Retrieval-Augmented Large-Language-Model-Based Time-Series Forecasting for Cross-Market Equity Analysis

Time-series foundation models and retrieval-based augmentation have recently emerged as relevant tools for financial forecasting, yet it remains unclear when explicit historical retrieval helps a pre-trained probabilistic forecaster in noisy equity settings. This study evaluates that question through Cross-Market Retrieval-Augmented Lag-Llama (CM-RAF-Lag-Llama), a controlled framework for comparing Lag-Llama-only forecasts with retrieval-augmented forecasts across emerging and developed market panels. The main validation uses three balanced seven-asset panels, IDX7, US7, and JP7, producing a 48-configuration panel-context-horizon grid across four context lengths and four prediction horizons. The hybrid system retrieves analogous historical windows from an indexed memory and blends the retrieved continuation with the Lag-Llama forecast while preserving the same forecasting backbone. In the same-backbone return validation, at least one retrieval variant reduces Lag-Llama MSE in all 48 configurations, with aggregate lower-error retrieval MSE reductions of 28.85\% for IDX7, 27.27\% for US7, and 33.60\% for JP7. A second OHLCV-derived stock-feature experiment evaluates log-price, log-volume, range-volatility, and 20-day realized-volatility targets across 72 matched configurations. In that experiment, validation-selected RAF reduces MSE relative to Lag-Llama-only in 68 of 72 configurations and records the lowest MSE among the compared systems in 22 configurations, including 17 of 18 non-log-price US7 settings. A mechanism-level single-series ablation further compares Lag-Llama-only against Lag-Llama + RAF under zero-shot and scheduler-based inference, showing that retrieval effects are horizon-sensitive. The results indicate that retrieval can serve as an external-memory correction for Lag-Llama under selected equity targets, especially when volatility or volume structure provides reusable historical analogs.