This paper introduces a novel retrieval-augmented forecasting framework called macro-contextual retrieval, designed to address the challenge of non-stationarity in financial markets. The key idea is to ground each prediction in historically analogous macroeconomic regimes by jointly embedding structured macro indicators (CPI, unemployment, yield spread, GDP growth) and financial news sentiment into a shared similarity space. During inference, the model retrieves the top-K most similar historical days (using FAISS with causal masking) and conditions its forecast on the average of their text embeddings, effectively learning to reason by precedent without retraining.

The framework is trained on 17 years of S&P 500 data (2007-2023) and evaluated out-of-distribution on AAPL and XOM in 2024. Results show that macro-conditioned retrieval consistently outperforms static baselines (numeric-only, text-only, naive multimodal fusion) under regime shifts. It is the only configuration achieving positive trading outcomes (Profit Factor > 1, Sharpe ratio > 0) on both OOD assets, while baselines collapse with negative risk-adjusted returns. Moreover, the retrieved neighbors form interpretable evidence chains that correspond to recognizable macro contexts (e.g., inflationary periods, yield-curve inversions), providing transparency and causal interpretability. The work demonstrates that macro-aware retrieval offers a principled path toward robust and explainable financial forecasting under distributional change, operationalizing the adage that 'financial history often rhymes.'