Self-Evolving Time-Series Forecasting Agents with Episodic Memory and Online Policy Learning

Junyi Wang; Yilin Wang; Wen Wu; Chao Zhang

cs.LGSep 26, 2026

Self-Evolving Time-Series Forecasting Agents with Episodic Memory and Online Policy Learning

Authors: Junyi Wang, Yilin Wang, Wen Wu, Chao Zhang

Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Tsinghua University

Abstract

LLM-based agents are increasingly used for time-series forecasting because they can organise contextual information, perform multi-step analysis, and guide the sequence of actions required to complete forecasting tasks. Most existing agents focus only on the current forecasting instance. However, in real-world deployments, forecasting commonly operates online, with new forecasts issued from the currently available history as the forecast origin advances and the ground-truth targets of earlier instances progressively become available. These targets provide feedback on the actions taken in earlier instances, yet existing agents generally do not preserve or utilise this information to adapt their subsequent actions. To address this limitation, we introduce FASE, a Feedback-Aware Self-Evolving forecasting agent that converts such feedback into task-specific experience for subsequent forecasting instances. FASE combines episodic memory, which retrieves relevant completed instances, with online policy learning, which summarises the feedback accumulated across instances into ranking guidance. The proposed framework is evaluated on 29 dataset configurations selected from the GIFT-Eval benchmark. Across these 29 configurations, FASE attains the strongest aggregate point forecasting performance among the evaluated methods and reduces the normalised MAE by 9.1% relative to the best individual foundation model baseline. The results further indicate that the cumulative advantage of FASE increases as delayed feedback accumulates. Together, these findings demonstrate that FASE can continually self-evolve through feedback from completed forecasting instances without updating the parameters of the LLM.

Figures & tables

Figure 1: Overview of FASE. For each forecasting instance, the LLM uses statistical context, retrieved experience, and policy guidance to invoke a forecasting tool. After the ground-truth target is revealed, FASE compares the candidate forecasts with the target and uses the resulting feedback to update episodic memory and periodically update the online policy for subsequent actions.

	MASE $\downarrow$		MAE $\downarrow$		RMSE $\downarrow$		CRPS $\downarrow$
Method	Norm.	Rank	Norm.	Rank	Norm.	Rank	Norm.	Rank
Seasonal Naive	1.000	7.79	1.000	7.72	1.000	7.86	1.000	7.97
Chronos-2 ( Ansari et al., 2025 )	0.752	4.55	0.729	5.03	0.720	4.59	0.545	4.90
TimesFM 2.5 ( Google Research, 2025 )	0.753	4.59	0.737	4.62	0.705	3.86	0.528	4.17
Moirai-2.0-R-Small ( Liu et al., 2025a )	0.774	5.97	0.739	5.21	0.756	6.24	0.542	5.38
TiRex-2 ( Podest et al., 2026 )	0.728	4.34	0.685	3.97	0.677	3.72	0.489	3.38

Table 1: Aggregate results over the 29 evaluated GIFT-Eval configurations. Bold and underlined values indicate the best and second-best results within each column, respectively.

	Long		Medium		Short
Method	MASE $\downarrow$	CRPS $\downarrow$	MASE $\downarrow$	CRPS $\downarrow$	MASE $\downarrow$	CRPS $\downarrow$
Seasonal Naive	1.000	1.000	1.000	1.000	1.000	1.000
Chronos-2 ( Ansari et al., 2025 )	0.782	0.547	0.782	0.563	0.729	0.538
TimesFM 2.5 ( Google Research, 2025 )	0.763	0.519	0.773	0.546	0.741	0.525
Moirai-2.0-R-Small ( Liu et al., 2025a )	0.823	0.560	0.831	0.567	0.733	0.525
TiRex-2 ( Podest et al., 2026 )	0.748	0.475	0.760	0.488	0.708	0.496

Table 2: Results aggregated by prediction length. Bold and underlined values indicate the best and second-best results within each column, respectively.

Figure 2: Cumulative relative MASE gain of FASE over the Base Agent across the first 500 forecasting instances in the replay. The curve begins at 25 processed instances, and the shaded band denotes the standard error across configurations.

Method	MASE $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$	CRPS $\downarrow$
Base Agent	0.758	0.718	0.713	0.543
Memory Only	0.683	0.631	0.655	0.483
Policy Only	0.687	0.629	0.649	0.505
FASE	0.679	0.622	0.643	0.497

Table 3: Ablation of episodic memory and online policy learning over the same 29 evaluated GIFT-Eval configurations. Base Agent uses neither mechanism; Memory Only uses episodic memory; Policy Only uses online policy learning; FASE combines both. Bold and underlined values indicate the best and second-best results within each column, respectively.

Figure 3: Distribution of forecasting performance relative to the Base Agent. Better and worse performance correspond to lower and higher MASE, respectively. Unchanged performance includes forecasting instances in which the evaluated method selects the same tool as the Base Agent or selects a different tool that yields the same MASE; the latter account for less than 0.3% of all evaluated instances. Results are averaged equally across configurations.

	MASE change relative to Base Agent (%)
Method	Improvement	Degradation	Net improvement
Memory Only	11.4	2.2	9.2
Policy Only	11.7	2.6	9.1
FASE	11.9	2.2	9.7

Table 4: MASE changes relative to the Base Agent. Improvement is the MASE reduction from forecasting instances with better performance. Degradation is defined analogously for instances with worse performance. Net improvement is their difference. Results are then averaged equally across configurations.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Domain	Frequency	Term	Pred. Length	Instances
bitbrains_fast_storage	Web/CloudOps	5T	medium	480	5,000
bitbrains_fast_storage	Web/CloudOps	5T	long	720	5,000
bitbrains_fast_storage	Web/CloudOps	H	short	48	5,000
bitbrains_rnd	Web/CloudOps	5T	medium	480	2,000
bitbrains_rnd	Web/CloudOps	5T	long	720	2,000
bitbrains_rnd	Web/CloudOps	H	short	48	2,000

Appendix

Table 5: GIFT-Eval configurations used for evaluation. Rows are grouped by source dataset. Prediction length is measured in time steps.

Component	Setting
LLM controller	Qwen3.8-27B-FP8; temperature $0.1$ ; top-p $1.0$ ; reasoning effort xhigh; sampling seed $1234$ .
Statistical context	18 features computed from up to the 15,360 most recent history positions.
Episodic memory	Recent pool capacity $100$ ; long-term pool capacity $900$ ; $10$ retrieved instances per action.
Policy input	Most recent $8H$ observations, normalised within each forecasting instance using the median and standard deviation, together with a binary observation mask.
Policy predictor	MLP with dimensions $2(8H)\rightarrow 2048\rightarrow 128\rightarrow 6$ and ReLU activations after the hidden layers.
Feedback pool	First-in, first-out pool with capacity $3{,}000$ ; uniform minibatch sampling.

Appendix

Table 6: Complete implementation settings used in the experiments.

Group	Statistic	Description
Data quality	Missing ratio	Fraction of missing target values.
Data quality	Zero ratio	Fraction of observed target values equal to zero.
Temporal dynamics	Global trend strength	Standardised slope over the full history.
Temporal dynamics	Recent trend strength	Standardised slope over the recent history.
Temporal dynamics	Recent level shift strength	Standardised shift between recent and prior means.
Temporal dynamics	Recent scale change strength	Normalised change between recent and prior scales.

Appendix

Table 7: Statistical features used to form the representation

s_{i}

Dataset	Setting	Chronos-2	TimesFM 2.5	Moirai-2.0- R-Small	TiRex-2	Toto-2.0- 313m	Uniform Ensemble	FASE
bitbrains_fast_storage	5T / medium	0.777	0.815	0.813	0.797	0.751	0.777	0.732
	5T / long	0.762	0.795	0.807	0.790	0.719	0.758	0.709
	H / short	0.754	0.818	0.822	0.823	0.717	0.774	0.723
bitbrains_rnd	5T / medium	0.959	0.969	0.969	0.967	0.949	0.958	0.948
	5T / long	0.933	0.948	0.952	0.945	0.926	0.934	0.921
	H / short	0.964	0.969	0.962	0.979	0.956	0.962	0.950

Appendix

Table 8: Normalised MASE for each evaluated GIFT-Eval configuration. Lower values indicate better point forecasting performance.

Dataset	Setting	Chronos-2	TimesFM 2.5	Moirai-2.0- R-Small	TiRex-2	Toto-2.0- 313m	Uniform Ensemble	FASE
bitbrains_fast_storage	5T / medium	0.519	0.634	0.579	0.497	0.479	0.514	0.607
	5T / long	0.636	0.706	0.685	0.584	0.500	0.593	0.666
	H / short	0.798	0.756	0.601	0.654	0.572	0.651	0.679
bitbrains_rnd	5T / medium	1.015	0.722	0.508	0.533	0.472	0.640	0.533
	5T / long	0.861	0.632	0.484	0.505	0.454	0.573	0.881
	H / short	0.830	0.490	0.538	0.480	0.472	0.546	0.548

Appendix

Table 9: Normalised CRPS for each evaluated GIFT-Eval configuration. Lower values indicate better probabilistic forecasting performance.

LLM-Based Time Series Forecasting

LLM-based time series forecasting uses large language models to predict future values from numerical histories and contextual data.

See all 100 papers

Sep 21, 2026

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time-Series Forecasting

Agentic time series forecasting concerns systems whose underlying mechanisms evolve, making the relative effectiveness of numerical models, reasoning strategies, and intervention rules inherently time-varying. Consequently, a time series agent must adapt the forecasts it produces and the orchestration policy that determines which components to trust and how to coordinate them. The deployment process naturally provides supervision for this adaptation as forecast horizons elapse and realized targets reveal the effectiveness of earlier decisions. Committing all numerical expert forecasts and candidate agent paths before target observation allows each realized outcome to evaluate the entire alternative set, providing delayed feedback without additional annotation. However, existing time series agents primarily incorporate prior experience through forecast refinement, reflection, or retrieval, without systematically converting realized outcomes into persistent updates to the joint orchestration policy governing later origins. To exploit this delayed feedback systematically, we introduce TimEvolve, a frozen-backbone time series agent that converts each realized outcome into persistent joint updates of expert trust, agent path selection, and intervention strength. A temporally ordered predict, reveal, and update protocol applies this feedback to subsequent forecasts. Experiments across eight Time-MMD domains show that TimEvolve achieves the best average MSE and MAE ranks among fifteen methods and the lowest errors on both metrics in seven domains. These results demonstrate the value of learning forecasting policies from the futures encountered during deployment.

Yifan Hu, Xilin Dai, Zhiyuan Qu +4

Ant International · Tsinghua University · The Chinese University of Hong Kong

Online Time Series Forecasting Time Series Forecasting Agentic Time Series Forecasting Self-Evolving Agents

Ask PDF

Aug 24, 2026

LLM-based Agents for Forecasting and Prediction: Methods, Training, Evaluation, and Applications

Large language models (LLMs) now support forecasting systems that combine language-based reasoning with temporal data, evidence retrieval, external tools, and iterative prediction. We investigate LLM-based forecasting agents, meaning systems in which a language model contributes to a scored prediction about a future or currently unobserved target. We organize architectures into three groups. Standalone LLM workflows operate on encoded time series or event context. Tool- and retrieval-augmented agents incorporate external evidence. Hybrid systems pair LLMs with statistical or foundation models. We then review training methods and evaluation protocols. We examine negative as well as positive evidence, including sensitivity to small input perturbations, ablations in which the LLM component does not improve accuracy, and benchmark gains that may reflect contamination instead of temporal reasoning. We cover applications in finance, weather, health, energy, and operations, and we summarize the benchmarks and datasets used for evaluation. The evidence indicates that measurement is a central limitation. Future work requires calibration under distribution shift, contamination-resistant live evaluation, explicit reporting of cost and accuracy together, and methods for handling feedback between deployed forecasts and the outcomes being forecast.

Xiaogang Xu, Jiaqi Tang, Jianmin Chen +12

Zhejiang University, China · The Hong Kong University of Science and Technology, Hong Kong · Northwestern Polytechnical University, China +4

LLM Agents LLM-Based Time Series Forecasting LLM Agent Evaluation LLM Agent Training Forecasting Benchmarks

Ask PDF

Jun 1, 2026

Bridging the Last Mile of Time Series Forecasting with LLM Agents

Time series forecasting has advanced rapidly, especially with the emergence of foundation models that show strong zero-shot performance on numerical extrapolation. However, in real-world forecasting settings, a statistically plausible baseline is rarely the final forecast used in practice. Before a forecast becomes decision-ready, it often needs to be revised using weakly structured business context such as holiday effects, campaign plans, external events, historical analogs, and expert feedback. This practical stage remains underexplored in the forecasting literature. In this paper, we formulate this stage as the last-mile forecasting problem and present an LLM-agent framework that sits on top of a forecasting backbone. Our system maintains a unified forecast workspace, invokes tools to retrieve contextual evidence, and converts reasoning trajectories into explicit forecast revision actions under structural safety constraints. It also supports long-horizon forecasting through map-reduce-style decomposition and post-hoc reflection through a memory bank. The resulting system is designed to be controllable and auditable. Through real-world case studies, we show how LLM agents can bridge the gap between statistical prediction and business-ready forecasting.

Yuhua Liao, Zetian Wang, Qiangqiang Nie +1

Trip.com Group, Shanghai, China

Time Series Forecasting LLM-Based Time Series Forecasting Agentic Time Series Forecasting

Ask PDF

Abstract

LLM-Based Time Series Forecasting

Explore similar work

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time-Series Forecasting

LLM-based Agents for Forecasting and Prediction: Methods, Training, Evaluation, and Applications

Bridging the Last Mile of Time Series Forecasting with LLM Agents