TS-Haystack: A Multi-Scale Retrieval Benchmark for Time Series Language Models

TS-Haystack addresses a critical gap in the emerging field of Time Series Language Models (TSLMs): the lack of rigorous evaluation for long-context retrieval. While 'Needle In A Haystack' (NIAH) tests are standard for LLMs, time-series data presents unique challenges due to signal noise, multi-scale features, and the continuous nature of the data. The project's defensibility is currently low (score 3) because it is a research artifact (0 stars, 10 forks, 5 days old) rather than a software product with a moat; its value depends entirely on community adoption and becoming a 'standard' for future TSLM papers. Frontier labs like Google (TimesFM) and researchers building models like Lag-Llama are the primary audience. The medium frontier risk reflects that these labs could easily develop internal evaluation suites that supersede this if it doesn't gain rapid academic traction. The 10 forks vs. 0 stars suggest initial interest from a small group of researchers (likely collaborators or early reviewers) rather than organic growth. If adopted by major benchmarks like Hugging Face's leaderboards, its defensibility would shift toward network effects.