Organizations: Alibaba Group · Chang’an University · University of Electronic Science and Technology of China · University of Bristol · Computer Science, University of California, Irvine · University of Pennsylvania · Xiaomi
Event-informed forecasting requires translating reports and historical responses into changes to a numerical forecast. We propose TIMBRE (Temporal Integration of Memory-Based Responses and Evidence), which combines source-aware representation, state-conditioned response transfer, and reliability-guided fusion before a frozen forecast head. A separate readout adjusts interval widths while preserving the median. In a single-seed, one-epoch development study of 13 tasks, TIMBRE improves MAE over ordinary fusion on eight tasks but over native Chronos-2 on only two. Disabling response transfer in the trained model reduces BTC and AULF MAE by 47.04% and 6.81%, respectively. These findings identify sensitivity to learned response transfer rather than a general forecasting advantage. Missing development-set scores and the absence of retrained ablations limit attribution to individual evidence mechanisms.
Figures & tables
Figure 1: Motivating BTC example. News and numerical history jointly inform a forecast adjustment. Target and forecast positions are schematic, not measured results.
Figure 2: TIMBRE’s numerical and event-memory paths. Source relations (A), response transfer (B), and reliability (C) inform horizon-aware fusion (F), updating future patch states before the native head. C controls median-preserving interval readout after shared postprocessing. Numerical conditioning of B/C/F is omitted; curves and weights are schematic.
Data family
FIT
TUNE
H
M+
BTC
1,999
128
6
128
AULF
2,805
705
48
705
NHC
4,999
807
40
807
Time-MMD (10)
3,990
599
6
588
Total
13,793
2,239
–
2,228
Table 1: Prepared development data. FIT counts follow boundary purging; H is the output-grid length. M+ counts TUNE windows with at least one supported memory point, not full-horizon coverage. NHC uses a masked grid.
Task
n
History mean
Persistence
Seasonal
Chronos-2
Ordinary
TIMBRE
AULF
705
473.04
469.12
270.08
245.52
250.31
268.80
BTC
128
1682.16
311.43
443.23
322.75
428.72
649.24
NHC
807
14.121
12.387
12.387
13.184
13.569
14.447
Time-MMD
Agriculture
64
9.508
3.240
5.952
3.938
4.003
3.999
Climate
64
11.005
4.089
16.006
4.033
4.240
4.187
Table 2: MAE on the common TUNE panel; lower is better. BTC is in USD, AULF in MW, and NHC in knots. Time-MMD rows retain source units; Security values are divided by 106 for display. n counts windows. Bold highlights TIMBRE; underlining marks the lowest value among six methods, using unrounded results. History mean and seasonal naive are exploratory controls added after the trained results were seen.
BTC
AULF
Method
PB
C90 (%)
PB
C90 (%)
Native Chronos-2
118.27
91.02
87.05
83.53
Ordinary fusion
168.10
97.27
88.78
87.09
TIMBRE
239.18
84.24
96.14
81.67
Table 3: Probabilistic TUNE results. PB is mean pinball loss (lower is better); C90 is coverage of the nominal 90% interval. WIS is 2× PB on the shared grid.
BTC
AULF
Inference mode
MAE
PB
MAE
PB
TIMBRE
649.24
239.18
268.80
96.14
w/o A: source
483.14
178.66
263.62
94.04
w/o B: transfer
343.86
126.60
250.51
88.79
w/o C: reliability
731.43
268.70
272.68
97.65
Table 4: Ablation by inference-time removal with the same trained TIMBRE weights, seed 17. All 833 BTC/AULF windows are retained; zero training updates. These are fixed-weight diagnostics, not retrained ablations.
Figure 3: Fixed-checkpoint diagnostics on 833 BTC/AULF TUNE windows: (a) module removal; (b) native–full output interpolation, native MAE = 100; (c) input-content masks retaining availability metadata. MAE changes are relative to FULL. Seed 17; no retraining, setting selection, or across-seed error bars.
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
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.
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
Most time series (TS) models are specialized for a single task, either understanding (i.e., returning text answers about a TS) or generation (i.e., returning a numeric forecast). Only recently have unified models begun to handle the two within a single architecture. Even these models, however, produce the two outputs as task-separated paths and cannot predict a series and explain why that prediction arises within a single coherent response. In this paper, we argue for a task-fused model that jointly produces 1) prediction (generation) and 2) selfexplanation (understanding), thereby integrating 1) numerical TS forecasting and 2) interpretable text reasoning within a single response. To enable the systematic study of this capability, we present both a benchmark and a recipe that jointly address the two tasks. The benchmark, ReasonTS-Bench, identifies five fundamental patterns underlying TS and enables the joint evaluation of both tasks. ReasonCast, our recipe for finetuning any LLM to perform both tasks jointly, yields a model that generates a reasoning chain and a forecast together in a single autoregressive pass. Extensive experiments show that ReasonCast outperforms both LLMs and TS models on prediction accuracy while producing verifiable, causal reasoning. Code is available at: https://github.com/seunghan96/reasoncast.