Forecasting and Simulation Scenarios Based on the TFT Model

As the global smart railways market is expected to grow from 149 million USD in 2025 to 1,124.5 billion USD by 2035 at a CAGR of 22.4%, railway enterprises in the digital transformation process shall require accurate and interpretable economic forecasting tools. Temporal Fusion Transformer (TFT) by Lim combines transformer-style self-attention, recurrent layers with gating networks to enjoy the capability of processing heterogeneous inputs while producing interpretable factor attribution over multiple time horizons. While TFT has been previously shown to be effective for energy and retail forecasting, it has not yet been used for railway transport economic forecasting in developing economies and also we are unaware of any other study that makes a singular–applied economics framework by integrating TFT-driven prediction with scenario-based simulation. 2015–2025 data for “Uzbekistan Railways” JSC, this study develops a log-sequence econometric proxy model compatible with the TFT and then computes predictions for gross transport revenue to 2031 for baseline, pessimistic and optimistic trajectories. The model scored an R² of 0.9797 (ADJ R²), an internal MAPE of 1.96% and a walk-forward back test MAPE of 6.22%. In the baseline scenario, revenue is forecast to grow from 12,312.13 billion UZS in 2026, climbing to 13,598.39 billion UZS in 2031, and improvements in international freight share, expansion of digital channel and deepening of automation are identified as sequential drivers of growth up to 2031. This is the first attempt to combine TFT-based interpretable forecasting with scenario simulation derived from walk-forward error bands as applied to railway economics in a Central Asian developing economy context. As a first step, the findings constitute a strategic forward looking forecasting tool that converts prediction from back casting to decision support. The next step for future research should be to expand the framework to high-frequency data, cross-country panel models, and ESG-embedded sustainability forecasting.