The study utilises a massive dataset comprising approximately two billion observations across 94 countries over 34 years. It compares the performance of various TSFM regimes—zero-shot inference, fine-tuning, and pre-training from scratch—against traditional benchmarks like linear models and tree-based ensembles. The findings indicate that while off-the-shelf TSFMs often underperform, those pre-trained specifically on financial data show significant predictive and economic improvements, particularly when augmented with synthetic data or additional financial factors.