Time series forecasting models often rely on historical patterns, but traditional retrieval methods that use only numerical similarity struggle when data conditions change—a problem known as non-stationarity. A new research paper introduces SERAF (Semantics-Enhanced Retrieval-Augmented Time Series Forecasting), a multimodal framework that retrieves both historical time series segments and their self-generated textual descriptions to guide future predictions.
The Challenge of Non-Stationarity
According to the paper by Zhou, Shiqiao, Wu, Zipeng, Schöner, Holger, Fouché, Edouard, Wilson, IAG, and Wang, Shuo, mainstream retrieval-augmented models for time series rely solely on time series similarity. Under non-stationarity—where the statistical properties of data shift over time—this one-dimensional approach often proves insufficient. The authors argue that incorporating semantic information can help bridge the gap between numerical patterns and the underlying context.
How SERAF Works
SERAF performs dual retrieval: it searches over both the raw time series data and their automatically generated textual descriptions. The framework retrieves two complementary sets of historical patterns and their corresponding future outcomes. These are then selectively and jointly used to enhance forecasting. The key innovation is that the textual descriptions provide a semantic view that can capture context not evident in the numeric sequence alone, such as seasonality shifts or external events.
"Unlike mainstream approaches that depend only on time series similarity, SERAF conducts dual retrieval over the time series and their self-generated textual descriptions."
The system uses a multimodal approach to combine numerical and semantic views, allowing it to better handle non-stationary environments where past patterns may not directly repeat numerically but share semantic similarities.
Experimental Validation
The researchers tested SERAF across seven real-world datasets and compared its performance against state-of-the-art baselines. The results demonstrated that SERAF effectively bridges numerical and semantic views, leading to improved forecasting accuracy. While the paper does not specify the domains of these datasets, the broad evaluation suggests the framework is generalizable.
Potential Enterprise Applications
Although the paper is focused on methodology, its implications for supply chain and logistics forecasting are clear. Many enterprise forecasting tasks—such as demand planning, inventory optimization, and freight volume prediction—face non-stationarity due to seasonality, market shifts, or disruptions. SERAF's ability to incorporate semantic context (e.g., textual descriptions of past events) alongside raw numeric data could make forecasts more robust. Enterprises using time series models in their planning systems may benefit from similar retrieval-augmented architectures that combine numeric and textual information.
| Feature | Description |
|---|---|
| Dual Retrieval | Searches over time series segments and self-generated text descriptions |
| Two Complementary Sets | Retrieves patterns and corresponding futures from both modalities |
| Selective Joint Use | Combines retrieved information to guide predictions |
| Evaluation | Tested on 7 real-world datasets, outperforming state-of-the-art baselines |
The framework is currently a research artifact, but its principles could be integrated into commercial forecasting platforms. As enterprise data increasingly includes unstructured text (notes, reports, logs), models that can fuse such information with numeric time series will become more valuable.
For CTOs and digital transformation leaders, SERAF represents a step toward more intelligent forecasting systems that understand not just what happened, but why it happened—and use that understanding to predict the future.