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

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

Problem Statement
The paper addresses the challenge of forecasting equity returns in noisy, partially nonstationary financial markets. While time-series foundation models like Lag-Llama offer reusable, probabilistic forecasting backbones capable of zero-shot generalization, it remains unclear whether explicit retrieval of analogous historical windows can consistently reduce forecast errors across different market regimes, assets, and horizons. Existing literature often focuses on task-specific models or single-market case studies, leaving a gap in understanding how retrieval-augmented mechanisms function as external-memory corrections for pre-trained foundation models in cross-market equity settings.

Methodology
The study introduces Cross-Market Retrieval-Augmented Lag-Llama (CM-RAF-Lag-Llama), a controlled framework that augments a fixed Lag-Llama probabilistic forecaster with a retrieval layer without retraining the backbone.

• Data and Panels: The evaluation utilizes three balanced seven-asset panels from emerging and developed markets: IDX7 (Indonesia), US7 (United States), and JP7 (Japan). Data consists of daily closing prices from Yahoo Finance (2019–2025), preprocessed with forward-filling, 5-day moving averages, and percentage return transformation.
• Hybrid Workflow:
  1. Base Forecast: Lag-Llama generates a probabilistic forecast based on a context window.
  2. Retrieval: Historical training windows are embedded using statistical descriptors (mean, standard deviation, extrema, quantiles, recent-window summaries) and indexed via FAISS.
  3. Similarity Search: The current context is queried against the memory bank using Euclidean distance to retrieve nearest-neighbor historical continuations.
  4. Fusion: The retrieved continuation signal is fused with the base forecast using a weighted average: $\hat{y}_H = (1 - \alpha)\hat{y}_{LL} + \alpha\hat{y}_R$.
• Experimental Design:
  • Cross-Market Grid: A full-grid evaluation across 48 configurations (3 panels × 4 context lengths [32, 64, 128, 256] × 4 prediction horizons [30, 60, 90, 120 days]).
  • Retrieval Variants: The study compares "Lag-Llama-only" against "RAF-raw" (fixed fusion) and "RAF-sel." (validation-selected fusion, where the optimal retrieval key and fusion scale are chosen based on validation windows).
  • Stock-Feature Extension: A secondary experiment evaluates OHLCV-derived targets (log-price, log-volume, range-volatility, 20-day realized volatility) across 72 configurations to test if richer equity representations improve retrieval efficacy.
  • Mechanism-Level Ablation: Detailed zero-shot and scheduler-based (learning rate 0.0005) evaluations on a single series (BBCA.JK) to isolate horizon and context-length sensitivities.
• Metrics: Performance is measured using Mean Squared Error (MSE) for point-forecast precision and Continuous Ranked Probability Score (CRPS) for probabilistic quality.

Key Results

• Cross-Market Validation: In the return-only experiment, at least one retrieval variant reduced MSE in all 48 configurations. Aggregate MSE reductions were 28.85% (IDX7), 27.27% (US7), and 33.60% (JP7) relative to Lag-Llama-only. The validation-selected retrieval variant consistently outperformed raw retrieval in 41 of 48 configurations.
• Stock-Feature Extension: For OHLCV-derived targets, validation-selected RAF reduced MSE relative to Lag-Llama-only in 68 of 72 configurations. It achieved the lowest MSE among all compared systems (including external baselines like Random Forest, KNN, and PatchTST) in 22 configurations, notably dominating in US7 volatility and volume targets (17 of 18 non-log-price settings).
• Mechanism-Level Insights:
  • Horizon Sensitivity: Retrieval was most effective at medium-to-long horizons (60–120 days). In zero-shot settings, 30-day forecasts often saw error increases, suggesting retrieval introduces noise when local dynamics dominate.
  • Context Length: Moderate context lengths (32 and 64) yielded the best results. Context length 256 often degraded performance, particularly for CRPS, implying that overly long windows dilute query representations with heterogeneous market regimes.
  • Adaptation: The scheduler-based setting stabilized performance, reducing MSE in 14 of 16 configurations compared to 12 of 16 in zero-shot, though zero-shot showed slightly higher average reductions when retrieval aligned well.
  • MSE vs. CRPS: Retrieval consistently improved point-forecast accuracy (MSE) more reliably than probabilistic calibration (CRPS), indicating the mechanism primarily corrects the forecast mean rather than reshaping the full uncertainty distribution.

Significance and Claims
The paper claims to provide controlled evidence that retrieval augmentation serves as an effective external-memory correction for foundation-model forecasting in equity markets. Its primary contributions are:

1. Validation of Retrieval as a Correction Layer: Demonstrating that explicit historical retrieval can reduce errors for a fixed, pre-trained Lag-Llama backbone without requiring full retraining or architectural changes.
2. Cross-Market Generalizability: Establishing that these error reductions are not artifacts of a single market but persist across emerging (IDX) and developed (US, JP) equity panels.
3. Configuration-Dependent Utility: Clarifying that retrieval is not universally beneficial; its value is contingent on forecast horizon (more useful for 60+ days) and context length (optimal at 32–64 steps).
4. Target-Specific Efficacy: Showing that the hybrid system is particularly competitive when forecasting targets with reusable volatility or volume structures (e.g., realized volatility) rather than raw price levels.

The authors position this work as a transparent, lightweight alternative to retraining large models, suitable for scenarios where computational resources are limited or where interpretable inference-time corrections are preferred over opaque model updates. They explicitly note that improved forecast error does not guarantee profitable trading strategies due to unmodeled transaction costs and market impact, framing the utility primarily as a decision-support tool for scenario analysis and risk planning.