSE, MPE and forecast bias are used to assess the accuracy and precision of GDP growth forecasts.  The proposed system has achieved better performance compared to existing methods.