Foundational Transformer Models for Financial Time Series Forecasting

Predicting short-horizon equity price movements remains a fundamental challenge in quantitative finance. Forecasting models typically use regression metrics like RMSE, measuring prediction fidelity rather than trading performance. Deep learning, particularly pretrained time-series foundation models, has renewed interest in this task through zero-shot forecasting. This thesis addresses that gap through a multi-phase empirical study of modern forecasting methods under a single, integrity-checked protocol. We compare prompt-based large language models, the Chronos family of pretrained forecasters, xLSTM-ts, StockMixer, and cgc_x, a novel hybrid architecture that couples a frozen Chronos-T5 backbone with an xLSTM covariate encoder trained under walk-forward optimisation. A systematic integrity audit identified global normalisation leakage in the StockMixer and xLSTM-ts pipelines; both were patched before evaluation. On standard regression metrics, patched baselines and Chronos variants achieve similar next-day price error (RMSE ~3.6). The main distinction emerges in cross-sectional ranking: cgc_x v1 achieves the highest information coefficient (IC = 0.093, IC IR = 0.261), indicating superior within-day asset ranking. Its RMSE remains comparable to zero-shot Chronos-T5, suggesting walk-forward domain adaptation strengthens cross-sectional signals without materially improving absolute price forecasts. ChronosX achieves the second-highest IC, indicating covariate injection also improves ranking relative to zero-shot inference. The audit framework, evaluation protocol, and patched reruns provide a reproducible baseline for future work. The study is deliberately bounded: next-day daily forecasting, a 10-symbol universe, and frictionless backtests.