EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution

Kaipeng Xu; Xianli Yan; Yan Wang; Xiang Liu; Shan Liu

cs.AIOct 3, 2026

EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution

Authors: Kaipeng Xu, Xianli Yan, Yan Wang, Xiang Liu, Shan Liu

Organizations: School of Automation and Intelligent Science, Jiangnan University, Wuxi 214122, China · School of Automation, Southeast University, Nanjing 210096, China

Abstract

Deep time-series forecasting models have rapidly diversified, yet adapting them to a specific task still requires extensive expert effort in model selection, mechanism diagnosis, architecture design, implementation, and evaluation. Existing AutoML methods are constrained by predefined search spaces, while general-purpose LLM research agents lack reliable control over experimental protocols and model promotion. We introduce EvoCast, a fully autonomous research-agent system for iterative forecasting architecture evolution. EvoCast first establishes and diagnoses a task-specific baseline through executed mechanism ablations, then generates evidence-grounded research directions from dataset characteristics, diagnostic results, prior rounds, and failure records. Its central design, cognition-authority separation, assigns open-ended hypothesis generation and code implementation to LLM agents, while deterministic program authorities control source-edit boundaries, canonical evaluation, and promotion decisions. Experimental outcomes are accumulated as evidence to guide subsequent rounds. Results show that EvoCast completes complex architecture modifications with higher implementation success and lower agent-side token/time cost, and develops task-specific architectures that outperform selected baselines, strong forecasting models, and agent baselines in three real-world forecasting cases. The code is available at https://github.com/18e0-x/EvoCast.

Figures & tables

Figure 1. EvoCast in the landscape of automated forecasting model development. Compared with fixed-space AutoML/NAS and open-ended LLM research agents, EvoCast combines evidence construction, bounded cognitive-agent exploration, and program-authority verification in a fixed-protocol loop for reliable architecture evolution. A three-part comparison of fixed-space AutoML and neural architecture search, open-ended LLM research agents, and EvoCast. EvoCast links task evidence, bounded cognitive-agent exploration, program-authority verification, canonical evaluation, and evidence-state updates in a closed loop.

Figure 2. Complete EvoCast workflow. The loop fixes the task protocol, establishes and diagnoses a task-specific baseline, executes formal research rounds under bounded implementation and canonical evaluation, updates the evidence-carrying research state, and produces a final HTML report. Workflow diagram with five stages: task grounding, baseline establishment, baseline diagnosis, formal research rounds, and final reporting. Research directions pass through a build contract, an isolated agent workspace, source audit, canonical evaluation, and a promotion gate before their outcomes update the evidence state.

Figure 3. Cognitive roles and program authority chain in one controlled research round. Cognitive agents generate diagnosis targets, research directions, candidate source states, and final narrative, while program authorities enforce source boundaries, canonical evaluation, promotion decisions, and state updates. Diagram of a controlled research round. Cognitive agents create research content and source edits. Program authorities audit the source boundary, materialize the candidate, run canonical evaluation, make promotion decisions, and update persistent state.

Method	Success	Fidelity	Repair	Avg. Token	Avg. Cost (USD)	Avg. Time (s)
Direct Independent Edit	33.3% (10/30)	90.0%	—	4,975	0.0037	31
mini-SWE-agent (2024)	90.0% (27/30)	92.6%	90.0% (27/30)	187,818	0.0206	573
AIDE (2025)	83.3% (25/30)	96.0%	54.5% (6/11)	38,483	0.0356	283
R&D-Agent (2025)	46.7% (14/30)	100.0%	42.9% (12/28)	60,932	0.0313	266
AutoResearch (2026)	86.7% (26/30)	96.2%	50.0% (4/8)	38,780	0.0327	258
EvoCast	96.7% (29/30)	100.0%	95.7% (22/23)	27,618	0.0137	153

Table 1. Overall method-level RQ1 results. Success is the proportion of tasks that pass the canonical validation pipeline; Fidelity is the proportion of evaluated implementations that realize the given idea without semantic mismatch; Repair is computed only over repair-eligible tasks. Higher is better for Success, Fidelity, and Repair; lower is better for Avg. Token, Avg. Cost, and Avg. Time. Best values are bolded and second-best values are underlined .

Task	Mode	Selected backbone	Lookback $\rightarrow$ Horizon	Forecasting characteristics
Air Quality	MM	FreTS	$168\rightarrow 24$	Cross-variable dependency, seasonal structure, and trend variation across 12 air-quality, sensor-response, and weather variables.
Melbourne Pedestrian T33	SS	PatchMLP	$168\rightarrow 24$	Hourly pedestrian counts with strong daily and weekly periodicity and long-term level change.
Steel Industry	MS	PatchMLP	$720\rightarrow 120$	Industrial power forecasting from 9 input variables to Usage_kWh, with a long lookback and horizon, regime transitions, seasonal structure, and distribution drift.

Table 2. RQ2 task protocols and forecasting characteristics. The Lookback

\rightarrow

Horizon column gives each task’s history and prediction windows. All tasks use chronological 7:1:2 train/validation/test splits.

Method	Air Quality (MM)		Melbourne Pedestrian T33 (SS)		Steel Industry (MS)
	MSE	MAE	MSE	MAE	MSE	MAE
Sundial (2025)	0.5481	0.5194	0.1814	0.2253	0.4413	0.3997
Toto-2.0-2.5B (2026)	0.4968	0.5061	0.2583	0.2643	0.4880	0.3702
Time-VLM (2025)	0.4737	0.4988	0.2570	0.3067	0.3565	0.4057
TimeCMA (2025)	0.5134	0.5240	1.2620	0.9147	0.3037	0.3658
AIDE (2025)	0.5309	0.5327	0.1728	0.2070	0.4112	0.4780

Table 3. RQ2 overall forecasting results on three real tasks. Best values are bolded and second-best values are underlined . Numbers in parentheses for EvoCast denote relative error reduction against the corresponding selected baseline.

Task	Method	Valid rounds $\uparrow$	Active time (s) $\downarrow$	Calls $\downarrow$	Tokens $\downarrow$	Cost (USD) $\downarrow$
Air Quality	AIDE	60% (12/20)	3,436	60	872,050	0.4915
	R&D-Agent	65% (13/20)	29,292	256	7,301,193	3.7618
	EvoCast	100% (20/20)	7,989	96	1,210,702	0.6089
Melbourne T33	AIDE	65% (13/20)	11,221	60	1,127,775	1.0902
	R&D-Agent	75% (15/20)	14,425	148	3,445,604	2.4359
	EvoCast	90% (18/20)	16,625	96	1,485,592	0.7798

Table 4. RQ2 valid formal rounds and end-to-end resource ledger. Valid rounds show the ratio and count over 20. Active time is cumulative workflow wall-clock time; Calls and Tokens count all LLM calls and input/output tokens; Cost is total API expenditure. Bold indicates the highest valid-round ratio for each task.

Figure 4. Task-adaptive final live-path architectures discovered by EvoCast on the three RQ2 real forecasting tasks. Gray modules are original backbone components, and colored modules are active modules accumulated during EvoCast research rounds. Three model diagrams for Air Quality, Melbourne Pedestrian T33, and Steel Industry. Gray blocks represent inherited backbone components, while colored blocks represent task-specific modules retained after accepted EvoCast research rounds.

Variant	Removed component	Final MSE	Valid rounds	Time (h)	Tokens (M)	Cost (USD)
Full	None	0.162579	18 / 20	4.62	1.49	0.7798
A1	Dataset characterization	0.173547	17 / 20	7.01	1.42	0.6879
A2	Baseline diagnosis	0.173473	15 / 20	6.80	2.79	1.4694
A3	Evidence-grounded planning	0.173250	20 / 20	5.07	1.88	0.9793
A4	Authority-controlled execution	0.185001	14 / 20	10.80	25.32	7.8792
A5	Iterative evidence update	0.178489	15 / 20	2.23	1.81	0.8145

Table 5. Overall RQ3 system-ablation results. Final MSE is the three-seed mean over seeds 2026, 2027, and 2028. Lower is better for MSE, time, tokens, and cost; higher is better for valid rounds.

Figure 5. RQ3 cumulative best trajectories (event-only). The plot records only events that refresh the best-so-far result, showing when each ablation chain finds a new best candidate and whether later rounds continue to improve. Line chart comparing event-only cumulative-best trajectories for the full EvoCast system and five ablations. The full system reaches a stronger error range earlier and records later best-result updates, while several ablations plateau sooner.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

System	Starting point	Modification target	Evidence for next step	Promotion authority
TimeSeriesScientist ( Zhao et al., 2025 )	forecasting workflow and model pool	workflow and model configuration	workflow observations and validation results	workflow-level selection
SEA-TS ( Xu et al., 2026 )	searched forecasting code state	generated forecasting code	search feedback, review results, and execution outcomes	execution-result-based selection
GenAutoML ( Poikarayil et al., 2026 )	generated architecture scaffold	task-specific model architecture	runtime checks and repair feedback	runtime-validation-based selection
EvoCast	task-specific repository baseline	repository source modules	executed baseline ablations, prior-round evidence, and failure memory	deterministic promotion gate outside the LLM

Appendix

Table A.1. Comparison of representative time-series agent systems.

Ablation result	Diagnostic label	Planning consequence
Clear degradation	essential	Preserve the core function and modify cautiously
Clear improvement	harmful or redundant	Prioritize replacement or restructuring
Small change	weak or uncertain	Keep it as an exploration target
Build/runtime failure	engineering failure	Do not draw a scientific conclusion
Protocol violation	invalid	Discard or rerun

Appendix

Table B.1. Diagnostic semantics of ablation outcomes.

Evidence group	Examples	Planning use
Task evidence	data features, horizon, variable count	identify task needs
Model evidence	baseline, current best, source structure	determine the starting point
Scientific evidence	ablations, positive results, negative results	update mechanism judgment
Engineering evidence	import, shape, training, repair history	identify implementation risks

Appendix

Table B.2. Planner evidence bundle.

Outcome	Status	Evidence	Failure memory	Update best	Later effect
Accepted candidate	confirmed positive	$\checkmark$	$\times$	$\checkmark$	next-round starting point
Valid rejection	confirmed negative	$\checkmark$	$\times$	$\times$	avoid the same mechanism
Unstable improvement	weak positive	$\checkmark$	$\times$	$\times$	weak clue for later rounds
Engineering failure	no scientific conclusion	$\times$	$\checkmark$	$\times$	constrain implementation paths

Appendix

Table B.3. Round outcome semantics and state updates.

Item	Value
Model	MiniMax-M3
Service tier	standard
Context cap	$\leq 512$ K input tokens
Uncached input price	0.30 USD / M input tokens
Cache-hit input price	0.06 USD / M input tokens
Output price	1.20 USD / M output tokens

Appendix

Table C.1. MiniMax-M3 API pricing used in our experiments.

Workflow	Pre-formal stage	Formal cap	Repair cap	Resource-accounting scope
AIDE	10-round draft stage	20 rounds	3	From draft-stage start through workflow termination.
R&D-Agent	Workflow initialization	20 rounds	3	From workflow start through termination.
EvoCast	Baseline establishment and mechanism diagnosis	20 rounds	3	From baseline-establishment start through workflow termination.

Appendix

Table C.2. Formal-improvement budget and end-to-end resource-accounting boundary.

Category	Count	Models
linear	7	CMoS ( Si et al., 2025 ) , CrossLinear ( Zhou et al., 2025a ) , DLinear ( Zeng et al., 2023 ) , Linear ( Zeng et al., 2023 ) , NLinear ( Zeng et al., 2023 ) , SparseTSF ( Lin et al., 2024 ) , TimeBase ( Huang et al., 2025a )
transformer	21	Autoformer ( Wu et al., 2021 ) , Crossformer ( Zhang and Yan, 2023 ) , DTAF ( Lu et al., 2026 ) , DUET ( Qiu et al., 2025 ) , ETSformer ( Woo et al., 2022 ) , FEDformer ( Zhou et al., 2022b ) , Informer ( Zhou et al., 2021 ) , MultiPatchFormer ( Naghashi et al., 2025 ) , Nonstationary_Transformer ( Liu et al., 2022c ) , PAttn ( Tan et al., 2024 ) , PDF ( Dai et al., 2024 ) , PatchTST ( Nie et al., 2023 ) , Pathformer ( Chen et al., 2024 ) , Pyraformer ( Liu et al., 2022b ) , Reformer ( Kitaev et al., 2020 ) , TemporalFusionTransformer ( Lim et al., 2021 ) , TimeBridge ( Liu et al., 2025b ) , TimePerceiver ( Lee and Lee, 2025 ) , Transformer ( Vaswani et al., 2017 ) , Triformer ( Cirstea et al., 2022 ) , iTransformer ( Liu et al., 2024 )
mlp	8	FreTS ( Yi et al., 2023 ) , HDMixer ( Huang et al., 2024b ) , LightTS ( Zhang et al., 2022 ) , PatchMLP ( Tang and Zhang, 2025 ) , TSMixer ( Chen et al., 2023 ) , TiDE ( Das et al., 2023 ) , TimeMixer ( Wang et al., 2024a ) , WPMixer ( Murad et al., 2025 )
cnn	4	MICN ( Wang et al., 2023 ) , ModernTCN ( Luo and Wang, 2024 ) , SCINet ( Liu et al., 2022a ) , TimesNet ( Wu et al., 2023 )
gnn	2	MSGNet ( Wen et al., 2024 ) , TimeFilter ( Hu et al., 2025 )
others	10	Amplifier ( Fei et al., 2025 ) , FITS ( Xu et al., 2024 ) , FiLM ( Zhou et al., 2022a ) , KANAD ( Zhou et al., 2025b ) , Koopa ( Liu et al., 2023 ) , MambaSimple ( Gu and Dao, 2023 ) , SRSNet ( Wu et al., 2025 ) , SegRNN ( Lin et al., 2026 ) , TimeKAN ( Huang et al., 2025b ) , xPatch ( Stitsyuk and Choi, 2025 )

Appendix

Table C.3. Supported baseline model pool grouped by architecture family.

Level	Definition	Typical forms	Main capability under test
Level 1: Local Modification	Modify an existing module locally without changing the surrounding execution structure.	Replace an activation; add a residual gate; adjust the normalization position; insert a lightweight channel mixer.	Small and precise local code edits inside an existing deep-learning block.
Level 2: Cross-Module Integration	Understand multiple source regions and connect existing modules through a new live path.	Add adaptive fusion between decomposition and the forecasting head; inject a frequency branch into the backbone; align multi-scale paths and then fuse them.	Cross-file and cross-module implementation ability, including interface matching and shape alignment.
Level 3: Novel Module Insertion	Add a new module or reorganize the main execution path so that the new mechanism participates in prediction.	Add a context-conditioned spectral gate; add a horizon-adaptive branch; add a variable interaction router.	New-module implementation and main-path restructuring under executable repository constraints.

Appendix

Table D.1. The three predefined idea-complexity levels used in RQ1.

Figure D.1. Success-rate trends across idea-complexity levels. Line chart of implementation success rates for several agent methods across low, medium, and high idea-complexity levels. EvoCast maintains the strongest success trend as complexity increases.

Dataset	Mode	Rows	Setting	Variables	Task characteristics
Air Quality ( Vito et al., 2008 )	MM	9,357	168 $\rightarrow$ 24	12 air-quality, sensor-response, and weather variables	Multivariate air-quality forecasting with cross-variable dependency, seasonal structure, and trend variation.
Melbourne Pedestrian T33 ( Godahewa et al., 2020 )	SS	47,315	168 $\rightarrow$ 24	Univariate hourly pedestrian count	Single-series pedestrian-flow forecasting with strong daily and weekly periodicity and long-term level change.
Steel Industry ( Sathishkumar et al., 2021 )	MS	35,040	720 $\rightarrow$ 120	9 input variables and target Usage_kWh	Multivariate-to-univariate industrial power forecasting with long lookback, long horizon, regime transition, seasonal structure, and distribution drift.

Appendix

Table D.2. RQ2 tasks and dataset characteristics.

Method	Air Quality		Melbourne Pedestrian T33		Steel Industry
	MSE	MAE	MSE	MAE	MSE	MAE
DLinear (2023)	0.5160	0.5344	0.2655	0.3149	0.3268	0.3722
NLinear (2023)	0.6004	0.5692	0.3188	0.3737	0.3476	0.3975
FreTS (2023)	0.4665	0.5004	0.1844	0.2446	0.3285	0.3847
FITS (2024)	0.6672	0.6151	0.4179	0.4360	0.3572	0.3966
HDMixer (2024)	0.4753	0.4950	0.2271	0.2663	0.3318	0.3395

Appendix

Table D.3. EvoCast versus general-purpose SOTA forecasting models on RQ2 tasks. Best values are bolded and second-best values are underlined .

Figure D.2. Round-level internal trajectories of agent workflows on Air Quality. EvoCast stays in a lower and more stable error band, while AIDE (2025) and R&D-Agent show larger fluctuations and more rounds without usable metrics. Round-by-round Air Quality error trajectories for EvoCast, AIDE, and R and D-Agent. EvoCast remains in a lower and more stable error band, while the other methods fluctuate and include missing usable metrics.

Figure D.3. Baseline and round-level trajectories on the three RQ2 real forecasting tasks. Red stars denote accepted rounds, and error bars show available seed-level stability information. Three task-level plots showing selected baselines and round-level forecast errors. Red stars mark accepted research rounds and error bars indicate available seed-level stability information.

Figure E.1. EvoCast terminal experiment console. The interface unifies live progress tracking, round-level status, metrics, elapsed time, token-aware execution context, and raw-log inspection in one view. Screenshot of the EvoCast terminal console. It shows task and stage status, the current best result, baseline and research-round summaries, elapsed time, token accounting, and live raw logs.

Explore similar work

SEA-TS: Self-Evolving Agent for Autonomous Code Generation of Time Series Forecasting Algorithms
Mar 5, 2026Longkun Xu, Xiaochun Zhang, Qiantu Tuo +1Time Series Forecasting Code Generation
Ecoflow Inc.
A Human-in-the-Loop Autonomous Agent for Industry Time Series Forecasting
Aug 31, 2026Xiaoyu Tao, Mingyue Cheng, Ze Guo +4Electricity Price Forecasting Time Series Forecasting
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei, China
CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting
Apr 30, 2026Bokai Pan, Mingyue Cheng, Zhiding Liu +6Time Series Forecasting LLM-Based Time Series Forecasting
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei 230026, China

Mar 5, 2026J/K move · Enter open · S save

SEA-TS: Self-Evolving Agent for Autonomous Code Generation of Time Series Forecasting Algorithms

Longkun Xu, Xiaochun Zhang, Qiantu Tuo, Rui Li

Ecoflow Inc.

Time Series Forecasting Code Generation Monte Carlo Tree Search Automated Machine Learning Self-Evolving Agents

Open paper AskPDF arXiv

Accurate time series forecasting underpins decision-making in many domains, yetconventional ML development often faces data scarcity, distribution shift, anddiminishing returns from manual iteration. We propose Self-Evolving Agent forTime Series Algorithms (SEATS), a framework that autonomously generates, val-idates, and optimizes forecasting algorithm code through an iterative self-evolutionloop. Our design combines three mechanisms: (1) Metric-Advantage MCTS(MA-MCTS), which replaces fixed rewards with a statistically normalized advan-tage score for search guidance, (2) code review with running prompt refinement,so every successfully executed solution is reviewed and the running prompt encodescorrective patterns for later iterations, and (3) global steerable reasoning, whichcompares each evaluated node to global best- and worst-performing solutions forcross-trajectory transfer. A MAP-Elites archive maintains architectural diversity.Across four datasets and two metrics, SEATS wins seven of eight comparisonsagainst strong baselines TimeMixer, Timer, and SEMixer

Abstract

Explore similar work

SEA-TS: Self-Evolving Agent for Autonomous Code Generation of Time Series Forecasting Algorithms

A Human-in-the-Loop Autonomous Agent for Industry Time Series Forecasting

CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting