Real-world time series are frequently driven by exogenous events and structural shifts, rendering conventional forecasting based solely on historical numerical observations insufficient. While language models can retrieve external news, standard retrieval-augmented approaches struggle with high noise, missing signals, and an inability to reason causally about event impacts. We propose SEER (Self-Evolving Event Reasoning and Retrieval), a closed-loop framework that dynamically optimizes event conditioning for time series forecasting. SEER translates prediction errors into two decoupled feedback mechanisms: (i) a reflective retrieval memory that refines subsequent search queries and filters spurious noise, and (ii) a persistent causal knowledge base that distills transferable domain dynamics. SEER enforces strict chronological boundaries across both event retrieval and reflection, preventing look-ahead bias and data leakage. Across six volatile time-series benchmarks, SEER consistently outperforms state-of-the-art time series foundation models and language model baselines.
Figures & tables
Figure 1: Performance and interpretability of SEER. (a) SEER demonstrates consistent performance improvements over prior forecasting paradigms across event-driven time series tasks (e.g., Memory/SSD prices, Electricity, and Weather). Detailed results and comparisons with additional baselines are provided in Table 2 (§ 4.3 ). (b) A memory price forecasting case study ( 64GB-DDR5 from Pangoly (2026) ) illustrates SEER’s interpretable predictions, driven by dynamic event retrieval and causal knowledge discovery. For details, see Figure 5 and 6 in Appendix B .
Figure 2: Overview of the SEER framework. At step t , an LLM Forecaster ( f ) generates a reasoning process rt , which contains the prediction y^t , based on historical numerical data xt and a dynamic textual context space. This context space supplies the LLM with augmented external events (further retrieved using Mtret and filtered via reflective memory Mtsel ) alongside persistent causal knowledge Kt . Upon observing the true outcome yt , SEER closes the feedback loop by reflecting on the prediction error. The updates are decoupled: Φmem refines the Event-Augmented Memory Mt+1 , while Φknow extracts transferable domain dynamics from past errors to form Kt+1 .
Domain
Entities
Learning Window
Evaluation Window
Horizon
Target (Metric)
Weather
10
3 seasonal ranges †
2026-02-22 – 2026-07-19
7 days
Daily Temp (MAE, MSE)
Electricity
3
3 seasonal ranges †
2026-02-22 – 2026-07-19
7 days
Daily Demand (MAE, MSE)
Stock
50
2025-10-01 – 2026-02-01
2026-02-01 – 2026-05-01
1 day
Daily Price Direction (Accuracy)
Polymarket
146
Pre-split resol. ‡
Post-split resol. ‡
To resol.
Probability (Brier Score)
Memory
20
2024-02-01 – 2025-05-01
2025-05-01 – 2026-06-01
6 m / 9 m
Monthly Price (MAE, MSE)
SSD
10
2024-02-01 – 2025-05-01
2025-05-01 – 2026-08-01
6 m / 9 m
Monthly Price (MAE, MSE)
Table 1: Summary of the six evaluation domains. The table outlines the number of entities, chronological splits for the learning and evaluation windows, forecasting horizons, and target evaluation metrics for each task. Multiple horizons (e.g., 6 / 9 months) represent distinct experimental settings.
Method
Memory 6M
Memory 9M
SSD 6M
SSD 9M
Stock
Weather
Electricity
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
Acc (%)
MAE
MSE
MAE
MSE
Autoformer
390.5
639.0
381.9
421.5
258.7
248.5
324.1
372.1
35.50
3.025
16.82
1751.6
6151.0
DLinear
317.8
474.7
400.8
570.4
211.8
221.8
340.6
479.8
29.67
3.020
16.69
1632.4
5354.9
TimesNet
376.7
628.7
435.1
556.4
210.7
200.4
336.6
446.2
33.42
3.016
16.78
1769.7
6000.1
PatchTST
350.1
558.5
427.2
988.2
215.9
222.2
434.2
714.9
31.17
3.044
17.10
1713.8
5582.4
Crossformer
300.9
386.2
318.1
369.0
211.2
188.0
323.2
400.9
31.33
2.948
15.96
1680.6
5496.1
Table 2: Performance comparison between SEER and baselines on time-series forecasting tasks. MSE values for Memory, SSD, and Electricity are scaled by 10−3 . For the Stock domain, we convert the exact price predictions of baseline methods into price directions to evaluate accuracy; in each column, the best result is shaded and set in bold, and the second best is lightly shaded. Results including confidence intervals are detailed in Table 5 in Appendix C .
Backbone
SEER-Full vs. LLM-RAG
SEER-Full vs. SEER-NoK
SEER-NoK vs. LLM-RAG
Win
Tie
Loss
Δ
Win
Tie
Loss
Δ
Win
Tie
Loss
Δ
Claude Opus 4.8
71.8
20.3
7.9
+64.0
67.6
26.4
6.1
+61.5
32.2
44.1
23.6
+8.6
Claude Sonnet 4.6
54.8
30.4
14.8
+40.0
44.4
37.3
18.3
+26.2
40.5
41.1
18.4
+22.0
Gemini Flash 3.1
54.4
29.1
16.5
+37.9
42.1
38.5
19.5
+22.6
51.5
29.6
19.0
+32.5
Gemini Pro 3.1
41.3
30.5
28.3
+13.0
41.0
32.7
26.2
+14.8
33.3
33.0
33.8
-0.5
All backbones
55.6
27.6
16.9
+38.7
48.8
33.7
17.5
+31.3
38.7
36.6
24.7
+14.1
Table 3: Pairwise LLM-as-a-judge (Grok-4.20-Reasoning) evaluation on forecasting reasoning quality. Results demonstrate that SEER significantly outperforms both the baseline (LLM (Event RAG), denoted as LLM-RAG) and the knowledge-ablated variant (SEER w/o Knowledge, denoted as SEER-NoK). Win / Tie / Loss denote the percentage (%) of records where the first setting is better, tied, or loses. Δ represents the net win rate (Win − Loss).
Backbone
Setting
Mem 6M
Mem 9M
SSD 6M
SSD 9M
Stock
Weather
Electricity
vs. LLM
vs. + Events
MAE
MAE
MAE
MAE
Acc. (%)
MAE
MAE
Δ (%)
Δ (%)
Claude Opus 4.8
LLM (No Events)
376.5
422.0
246.7
367.2
36.58
2.876
1642.2
–
–
LLM (Event RAG)
287.9
372.8
234.5
296.6
36.92
2.590
1620.0
+10.2
–
SEER w/o Knowledge
303.7
370.1
218.6
276.6
37.00
2.589
1659.6
+11.1
+0.9
SEER
240.3
291.5
194.2
285.8
38.25
2.547
1473.3
+19.5
+10.5
Claude Sonnet 4.6
LLM (No Events)
387.5
441.2
259.0
403.7
35.58
3.003
1734.3
–
–
Table 4: Ablation study of SEER’s components across multiple LLM backbones. The best configuration per backbone is highlighted in bold . The final two columns, vs. LLM and vs. + Events , denote the average relative performance improvement across all seven tasks compared to the LLM (No Events) and LLM (Event RAG) baselines. Results including 95% confidence intervals are detailed in Table 6 (Appendix C ).
Figure 3: SEER transfers to non-time-series forecasting. Performance comparison of different agent configurations on the Polymarket event forecsating task (evaluated by Brier Score) and the Selective Stock Forecasting task (evaluated by Accuracy). The progressive improvements highlight that SEER’s evolving method, i.e., integrating both refined event retrieval and causal knowledge, consistently enhances predictive reasoning on event-driven tasks beyond standard numerical data.
Figure 4: Interpretable forecasting of SEER. (a) Memory price forecasting at the 2025-06-01 cut-off. In contrast to numerical models and the baseline LLM (Event RAG) that miss the regime shift, SEER leverages distilled causal knowledge (e.g., UPSIDE SQUEEZE) to accurately predict the price surge. (b) Selective stock forecasting. (Top) For the 29 mutually selected stocks, SEER and LLM (Event RAG) achieve similar accuracies (51.7% vs. 48.3%). However, SEER selectively targets 43 different tickers that it evaluates as more predictable, on which it achieves a higher prediction accuracy than the baseline (51.2% vs. 32.7%), leading to an improved overall performance. (Bottom) An example illustrating how SEER leverages learned causal knowledge to filter out uncertain stocks. The details for the cases are provided in Figures 15 to 18 in Appendix B .
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Memory 6M
Memory 9M
SSD 6M
SSD 9M
Stock
Weather
Electricity
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
Acc (%)
MAE
MSE
MAE
MSE
Autoformer
[258.4, 548.5]
[244.3, 1279.3]
[258.3, 522.7]
[208.9, 675.9]
[89.9, 476.0]
[33.4, 579.8]
[124.8, 606.9]
[66.5, 936.9]
[31.67, 39.42]
[2.713, 3.357]
[13.47, 20.45]
[1392.1, 2150.3]
[3785.6, 8892.5]
DLinear
[198.0, 467.0]
[158.3, 990.3]
[250.0, 575.5]
[245.4, 976.4]
[70.8, 395.1]
[29.4, 521.7]
[122.1, 658.4]
[72.9, 1209.7]
[25.92, 33.33]
[2.700, 3.352]
[13.38, 20.16]
[1325.1, 1978.5]
[3379.1, 7717.5]
TimesNet
[245.6, 535.1]
[232.0, 1284.1]
[300.4, 587.4]
[270.6, 900.1]
[72.2, 392.0]
[26.9, 467.2]
[122.9, 649.6]
[67.7, 1128.0]
[30.17, 36.67]
[2.724, 3.324]
[13.56, 20.27]
[1490.4, 2077.2]
[4191.4, 8052.9]
PatchTST
[223.6, 504.0]
[197.1, 1151.6]
[226.6, 735.0]
[181.0, 2636.1]
[70.7, 407.9]
[30.4, 517.6]
[155.3, 855.6]
[102.6, 1824.8]
[28.00, 34.33]
[2.732, 3.380]
[13.69, 20.78]
[1430.9, 2025.5]
[3843.4, 7631.5]
Crossformer
[196.8, 434.9]
[143.2, 756.5]
[206.4, 458.3]
[146.9, 714.4]
[70.0, 399.1]
[23.6, 440.3]
[121.3, 611.8]
[70.7, 988.4]
[27.83, 34.83]
[2.646, 3.263]
[12.82, 19.29]
[1379.2, 2016.3]
[3584.9, 7771.8]
Appendix
Table 5: 95% confidence intervals for every cell of Table 2 . Each entry is the cluster-bootstrap interval (20,000 resamples; clusters are parts for Memory and SSD, cut-off days for Stock, forecast windows for Weather and Electricity) of the corresponding point estimate; all methods in a column are resampled with the same draws, so the intervals are paired. MSE for Memory, SSD and Electricity is scaled by 10−3 . Bold and shading are carried over from Table 2 (best and second-best point estimate).
Backbone
Setting
Mem 6M
Mem 9M
SSD 6M
SSD 9M
Stock
Weather
Electricity
vs. LLM
vs. + Events
MAE
MAE
MAE
MAE
Acc. (%)
MAE
MAE
Δ (%)
Δ (%)
Claude Opus 4.8
LLM (No Events)
[240.2, 541.9]
[277.9, 589.9]
[81.1, 461.5]
[134.6, 693.7]
[33.08, 40.25]
[2.559, 3.212]
[1303.5, 2017.5]
–
–
LLM (Event RAG)
[187.3, 412.4]
[247.2, 517.5]
[72.2, 456.2]
[114.3, 553.7]
[34.08, 39.67]
[2.290, 2.889]
[1308.9, 1964.1]
[+6.2, +14.8]
–
SEER w/o Knowledge
[191.8, 444.9]
[245.6, 513.0]
[72.6, 411.9]
[102.7, 531.5]
[33.33, 40.75]
[2.315, 2.864]
[1371.8, 1976.6]
[+6.9, +15.9]
[-1.6, +3.7]
SEER
[158.3, 335.7]
[191.8, 409.1]
[69.1, 349.6]
[111.5, 529.7]
[34.00, 42.33]
[2.280, 2.810]
[1217.1, 1768.8]
[+14.0, +26.1]
[+6.4, +15.9]
Claude Sonnet 4.6
LLM (No Events)
[254.6, 554.1]
[292.3, 615.4]
[86.5, 488.3]
[152.5, 757.7]
[32.58, 38.50]
[2.677, 3.367]
[1392.7, 2122.0]
–
–
Appendix
Table 6: 95% confidence intervals for Table 4 . Every cell shows the cluster-bootstrap interval (20,000 paired resamples) of the corresponding entry; the last two columns are the intervals of the average relative improvement over the seven tasks against LLM (No Events) and LLM (Event RAG), obtained by averaging the per-task paired bootstrap replicates. Bold follows the point estimates of Table 4 .
Figure 6: SEER’s reasoning and forecast for the prompt in Figure 5 . It walks the Rule-1 regime checklist to an upside squeeze, citing the events by number, freezes the settled May print as the base (Rule 3, 5), places one × 1.780 step at the next quarter boundary (Rule 4, 5) and a × 1.43 second leg at the following one (Rule 6), then emits nine prices inside <prediction> tags.
Figure 16: SEER’s reasoning and forecast for the prompt in Figure 15 (backbone: Claude-Opus-4.8). SEER applies Rule-1 criteria to predict an upward squeeze, citing specific events ([132], [156], [173], [66–68]). It sets the settled June price (875.18) as the baseline, treating the Q2 fluctuation as non-structural (Rules 3, 6). It then applies a × 1.780 multiplier for Q3 (July 2025; Rules 4, 5) and a × 1.43 multiplier for Q4 (October 2025; Rule 6), maintaining this plateau into Q1 2026.
Figure 18: SEER’s response to Figure 17 (Claude-Opus-4.8). Using learned causal rules, SEER rejects risky candidates—e.g., an M&A-driven spike (MRK, Rule 19) and a headline reversal (CVX, Rules 3/41)—before selecting MU Down (Rule 59) and AMD Down (Rules 5/64). Both predictions correctly closed “Down” the next session, whereas all four baseline picks (MRK Up, AMD Up, CVX Down, XOM Down) failed.
Time series forecasting models often benefit from historical patterns. Inspired by Retrieval-Augmented Generation (RAG), recent research explored retrieving relevant historical time series segments to enhance forecasting. However, relying solely on time series similarity is often insufficient for retrieval under non-stationarity. To address this, we propose a multimodal approach: a Semantics-Enhanced Retrieval-Augmented Time Series Forecasting framework, SERAF. Unlike mainstream approaches that depend only on time series similarity, SERAF conducts dual retrieval over the time series and their self-generated textual descriptions. It retrieves two complementary sets of historical patterns and corresponding futures, which are selectively and jointly used to guide future predictions. Experiments across seven real-world datasets demonstrate the effectiveness of SERAF in bridging numerical and semantic views of time series compared with state-of-the-art baselines.
Shiqiao Zhou, Zipeng Wu, Holger Schöner +3
School of Computer Science, University of Birmingham, Birmingham, United Kingdom · School of Mathematics, University of Birmingham, Birmingham, United Kingdom · Siemens AG, Munich, Germany +1
Time series forecasting relies on historical patterns, but real-world series often exhibit non-stationarity and regime shifts that challenge fully parametric forecasters. Inspired by Retrieval-Augmented Generation (RAG), recent work augments forecasters by retrieving relevant historical segments and using them as external evidence at inference time. However, due to the intrinsic non-stationarity of real-world time series, a highly similar past segment does not necessarily imply a similar future, rendering similarity-only retrieval brittle and prone to redundancy. We propose Stationarity-Aware Retrieval-Augmented Time Series Forecasting (SARAF), a framework that adaptively balances relevance and diversity in retrieval. SARAF first forms a candidate pool via temporal similarity with time-aligned enhancement, then applies a diversity-aware selection strategy to cover heterogeneous historical regimes, with the diversification strength automatically modulated by dataset-level stationarity. Moreover, SARAF uses stationarity-aware aggregation to fuse the retrieved futures. Extensive experiments on eight real-world datasets show that SARAF achieves competitive forecasting performance and improves average accuracy and robustness over strong baselines, with particularly clear benefits under challenging non-stationary settings. Code: https://github.com/ShiqiaoZhou/SARAF.
Shiqiao Zhou, Holger Schöner, Zipeng Wu +3
University of Birmingham Birmingham, United Kingdom · Siemens AG Munich, Germany
Real-world time-series applications increasingly require models that can handle time series forecasting, context-conditioned prediction, and language-based temporal reasoning. Yet current time-series foundation models remain fragmented across these capabilities: numerical specialists often provide the strongest forecasts, while language-based models offer broader contextual understanding and analysis. A central challenge is to unify these heterogeneous capabilities without reducing their individual performance. We introduce OpenTSLM TeeMoE, a generalist time-series language model that can forecast directly from observed time series, reason over textual context and temporal patterns, and synthesize and refine predictions from external numerical forecasting specialists. We independently train three low-rank experts for forecast aggregation, native forecasting, and temporal analysis over a shared backbone. A learned LoRA mixture-of-experts controller then weights their frozen parameter updates for each request. Our proposed model achieves strong performance on widely used benchmarks for time series forecasting, context-conditioned prediction, and language-based temporal reasoning, ranking among the top three on GIFT-Eval by mean MASE rank, Context is Key by RCRPS, and TimeSeriesExam by accuracy.
Tony Chen, Timo Stoffregen, Maxwell Xu +8
Columbia University, USA · Stanford University, USA · Aionic Labs, Switzerland +4