A time-series shape query gets a new alphabet: TT-SAX adds a trend letter per segment and lifts recall and candidates together across six LLMs
Synopsis
When users query a time-series database in plain language, shape queries ask how a series moves, such as when a price plunged and then recovered quickly, yet existing symbolizations compress a series into letters that carry no direction. TT-SAX keeps the SAX level letters and adds one trend letter per segment, measured across neighboring segments and graded to match the adjectives queries use, giving two aligned letter tracks. On the 220 composite-trend queries of the public NLQTSBench benchmark, TT-SAX reaches a higher lift than six classical symbolizations under all six LLMs, at 2.42 without skills and 3.04 with them.
Interpretation
Once a time series is compressed into a letter string, the direction of movement is written as its own letter rather than only implied by the order of level letters. Level-only symbolizations such as SAX and ESAX write no direction, while 1d-SAX, TFSAX and TSAX write direction but measure it inside a single short segment, so noise inside a long trend breaks one sustained rise into a broken run of trend letters. On the 220 composite-trend queries of NLQTSBench, TT-SAX reaches a lift of 2.42 without skills and 3.04 with skills, while under the same setting the six classical symbolizations top out at 1.42 without skills and 2.16 with skills.
Trend letters are measured across neighboring segments, decoupling the measurement width from the level grid so the smoothing scale can follow the granularity at which natural language describes a shape. Measuring inside a segment lets noise dominate the change, while lengthening segments coarsens the level letters and erases short shapes, and the two could not previously be reconciled. In the ablation that widens the trend measurement step by step, TT-SAX recall first rises and then falls, and is best at two neighboring segments; removing the trend word drops lift to 1.00, while removing the level word leaves lift at 2.03.
The trend alphabet is designed for LLM readability, spelling steep fall, slow fall, stable, slow rise and steep rise as F, f, s, r and R, with case separating slow from steep. Letters in classical symbolizations are just names inside distance computations, where renaming changes nothing, so whether the alphabet itself helps retrieval had received little study. Across thirty alphabets and two LLMs, the group whose letters carry direction-word meaning is best with both models; rewriting FfsrR as the case-swapped fFsRr lowers recall, with an average loss of five points across the eight swapped pairs.
Perspective
Evaluation centers on the composite-trend subset of NLQTSBench, 220 queries over multi-domain real-world channels, plus a stress test on 1000 deduplicated queries synthesized with the benchmark's public generator; symbolic search is only the first stage of a search-then-verify pipeline, and returned candidate windows still need verification on raw values.
The best neighbor count is not universal and depends on the query distribution, the data and the symbolization's other parameters, so the experiments keep the benchmark's shared settings throughout and do not search this space; the two trend thresholds were fixed before the experiments, with a post-hoc sensitivity check over five other threshold pairs.
