LLM-driven time series forecasting has become increasingly promising by leveraging contextual features to generate future values directly or refine forecast priors. However, incorporating contextual features alone does not ensure that models can identify which features matter for forecasting and understand what they suggest about future dynamics. We argue that the fundamental challenge is how to extract predictive evidence from contextual features and when and how to leverage it to guide forecasting decisions. In this paper, we propose CastReasoner, an evidence-grounded reasoning model for context-aware time series forecasting. Specifically, CastReasoner uses a structured chain of thought (CoT) to extract predictive evidence from contextual features, infer the future dynamics suggested by this evidence, and assess these dynamics against the forecast prior to guide adaptive forecasting decisions. To enable this capability, we design a reasoning-centric two-stage post-training framework including supervised fine-tuning (SFT) and reinforcement learning (RL). First, we synthesize reasoning trajectories through controllable multi-teacher guidance and select high-quality ones using a novel student-aware difficulty metric. Second, we design fine-grained multi-objective rewards to further improve reasoning and forecasting. To address heterogeneous learning needs across domains in both stages, we introduce a convergence-guided data balancing strategy. Through this two-stage training, CastReasoner improves both reasoning quality and predictive accuracy. Extensive experiments across multi-domain forecasting benchmarks demonstrate that CastReasoner achieves superior forecasting performance.
Figures & tables
Figure 1. Illustrative demand forecasting with and without evidence extraction. Context and a forecast prior support raising the afternoon peak and retaining the evening decline. The same context and forecast prior support two illustrative forecasting paths. Without extracting predictive evidence, a broad uplift misses the afternoon peak and overestimates evening demand. Evidence from afternoon heat and a weak prior peak supports raising the peak, while evening cooling and a declining prior support retaining the decline.
Figure 2. Overview of CastReasoner. Evidence-grounded forecasting links history, context, and a forecast prior to evidence extraction, strategy decision, and forecast construction. Structured CoT synthesis, student-aware CoT selection, cross-domain data balancing, and fine-grained multi-objective RL rewards form the post-training pipeline. The upper panel connects target history, context, and a forecast prior to evidence analysis, selection among retain, revise, and reconstruct, and a complete forecast. The lower panel shows teacher-guided candidate trajectories selected by accuracy and learnability; an equal-mixture pilot estimates each domain's best-validation step, whose square root determines fixed quotas for SFT and RL; reinforcement learning combines fine-grained forecasting rewards with a group-relative policy update and KL regularization.
Model
Load
Price
Solar
Wind
BS
BAQ
HS
MetroPT-3
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
ARIMA
1.8470
7.5514
0.5963
1.0713
1.9798
5.8391
1.0234
1.7691
0.9491
1.7476
0.5808
0.9649
1.2498
3.6027
1.2446
2.4615
Prophet
0.6074
0.7953
0.4398
0.6305
1.3154
3.8053
1.3100
3.0755
0.4855
0.4245
0.3699
0.4783
0.9538
1.6255
0.4238
0.4370
PatchTST
0.3813
0.2884
0.3693
0.6078
0.3315
0.3192
0.8658
1.1368
0.2783
0.1760
0.6104
0.8065
0.9748
1.6576
0.9381
1.2385
iTransformer
0.4431
0.3644
0.4289
0.7556
0.3797
0.3688
0.8724
1.1412
0.3040
0.1971
0.6395
0.8407
0.9245
1.5275
0.9183
1.1653
TimeXer
0.3931
0.3022
0.4041
0.7116
0.4454
0.4239
0.8569
1.1071
0.2832
0.1815
0.6557
0.8716
0.9606
1.8368
0.8939
1.0758
Table 1. Overall forecasting performance of CastReasoner (9B) and baselines on eight datasets, measured by nMAE and nMSE. The best and second-best results in each column are shown in bold and underlined , respectively.
Method
Load
Wind
BAQ
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
CastReasoner
0.2120
0.1299
0.4394
0.3978
0.1317
0.0506
w/o Analysis
0.2504
0.1670
0.4635
0.4266
0.1456
0.0667
w/o Decision
0.2473
0.1616
0.4481
0.4037
0.1353
0.0587
w/o Construction
0.2469
0.1632
0.4539
0.4078
0.1394
0.0573
w/o SFT
0.2625
0.1742
0.6720
0.8196
0.1802
0.0782
Table 2. Ablations of CastReasoner (4B) on Load, Wind, and BAQ. We remove evidence analysis, strategy selection, forecast construction, SFT, or RL one at a time.
Method
Load
Wind
BAQ
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
Original
0.2120
0.1299
0.4394
0.3978
0.1317
0.0506
No Think
0.2331
0.1360
0.4526
0.4087
0.1431
0.0611
Shuffled Think
0.2432
0.1436
0.4677
0.4213
0.1589
0.0751
Corrupted Think
0.2496
0.1514
0.4739
0.4332
0.1657
0.0824
Table 3. Reasoning-trace interventions on Load, Wind, and BAQ. With forecasting inputs fixed, we compare original traces against removed, shuffled, and corrupted traces.
Figure 3. Contextual-input ablations on Load, Wind, and BAQ. We separately remove timestamps, exogenous covariates, and domain knowledge to measure each source’s contribution.
Figure 4. Single- and multi-teacher SFT on Load, Wind, and BAQ. We compare one teacher with diverse teachers under the same trajectory budget and trajectory selection strategy.
Method
Load
Wind
BAQ
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
Random
0.2604
0.1657
0.5105
0.5068
0.1839
0.0697
Accuracy Only
0.2528
0.1606
0.4714
0.4382
0.1586
0.0602
Ours
0.2485
0.1424
0.4634
0.4254
0.1503
0.0537
Table 4. SFT trajectory selection: Random samples one of four teacher traces; Accuracy Only selects the lowest-MAE trace; Ours combines accuracy with student learnability.
Figure 5. Eight-domain mean RL reward: (a) equal (Balanced) versus convergence-guided (Ours) sampling; (b) 0.8B, 4B, and 9B students with fixed data allocation.
Method
Load
Wind
BAQ
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
CastReasoner
0.2120
0.1299
0.4394
0.3978
0.1317
0.0506
w/o Rformat
0.2135
0.1327
0.4413
0.4119
0.1396
0.0587
w/o Raccuracy
0.3786
0.2913
0.7145
0.6539
0.2431
0.1022
w/o Ruplift
0.2321
0.1375
0.4705
0.4277
0.1564
0.0748
w/o Rinterval
0.2301
0.1329
0.4609
0.4211
0.1485
0.0721
Table 5. Reward ablation of CastReasoner (4B) on Load, Wind, and BAQ in nMAE/nMSE. Each variant removes one reward during RL; the full model uses all five.
Figure 11
Figure 8. Reasoning and forecasts for BAQ (retain), Solar (revise), and BS (reconstruct). Curves compare ground truth, forecast priors, and final predictions, showing how contextual evidence guides the scope of forecast intervention.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Domain
Freq.
Time points
Train
Validation
Test
Exogenous Variables
Load
Energy
1 h
61,344
2005-01–2009-11
2009-11–2010-12
2010-12–2012-01
25
Price
Energy
1 h
25,968
2011-01–2013-01
2013-01–2013-07
2013-07–2013-12
2
Solar
Energy
1 h
19,704
2012-04–2013-10
2013-10–2014-02
2014-02–2014-07
12
Wind
Energy
1 h
17,544
2012-01–2013-05
2013-05–2013-09
2013-09–2014-01
4
BS
Urban Mobility
1 h
17,455
2011-01–2012-07
2012-07–2012-09
2012-09–2012-12
8
BAQ
Environment
1 h
35,064
2013-03–2015-12
2015-12–2016-07
2016-07–2017-02
11
Appendix
Table 6. Dataset domains, sampling frequencies, time points, chronological split periods, and available exogenous-variable counts.
Model
Load
Price
Solar
Wind
BS
BAQ
HS
MetroPT-3
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
nMAE
nMSE
0.8B
0.2769
0.1914
0.3352
0.5993
0.1743
0.1576
0.5057
0.5392
0.2083
0.1203
0.1472
0.0624
0.7848
1.2904
0.2445
0.1764
4B
0.2120
0.1299
0.2811
0.4066
0.1548
0.1226
0.4394
0.3978
0.1754
0.0822
0.1317
0.0506
0.7496
1.1267
0.1912
0.0944
9B
0.2123
0.1101
0.2623
0.3658
0.1579
0.1277
0.4443
0.3821
0.1773
0.0858
0.1290
0.0478
0.7215
1.0828
0.1877
0.0998
Appendix
Table 7. Forecasting nMAE and nMSE of CastReasoner with 0.8B, 4B, and 9B student models across eight datasets. The best result for each metric is shown in bold.