Cast-R1RESEARCH
TIME SERIES × AGENTIC LEARNING

Learning to forecast,
one decision
at a time.

An 8B language model that learns to analyze time series, choose forecasting tools, and refine predictions using accumulated evidence.

Read the paper Code / GitHub
Explore a forecasting episode
A FORECASTING EPISODEETTh1 · 96 → 96
From observations
to an informed forecast.
01 Diagnose→02 Forecast→03 Refine
Observed history Cast-R1 forecast
Cast-R1 project logo, featuring the R1 robot and trajectory distillation, memory, and reinforcement learning.

Learning Agentic Time Series Forecasting via Trajectory Distillation
and Memory-Based Reinforcement Learning

Xiaoyu Tao1·Mingyue Cheng1·Yaguo Liu1·Xingpeng Gao1·Qingsong Wen2·Qi Liu1·Shijin Wang3·Enhong Chen1

1 University of Science and Technology of China2 Squirrel Ai Learning3 iFLYTEK Research

Revised manuscript · Public preprint (earlier version) ↗

8BForecasting policy
10Benchmark datasets
9/10Lowest MSE
8.41%Mean relative MSE reduction ↗
THE IDEA

Forecasting is a
sequence of decisions.

What should we inspect? Which forecaster fits the evidence? How should its prediction be revised?

Cast-R1 learns these decisions through forecasting-specific post-training. A single policy coordinates diagnostic tools, routes among forecasting models, and revises their outputs within a structured workflow.

Acquire evidence

Inspect data quality, statistical properties, temporal structure, events, and residuals.

Choose a forecaster

Route to ARIMA, Chronos, PatchTST, or iTransformer using diagnostic evidence.

Refine the prediction

Combine recent observations and stored evidence to revise the intermediate forecast.

CAST-R1 IN ACTION

Follow the forecast.

One ETTh1 episode. From historical evidence to a revised 96-step prediction.

FIGURE-BASED REPLAY
01 / 04
TRANSFORMER OIL TEMPERATURE

Start with the observed history.

ETTh1
History Initial forecast Refined forecast Held-out target
DIAGNOSTIC EVIDENCE

Statistical and temporal diagnostics characterize recurring fluctuations in the 96-hour history. Their summaries are retained in Analysis History for later decisions.

Replay adapted from the paper’s qualitative case. Curves are extracted from the original vector figure; summaries describe its stages. Future targets appear only in the final comparison. View original figure ↗

HOW IT LEARNS

Distill the interactions.
Learn from the outcomes.

Two training stages turn a Qwen3-8B backbone into a memory-conditioned forecasting policy.

STAGE 01SUPERVISED FINE-TUNING

Trajectory distillation

A DeepSeek-V3.2 teacher interacts with the forecasting harness. Complete, valid trajectories are ranked by forecast error within each dataset.

Teacher rollouts→Evaluate & select→SFT
1,200selected demonstrations
120 from each dataset
STAGE 02MULTI-TURN GRPO

Reinforcement learning

The distilled policy learns from terminal rewards for accuracy, temporal structure, and output completeness, with a difficulty-aware curriculum.

Group rollouts→Forecast rewards→GRPO
8trajectories per input
for group-relative optimization
STRUCTURED, EPISODE-LOCAL MEMORY

The next decision starts with the evidence so far.

Analysis History retains diagnostic outputs. Prediction Results retains intermediate forecasts. Each new policy state is constructed from these two components and the task inputs.

Mk = (Dk, Qk)
Explore the complete frameworkThe paper's complete framework showing memory-guided forecasting and two-stage policy training

Memory-guided forecasting (a) and two-stage policy learning (b). Select the figure to open its PDF.

BENCHMARK RESULTS

One policy.
Ten forecasting benchmarks.

Evaluated against 15 baselines across long-term forecasting and electricity price forecasting.

MEAN RELATIVE MSE REDUCTION8.41%

vs. the best baseline on each dataset

12.65%

Long-term forecasting

4.16%

Electricity prices

Arithmetic mean of per-dataset relative reductions, including the negative result on PJM.

Relative MSE reduction

Compared with each dataset’s strongest baseline · higher is better

Lower error than best baseline Higher error than best baseline

Reported point estimates from the revised manuscript. A single agent is trained across all ten datasets. Errors use the raw data scale; the two benchmark families have different supported input regimes.

Full benchmark table 16 models · raw-scale errors
MSE · lower is better

Bold = best. Underlined = second best. The selected metric and benchmark family also apply to this table.

WHAT CONTRIBUTES?

Both training stages
and retained evidence matter.

The reported ablations favor combining trajectory distillation and RL. Removing memory also increases MSE on the four displayed benchmarks.

Read the ablation analysis
Paper ablations comparing memory, SFT and RL, curriculum, and backbone sizes
EXPLORE & REPRODUCE

Take the next step.

The implementation includes the shared forecasting harness, trajectory selection, SFT, multi-turn GRPO, and evaluation.

Read the manuscriptMethod, experiments, and implementation details Explore the codeTraining pipeline and forecasting harness ↳Prepare data & modelsBenchmark sources, context, and checkpoints ⌘Reproduce the workflowFrom teacher trajectories to policy evaluation

Python 3.11. Full policy training and serving require a compatible CUDA environment.

CITATION

Build on this work.

For the public arXiv preprint, use its published title and author list. The revised manuscript linked above has updated metadata.

arXiv:2602.13802
BIBTEX · PUBLIC PREPRINT
@article{tao2026cast,
  title   = {Cast-R1: Learning Tool-Augmented Sequential
             Decision Policies for Time Series Forecasting},
  author  = {Tao, Xiaoyu and Cheng, Mingyue and Jiang, Chuang
             and Gao, Tian and Zhang, Huanjian and Liu, Yaguo},
  journal = {arXiv preprint arXiv:2602.13802},
  year    = {2026},
  doi     = {10.48550/arXiv.2602.13802},
  url     = {https://arxiv.org/abs/2602.13802}
}