Training‑Free Beats 14B Model: Sonar‑TS Fills Scale Gap in Time‑Series QA

Introduction

Recent advances in large language models and time‑series foundation models have enabled natural‑language questioning of time‑series data, yet a fundamental query remains unsolved: given a database‑scale time‑series, how can one retrieve all segments that match a specific shape description, such as “days in the past year that first rose quickly and then entered a plateau”?

Problem Definition (NLQ4TSDB)

The authors formalize this gap as a new task, NLQ4TSDB (Natural Language Query for Time‑Series Databases), which requires three capabilities simultaneously: shape understanding (MP), handling database‑scale data (MS), and natural‑language grounding (NLG).

Benchmark (NLQTSBench)

To measure the task, they construct NLQTSBench, the first benchmark for NLQ4TSDB. Each query’s search space averages about 12 000 points, far larger than prior time‑series QA benchmarks that contain only a few hundred points. The benchmark comprises nine sub‑task categories (basic value retrieval, shape detection, periodicity, subsequence matching, composite trends, contextual anomalies, causal anomalies, and comprehensive insights) totaling 1 153 queries.

Method: Sonar‑TS (Search‑Then‑Verify)

Sonar‑TS follows a three‑stage neural‑symbolic pipeline, all without any training:

Offline indexing: Raw series are windowed at yearly, monthly, and daily scales. For each window, statistical primitives (e.g., slope, std) and SAX symbols are computed and stored in a SQL‑compatible feature table. Shapes become indexable symbols, e.g., “rapid rise then rapid fall” maps to the regular expression [ab]+.*[de]+.*[ab]+.

Online coarse search: An LLM parses the natural‑language question into sub‑steps, generates a SQL query that retrieves candidate windows from the symbol index, favoring high recall.

Fine verification: The LLM then emits Python code that applies mature operators such as DTW, PELT, and Theil‑Sen on the original signal to validate each candidate. Failed verifications trigger a fallback to raw data and a self‑correction loop.

The final answer is assembled from verified intervals, presented as natural language and visualizations.

Experimental Results

On the full‑scale setting, Sonar‑TS attains an overall score of 0.6144, roughly 3.8× the best Text‑to‑SQL baseline MAC‑SQL (0.16). In the short‑context (Lite) setting, the training‑free system scores 0.30, surpassing the 14B parameter model ChatTS‑14B (0.18). Ablation studies show that both the symbolic coarse search and the executable verification are indispensable.

However, even the strongest Sonar‑TS still scores low on the most shape‑intensive sub‑tasks, indicating that the problem remains far from solved.

Discussion and Limitations

Coarse search may miss candidates when SAX approximations are too rough; the authors provide a fallback to raw data.

Shape‑heavy tasks (e.g., composite trends, causal anomalies) still exhibit weak performance.

Domain adaptation may require recalibrating symbol granularity and verification operators for out‑of‑distribution data.

Online verification relies on large‑model code generation, incurring latency and cost that could be optimized.

Conclusion

Sonar‑TS demonstrates that a “search‑then‑verify” neural‑symbolic approach can reconcile scale and shape understanding in time‑series QA, establishing a new problem definition, a benchmark, and a viable baseline. While the method outperforms existing baselines, the results also reveal substantial room for improvement, especially in fine‑grained shape comprehension.