A HYBRID TEMPORAL FUSION TRANSFORMER FOR EXPLAINABLE AND EARLY RECESSION PREDICTION USING MIXED-FREQUENCY MACRO-FINANCIAL DATA

Abstract

Economic recessions, marked by soaring unemployment, business failures, and widespread financial distress, pose severe challenges to global economies, as evidenced by the 2008 financial crisis, which saw U.S. unemployment hit 10% and India’s GDP growth drop to 3.1%, and the 2020 pandemic-induced global GDP contraction of 3.5%. Accurate and timely recession forecasting is critical for policymakers, central banks, and businesses to implement preemptive measures that mitigate these impacts. This thesis introduces the Hybrid Temporal Fusion Transformer (TFT), a novel AI framework designed to predict recessions in India and the United States with exceptional accuracy and interpretability, addressing the shortcomings of traditional econometric models and opaque AI systems. Leveraging 2020–2024 macro-financial data, the Hybrid TFT integrates mixed-frequency indicators, including 1.7 million daily GST invoices, UPI transactions, monthly GST collections (Rs. 1.8 trillion in April 2024), and quarterly GDP growth (7.8% in Q1 2024 for India), alongside U.S.-specific metrics like the May 2024 yield curve inversion (-0.35%). The model incorporates India’s unique economic factors, such as monsoon variability affecting 50% of agricultural output and a 40% informal sector contribution to GDP, ensuring contextual relevance. The methodology comprises three core components: a data harmonization framework using Empirical Mode Decomposition to align heterogeneous data streams, an interpretable Temporal Fusion Transformer prioritizing key variables like the RBI’s 6.5% repo rate hike, and economic regularization embedding principles like Okun’s Law to ensure theoretical coherence. Backtesting with 2020–2024 data demonstrated the Hybrid TFT’s superior performance, achieving 93.2% accuracy for U.S. forecasts and 90.1% for India, outperforming Long Short-Term Memory models (77.3% India accuracy, 18.2% false alarms) and traditional RBI systems (80.2% India accuracy, 9.7% false alarms) by 10–13% and reducing false positives by 40%. A standout achievement was the early detection of India’s March 2024 election-related market stress, flagged 11 days in advance with 68% confidence, enabling proactive liquidity measures. The model’s integration of digital economy data captured small business activity, constituting 73% of India’s workforce, and uncovered insights like a 14% UPI payment dip signaling distress and GST collections below Rs. 1.8 trillion predicting liquidity crunches. The open-source platform and interactive dashboard empowered RBI analysts, businesses, and farmers with tools for policy simulations, inventory planning, and monsoon-adjusted crop strategies. Despite its success, the model faced limitations, including informal sector blind spots, underestimation of black swan events like the May 2024 oil price shock, and dashboard complexity requiring analyst training. Future research will focus on modeling tier-2 and tier-3 city economies, integrating climate data for rural forecasts, and enhancing explainability through natural language summaries for non-technical users. By bridging AI innovation with India’s economic realities—where Diwali and monsoons shape markets—the Hybrid TFT sets a new benchmark for recession forecasting, fostering resilience and informed decision-making across diverse stakeholders.