Neural–Volatility Intelligence (NVI): Code and Data for Theory-Informed Hybrid Remittance-Cost Forecasting

Abstract

This repository contains the full data-acquisition, preprocessing, and modelling pipeline supporting the paper "Neural–Volatility Intelligence: A Theory-Informed Hybrid Approach for Remittance Cost Forecasting and Policy Optimisation in Digital Financial Ecosystems" (Moroke & Makatjane). NVI is a hybrid forecasting framework that combines an Attention-Based Bidirectional LSTM with a GJR-GARCH volatility model through a trainable, market-state-conditioned softmax ensemble gate, with theory-derived economic sign restrictions embedded directly in the loss function as differentiable penalty terms rather than checked post-estimation. The analysis uses 5,204 real daily observations (2011–2025) from the Philippines–United States remittance corridor, drawn from the World Bank's Remittance Prices Worldwide database, the World Bank Global Findex Database, the Frankfurter exchange-rate API, and the World Bank CPI indicator. No synthetic or simulated data are used. Included: data-download and dataset-construction scripts; the full benchmark suite (Naive, ARIMA, ARIMA-GARCH, GJR-GARCH, Random Forest, XGBoost, GARCH-ANN, Attention-BiLSTM); stationarity and hypothesis-testing code; and the NVI model itself.