TimeART: Towards Agentic Time Series Reasoning via Tool-Augmentation

Xingjian Wu; Junkai Lu; Zhengyu Li; Xiangfei Qiu; Jilin Hu; Chenjuan Guo; Christian S. Jensen; Bin Yang

cs.LGJan 20, 2026

TimeART: Towards Agentic Time Series Reasoning via Tool-Augmentation

Authors: Xingjian Wu, Junkai Lu, Zhengyu Li, Xiangfei Qiu, Jilin Hu, Chenjuan Guo, Christian S. Jensen, Bin Yang

Organizations: East China Normal University, Shanghai, China · Aalborg University, Aalborg, Denmark

Abstract

Time series data widely exist in real-world cyber-physical systems. Though analyzing and interpreting them contributes to significant values, e.g, disaster prediction and financial risk control, current workflows mainly rely on human data scientists, which requires significant labor costs and lacks automation. To tackle this, we introduce TimeART, a framework fusing the analytical capability of strong out-of-the-box tools and the reasoning capability of Large Language Models (LLMs), which serves as a fully agentic data scientist for Time Series Question Answering (TSQA). To teach the LLM-based Time Series Reasoning Models (TSRMs) strategic tool-use, we also collect a 100k expert trajectory corpus called TimeToolBench. To enhance TSRMs' generalization capability, we then devise a four-stage training strategy, which boosts TSRMs through learning from their own early experiences and self-reflections. Experimentally, we train an 8B TSRM on TimeToolBench and equip it with the TimeART framework, and it achieves consistent state-of-the-art performance on multiple TSQA tasks, which pioneers a novel approach towards agentic time series reasoning.

Figures & tables

Figure 1: Common dilemmas in time series reasoning, i.e., numerical hallucination and cognitive defect .

Figure 2: Common dilemmas in behaviour cloning and reinforcement learning, causing low generalization and entropy collapse .

Figure 3: Ratios of data sources in TimeToolBench (100k), the ReAct-style training corpus for agentic time series reasoning.

Figure 4: The overviews of the data pipeline for TimeToolBench (left), and the training pipeline for TimeART (right).

Model	Type	Stock Price Forecasting				Stock Indicator Forecasting				Temperature Forecasting
		7-day		30-day		7-day		30-day		7-day		14-day
		MAE	MAPE	MAE	MAPE	MACD	BB	MACD	BB	MSE	MAE	MSE	MAE
GPT-4o	closed	1.596	2.544	2.338	3.520	0.365	1.082	0.897	2.068	17.550	3.110	40.430	4.490
Gemini 3 Pro	closed	1.622	3.515	2.438	3.274	0.383	1.155	0.971	2.258	24.331	3.664	29.559	4.029
Claude Opus 4.5	closed	1.415	2.097	2.070	2.861	0.372	1.250	1.173	2.340	12.535	3.494	24.978	3.752
DeepSeek-V3.2	closed	1.713	2.144	2.127	3.299	0.353	1.192	1.070	2.190	29.274	4.058	101.655	6.592

Table 1: Results on Forecasting tasks in MTBench. Lower MSE, MAE, MAPE, MACD (MSE), and BB (MSE) values indicate better performance. Red : the best and Blue : the second best among open-source models ; closed-source models are reported as a reference and are not ranked. We also present comparisons with baselines equipped with TimeART’s framework in Table 4 .

Model	Type	Stock Trend Classification				News-stock Correlation				News-finance MCQA		Temperature Trend Classification
		7-day		30-day		7-day		30-day		7-day	30-day	Past	Future
		3-way	5-way	3-way	5-way	3-way	5-way	3-way	5-way	Acc	Acc	Acc	Acc
GPT-4o	closed	42.81	36.45	47.35	30.58	53.60	31.00	57.60	34.60	65.10	52.80	66.36	43.54
Gemini 3 Pro	closed	47.11	41.64	44.78	29.71	51.85	26.39	59.43	34.82	63.38	50.29	56.91	51.54
Claude Opus 4.5	closed	44.85	33.48	52.00	31.60	50.39	29.01	57.99	34.21	75.59	61.15	59.77	57.12
DeepSeek-V3.2	closed	45.05	35.44	48.41	29.59	49.98	27.01	57.53	35.06	77.47	69.50	26.56	25.15

Table 2: Results on Reasoning tasks in MTBench. Higher accuracy rates (Acc, 3-way, 5-way) indicate better performance. Red : the best and Blue : the second best among open-source models ; closed-source models are reported as a reference and are not ranked. We also present comparisons with baselines equipped with TimeART’s framework in Table 5 .

Panel of Figure 5 — Figure 5: Ablations on the TimeART framework.

Panel of Figure 6 — Figure 6: Ablations on training stages.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Type	Understanding	Perception	Reasoning	Estimation
GPT-4o	closed	50.86	69.65	50.00	66.66
Qwen3-max	closed	53.45	58.27	46.25	49.33
DeepSeek-R1 7B	open	7.76	12.69	16.89	7.41
Llama3 8B	open	33.62	36.27	27.03	25.92
ChatTS 7B	open	25.00	28.68	38.51	12.03
ITFormer 7B	open	40.52	45.24	30.40	44.44

Appendix

Table 3: Accuracy rates on tasks in TimeMQA. Higher values indicate better performance. Red : the best and Blue : the second best among open-source models ; closed-source models are reported as a reference and are not ranked.

Model	Type	Stock Price Forecasting				Stock Indicator Forecasting				Temperature Forecasting
		7-day		30-day		7-day		30-day		7-day		14-day
		MAE	MAPE	MAE	MAPE	MACD	BB	MACD	BB	MSE	MAE	MSE	MAE
GPT-4o	closed	0.805	1.720	1.142	3.080	0.352	1.052	0.870	2.006	4.145	1.665	5.170	1.835
GPT-5.2	closed	0.802	1.650	1.108	3.070	0.350	1.048	0.868	2.004	4.125	1.658	5.145	1.822
Gemini 3 Pro	closed	0.795	1.616	1.133	3.066	0.3508	1.046	0.8654	1.994	4.074	1.652	5.052	1.757
Qwen3-max	closed	0.810	1.780	1.150	3.090	0.345	1.060	0.872	2.018	4.110	1.665	5.150	1.850

Appendix

Table 4: Results on Forecasting tasks in MTBench. Note that we equip all the baselines with TimeART’s tool-use framework, and they are prompted with few-shot examples. Lower MSE, MAE, MAPE, MACD (MSE), and BB (MSE) values indicate better performance. Red : the best and Blue : the second best among open-source models ; closed-source models are reported as a reference and are not ranked.

Model	Type	Stock Trend Classification				News-stock Correlation				News-finance MCQA		Temperature Trend Classification
		7-day		30-day		7-day		30-day		7-day	30-day	Past	Future
		3-way	5-way	3-way	5-way	3-way	5-way	3-way	5-way	Acc	Acc	Acc	Acc
GPT-4o	closed	49.85	42.76	53.40	38.15	68.45	40.62	66.84	38.92	86.75	75.40	73.85	55.92
GPT-5.2	closed	52.38	45.62	54.86	40.72	70.82	41.95	68.74	40.18	90.65	81.25	75.90	58.10
Gemini 3 Pro	closed	51.80	44.85	54.45	39.86	70.35	41.56	68.28	39.84	89.72	79.85	75.35	57.62
Qwen3-max	closed	50.76	43.96	53.92	39.20	69.88	41.20	67.95	39.50	89.10	79.20	75.10	57.30

Appendix

Table 5: Results on Reasoning tasks in MTBench. Note that we equip all the baselines with TimeART’s tool-use framework, and they are prompted with few-shot examples. Higher accuracy rates (Acc, 3-way, 5-way) indicate better performance. Red : the best and Blue : the second best among open-source models ; closed-source models are reported as a reference and are not ranked.

Tool-use policy	TimeMQA Acc	MTBench Reason. Acc	MTBench Forec. MAE
Random selection	28.40	35.24	3.09
Fixed heuristic policy	38.70	46.54	1.93
Qwen3 8B + tools (zero-shot)	40.31	43.84	2.15
Qwen3 8B + tools (few-shot)	41.20	51.60	1.86
TimeART (trained)	57.05	59.87	1.33

Appendix

Table 6: Attribution of improvements. All variants use the identical 21-tool framework and differ only in the tool-use policy. Higher accuracy and lower MAE indicate better performance. Red : the best, Blue : the second best.

Model	TimeMQA Acc	MTBench Reason. Acc
Qwen3 8B + tools (few-shot)	41.20	51.60
Qwen3.5 9B + tools (zero-shot)	43.50	53.24
Qwen3.5 9B + tools (few-shot)	46.80	55.44
TimeART (Qwen3 8B, trained)	57.05	59.87

Appendix

Table 7: A newer backbone with stronger agentic pretraining narrows but does not close the gap to a trained model. All variants share the 21-tool framework. Red : the best, Blue : the second best.

# Tools	Qwen3 8B (untrained)	TimeART	Gap
5 (core)	43.48	48.50	+5.02
10	42.08	53.70	+11.62
15	39.78	55.20	+15.42
21 (full)	40.31	57.05	+16.74

Appendix

Table 8: TimeMQA accuracy as a function of the number of available tools. More tools hurt the untrained model but help the trained one. Red : the best per column.

Criterion	Qwen3 8B + tools (few-shot)	TimeART
Correct tool selection	62.3%	84.7%
Correct parameter configuration	54.8%	79.2%
Correct output interpretation	58.1%	82.5%
Inter-annotator agreement ( $\kappa$ )	0.71	0.76

Appendix

Table 9: Human evaluation of tool-use quality on 200 sampled evaluation trajectories, judged per tool invocation by two independent annotators. Red : the best.

Metric	Seed 1	Seed 2	Seed 3	Mean $\pm$ Std
TimeMQA Acc	57.05	56.35	57.75	57.05 $\pm$ 0.70
MTBench Reasoning Acc	59.87	58.97	60.27	59.70 $\pm$ 0.67
MTBench Forecast MAE	1.33	1.38	1.30	1.34 $\pm$ 0.04

Appendix

Table 10: TimeART retrained with three independent seeds. Aggregates follow the definitions of Appendix A.3 .

Panel of Figure 7 — Figure 7: Analyses on rationale $\mathrm{C}$ .

Hyperparameter	Assignment
Base model	Qwen3 8B
Computing infrastructure	8*3090-24GB GPU
Epochs	5
Warm-up ratio	0.02
Batch size per device	4
Gradient accumulation steps	16

Appendix

Table 11: Hyper-parameters and training time of fine-tuning the Qwen3 8B based on TimeToolBench.

Evaluation set	Trajectories in TimeToolBench	Exclusion method
MTBench (all tasks)	22k	entirely removed
TimeMQA (four subsets)	2.6k	held out by domain
Remaining training set	75.4k	non-overlapping domains

Appendix

Table 12: Separation between TimeToolBench and the evaluation benchmarks. No raw series, question template, or answer label from an evaluation set appears in training.

Category	Associated Tools
Forecasting	forecaster
Anomaly Detection	anomaly_detection , spike_detector
Classification	trend_classifier , noise_profile , stationarity_test
Seasonality	seasonality_detector , autocorr
Trend	trend_classifier , rolling_stat , summary_stats
Stationarity	stationarity_test

Appendix

Table 13: Categorization of TimeART Tools based on Analytical Tasks

Analysis step	Tools invoked by TimeART	Manual counterpart
Correlation shift detection	channel_correlation ,	correlation matrix
	change_point_detector	inspection
Return forecasting	forecaster (7-day horizon)	judgement + spreadsheet
Risk identification	volatility , spike_detector	visual chart inspection

Appendix

Table 14: Steps taken autonomously by TimeART on the financial portfolio risk monitoring case, and their manual counterparts.

Figure 11: Failure Case c.

Figure 12: Failure Case d.

Time Series Reasoning

Time series reasoning studies how models interpret temporal patterns, answer questions about sequences, and use evidence to make forecasts.

See all 44 papers

T-SMART: Mechanism-Level Attribution for Tool-Augmented Time-Series Question Answering
Sep 12, 2026Ivan Delgado, Himansi Gupta, Bishal Khatri +4Time Series QA Time Series Reasoning
University of California, San Diego, CA, USA · West Virginia University, Morgantown, WV, USA
TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale
Apr 11, 2026Malgorzata Gwiazda, Yifu Cai, Mononito Goswami +2Time Series Reasoning LLM Evaluation
Technical University of Munich, Munich, Germany · Auton Lab, School of Computer Science, Carnegie Mellon University, Pittsburgh, US
TimeSage-MT: A Multi-Turn Benchmark for Evaluating Agentic Time Series Reasoning
May 31, 2026Yaxuan Kong, Qingren Yao, Yuqi Nie +7Time Series Reasoning LLM Agent Evaluation
University of Oxford · VulpiVox Intelligence · Eindhoven University of Technology +3

Sep 12, 2026J/K move · Enter open · S save

T-SMART: Mechanism-Level Attribution for Tool-Augmented Time-Series Question Answering

Ivan Delgado, Himansi Gupta, Bishal Khatri, Niharika Sapre+3

University of California, San Diego, CA, USA · West Virginia University, Morgantown, WV, USA

Time Series QA Time Series Reasoning Numerical Reasoning in Language Models Neuro-Symbolic Reasoning

Open paper AskPDF arXiv

Large language models (LLMs) can struggle with time-series question answering (TS-QA), especially when numerical signals are serialized as text and require explicit computation. Tool-augmented approaches improve performance, but existing systems often intertwine language reasoning, computation, and perception, making it difficult to determine which components drive the gains. We present T-SMART, a neurosymbolic framework that separates these roles: a frozen LLM interprets questions and selects operations, deterministic tools perform numerical computation, and structured perception is invoked only when needed. Controlled paired ablations show that deterministic computation provides the dominant benefit, improving accuracy by 31.7 percentage points over direct LLM reasoning on serialized time series, while language understanding and perception offer smaller complementary gains. These results indicate that tool-augmented TS-QA benefits primarily from reliable numerical execution rather than additional language-model reasoning and provide a controlled framework for analyzing component contributions in neurosymbolic time-series systems.

Abstract

Time Series Reasoning

Explore similar work

T-SMART: Mechanism-Level Attribution for Tool-Augmented Time-Series Question Answering

TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale

TimeSage-MT: A Multi-Turn Benchmark for Evaluating Agentic Time Series Reasoning