acceptodds
Under review as a conference paper at ICLR 2027

Adapting Time Series Foundation Models for Event-Conditioned Multimodal Forecasting

16 views

Abstract

Time series foundation models (TSFMs) have demonstrated strong zero-shot forecasting capabilities through large-scale pretraining on heterogeneous time series data. However, they cannot natively incorporate natural-language descriptions of external events, such as promotional events, holidays, and policy changes, that provide predictive information not recoverable from historical observations alone. Existing work either relies on LLM-based reasoning that does not guarantee accurate numerical forecasts, or requires costly joint multimodal pretraining from scratch. To address this, we propose MM-Time, a parameter-efficient framework that equips pretrained TSFMs with event-aware forecasting through multimodal continued pretraining. To bridge the semantic–effect gap in which semantically distinct events exert identical numerical impacts, we introduce an Event-Effect Prototype Router. It maps event descriptions and temporal attributes to sparse combinations of shared learnable prototypes under forecasting supervision, encouraging the sharing of forecast-relevant information across events. Furthermore, to address temporal alignment over event durations, we propose Event Interval Rotary-Bias Attention, which conditions temporal attention on event boundary positions and incorporates an interval-distance bias to favor temporally relevant events. We adapt three state-of-the-art TSFMs, Chronos-2, Falcon-2, and TiRex-2, and show that MM-Time consistently improves event-conditioned forecasting across multiple downstream datasets.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series foundation models, Time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_uUB1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MM-Time, a framework that enables pretrained time series foundation models to use event descriptions, such as promotions and holidays, when forecasting. Its Event-Effect Prototype Router maps event descriptions and temporal attributes to shared learnable prototypes. Event Interval Rotary-Bias Attention uses event start and end times to guide when this information should influence the forecasts. The framework adapts existing models by training event-conditioning modules, DoRA adapters, and the forecasting head while freezing the remaining backbone parameters. Experiments on e-commerce and retail datasets show improved forecasting with Chronos-2, Falcon-2, and TiRex-2.

Critical strengths
  • S1. MM-Time incorporates event information into pretrained TSFMs by translating event descriptions into useful forecasting information and aligning that information with prediction times.

  • S2. The experiments support the framework’s effectiveness across three TSFM backbones.

Critical weaknesses
  • W1. The dataset descriptions are insufficient. Specifically, it is unclear where the data came from, how the event descriptions were collected or generated, and how the authors verified that this information was available at forecast time.

  • W2. Although the ablation studies show that the Prototype Router improves forecasting accuracy, they do not clearly demonstrate that it learns meaningful numerical effects shared across semantically different events, as claimed in the paper.

  • W3. Although the ablation studies show that Interval RoPE and Locality Bias improve forecasting accuracy, it remains unclear whether the attention mechanism actually learns meaningful temporal relationships between events and forecast steps.

  • W4. Forecasting results are provided without statistical significance tests. Since some improvements in the ablation studies are relatively small, it is unclear whether these differences are statistically reliable.

  • W5. The use of the term "zero-shot forecasting" may be potentially misleading. Unlike the original TSFMs, MM-Time undergoes additional multimodal continued pretraining on time series paired with event descriptions.

  • W6. The downstream evaluation is limited to e-commerce and retail datasets.

Questions for authors
  • Q1. Could the authors clarify the sources of the continued-pretraining and evaluation datasets, how event descriptions were collected or generated, and how their availability at forecast time was verified?

  • Q2. Could the authors provide further evidence that the learned prototypes capture shared numerical effects rather than simply grouping semantically similar events?

  • Q3. Could the authors analyze the learned attention weights to show whether the model actually attends to relevant events at appropriate forecast steps?

  • Q4. Could the authors report standard deviations or confidence intervals across multiple runs, particularly for the ablation studies?

  • Q5. Could the authors clarify why the evaluation is described as zero-shot forecasting, given that MM-Time undergoes additional multimodal continued pretraining?

  • Q6. Could the authors extend the evaluation to additional domains beyond e-commerce and retail to better demonstrate the generalizability of MM-Time?

5 comments
AuthorsOpenReview ↗

Response to Reviewer uUB1(Part I)

We thank the reviewer for the constructive comments on data transparency, the proposed mechanisms, and the clarity and scope of our evaluation.

W1, Q1. Data Sources, Event Construction, and Availability at Forecast Time.

A1: Data sources and event construction. In the original experiments, both the continued-pretraining and evaluation datasets consist of real-world business time series from our company’s internal datasets, covering e-commerce and retail scenarios. The accompanying event information is collected from published announcements and public calendars, including promotional campaign announcements and holidays in the countries or regions associated with the time series. Event descriptions are compiled from these sources and anonymized to remove sensitive information.

Information availability at forecast time. We define the availability time of event information as its publication timestamp on the source website, distinguishing it from the time when the event actually occurs. For each forecast origin, we include only event information published before that origin. Accordingly, promotions or holidays announced in advance can be used as known information even when they occur within the forecast horizon. Information published after the forecast origin is excluded from the model inputs. We have added these descriptions of data sources, event construction, and information availability to Appendix A of the revised paper‾\underline{\text{Appendix A of the revised paper}}.

W2, Q2. Shared Numerical Effects versus Semantic Similarity in Prototype Routing.

A2: Thank you for this suggestion. In addition to the forecasting ablations, Appendix E.2 and Figures 6–8 of the revised paper‾\underline{\text{Appendix E.2 and Figures 6–8 of the revised paper}} provide a qualitative analysis of event-to-prototype routing. With the trained MM-Time-C2 parameters held fixed, we compare event descriptions, standardized observed time series, standardized residuals relative to the original Chronos-2 forecasts, and the full 32-dimensional prototype routing weight vectors.

For example, the season-end sale and Valentine’s Day gift promotion in Figure 6 have different campaign themes, yet exhibit a residual cosine similarity of 0.885 and a routing cosine similarity of 0.811, sharing three of their four selected prototypes. In Figure 7, tiered discounts and free shipping combined with coupons involve different promotional mechanisms but exhibit similar residual patterns, with a residual cosine similarity of 0.971. They also share three selected prototypes while retaining distinct routing weights and a different fourth prototype. Figure 8 further illustrates prototype reuse across different entities.

These examples provide qualitative support for the learned prototypes capturing shared numerical response patterns relevant to forecasting across events with different themes or promotional mechanisms.

W3, Q3. Temporal Alignment Between Events and Forecast Steps in Learned Attention.

A3: Thank you for the suggestion. We have added attention analyses in Appendix E.3, Figures 9–10‾\underline{\text{Appendix E.3, Figures 9–10}}. Figure 9 shows attention shifting from E08 and E10 toward E11 across rolling forecast rounds, while Figure 10 shows different attention weights for events with overlapping intervals. These observations provide qualitative evidence that event weighting varies across forecast positions and distinguishes among overlapping events, complementing the ablation results for Interval RoPE and Locality Bias.

AuthorsOpenReview ↗

Response to Reviewer uUB1(Part II)

W4, Q4. Statistical Reliability of Forecasting and Ablation Results.

A4: Since each ablation setting requires continued pretraining, repeating all settings across multiple random seeds would incur substantial computational costs. We therefore conducted three independent continued-pretraining runs with different random seeds for both the full MM-Time-C2 model and its w/o Interval RoPE variant on the Chronos-2 backbone. This variant performed closest to the full model in the original ablation experiments.

The table below reports the mean ± sample standard deviation of MASE across the three runs. The full model achieves lower mean MASE on both E-Commerce and Retail, with approximate relative reductions of 3.758% and 1.648%, respectively. These repeated-run results provide further support for the benefit of Interval RoPE. The Overall comparison is included in Section 4.2 of the revised paper, and the domain-level results are provided in the table below.

Evaluation scopeMM-Time-C2: MASE ↓w/o Interval RoPE: MASE ↓
E-Commerce0.717 ± 0.0060.745 ± 0.004
Retail0.716 ± 0.0030.728 ± 0.002
Overall0.716 ± 0.0040.736 ± 0.002

W5, Q5. Clarifying Zero-Shot Forecasting after Multimodal Continued Pretraining.

A5: Thank you for pointing out this potential ambiguity. In this paper, zero-shot forecasting refers to how MM-Time is evaluated on target downstream datasets after multimodal continued pretraining.

MM-Time does undergo an additional multimodal continued pretraining stage using paired time series and event descriptions that are kept separate from the target evaluation data. After this stage, we freeze the model parameters and directly forecast on target datasets that were not used during continued pretraining, without further training or fine-tuning on those datasets. Historical observations of the target series and event information available at forecast time are used solely as inference inputs.

To clarify the distinction between the training procedure and the evaluation setting, we have described this setting as “zero-shot downstream forecasting after multimodal continued pretraining” in Section 4 (Baselines) of the revised paper‾\underline{\text{Section 4 (Baselines) of the revised paper}}.

W6, Q6. Generalization Beyond E-Commerce and Retail.

A6: To broaden the domain coverage of our evaluation, we have added experiments on three public datasets (Fidel-TS_Germany, CGTSF_LEU, and CiK_Transport), extending the evaluation beyond e-commerce and retail to energy and transportation.

For these additional experiments, we incorporated other datasets or tasks from Fidel-TS and CiK into the continued-pretraining corpus while holding out the three target evaluation datasets. The target evaluation data were not used during continued pretraining. After training, the models were evaluated directly on these datasets without additional target-specific training or fine-tuning.

The table below reports MASE for the three MM-Time variants, their corresponding backbones, and other foundation models. All three variants outperform their respective backbones on all three datasets, with relative MASE reductions of 2.637%–4.260% for MM-Time-C2, 2.057%–30.993% for MM-Time-F2, and 2.718%–7.054% for MM-Time-T2. These results provide further evidence of MM-Time’s effectiveness in domains beyond e-commerce and retail. We have included these experiments in Section 4.1 and Table 1 of the revised paper‾\underline{\text{Section 4.1 and Table 1 of the revised paper}}.

DatasetMM-Time-C2MM-Time-F2MM-Time-T2Chronos-2Falcon-2TiRex-2TimesFM2.5Moirai-2Timer-S1
Fidel-TS_Germany0.7810.8510.7430.8100.8950.7870.8370.9050.819
CGTSF_LEU0.6790.7840.7020.7091.1360.7210.7270.7250.721
CiK_Transport0.5320.5950.5840.5460.6080.6280.6690.5960.699

Thank you for the detailed rebuttal and additional analyses. However, I still feel that the proposed framework remains somewhat of a black box, and I find it difficult to understand what the model actually learns and how the proposed components contribute to forecasting.

  • W2. For the Prototype Router, the examples suggest that events with similar residual patterns may share prototypes, but it remains unclear whether the prototypes systematically capture shared numerical effects rather than semantic similarities.

  • W3. The attention visualizations show that attention weights change across forecast steps and events. However, since the attention mechanism explicitly incorporates temporal information through Interval RoPE and Locality Bias, it remains unclear to what extent the observed attention patterns reflect learned temporal relationships.

Reviewer_1ZehOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MM-Time, a parameter-efficient framework for adapting pretrained unimodal TSFMs through multimodal continued pretraining. This paper introduces a shared dictionary of latent effect prototypes as an intermediate representation between event descriptions and numerical forecasts, bridging the semantic–effect gap. The proposed Event-Effect Prototype Router maps event text and structured temporal features to sparse combinations of these prototypes, yielding event-specific effect tokens. Rather than relying on superficial semantic matching, the router and prototypes are jointly trained under forecasting supervision to capture predictive structures shared across events, effectively decoupling surface text from its underlying numerical impact. The proposed Event Interval Rotary-Bias Attention addresses temporal alignment by aligning these tokens with future forecast patches through cross-attention. Interval RoPE explicitly handles interval-based events by encoding the relative timing between forecast patches and event boundaries, while Interval Bias introduces a soft locality prior by penalizing attention according to distance outside the event interval. This design allows the pretrained numerical backbone to incorporate event information according to both its forecast-relevant content and temporal relevance.

Critical strengths

First, this paper innovatively introduces the concept of “MM-Time,” independently constructs a large-scale dataset in the relevant field, and validates the proposed method; the selected baseline methods are all state-of-the-art (SOTA) research results. Furthermore, the paper is written in a clear and accessible style, and the experimental results are presented very clearly. The main experiments and ablation studies are both very comprehensive.

Critical weaknesses

I believe one of the main shortcomings of this paper is that its evaluation metrics are too narrow, relying solely on MASE; meanwhile, among the baseline methods selected by the authors, most relevant methods typically test multiple point-wise prediction error metrics. For example, UNICA’s error metrics include not only point-wise error measures such as MAE, MAPE, and MSE, but also the probabilistic error measure CRPS. Furthermore, why is this point worth noting? It is because, although the authors indicate in the problem definition that the model is a probabilistic representation, they did not use any probabilistic error metrics to evaluate the results in the actual evaluation. From this perspective, the authors’ choice of a single error metric—MASE—seems somewhat peculiar. Furthermore, the appendix does not provide a complete table of results; it is unclear whether the reported results are from a single run or the average of multiple runs. If the authors are reporting the average of multiple runs, they should provide more detailed information, including the upper and lower bounds of the average results, standard deviation, and so on; however, the authors do not explicitly state that they conducted repeated experiments. Furthermore, given that the authors used an unspecified proprietary dataset and a non-public benchmark—and considering that the selected dataset is quite large, involving multiple datasets from different domains and with varying prediction horizons—disclosing these specific results would help in understanding the model’s performance capabilities or describing its performance on specific domain tasks. The authors neither published a complete results table in the appendix nor disclosed the experimental configurations used for the various baseline models;

Questions for authors

I would like to thank the authors for their work; this is an excellent piece of research. I have a few questions regarding the evaluation and reproducibility:

  1. Why was MASE the sole metric chosen? Furthermore, given that the problem was defined in terms of probabilistic output and probabilistic modelling, why was the probability error not calculated? Could the authors calculate or disclose the results for other error metrics—particularly those relating to probability error—as well as the complete results from multiple repetitions of all experiments?
  2. What was the specific experimental setup for the relevant baseline?
  3. How much training time was used in total? How many graphics cards were used?

Once it has been confirmed that the author’s work is highly reproducible, and provided that the data, code and results have been made publicly available, I am inclined to raise the score to 3 or 4 marks.

4 comments
AuthorsOpenReview ↗

Response to Reivewer 1Zeh (Part I)

We appreciate the reviewer's careful assessment and helpful suggestions for improving the completeness and reproducibility of our experiments.

W1, Q1. Metric Selection, Probabilistic Evaluation, and Results Across Multiple Runs.

A1: Thank you for raising these questions about the evaluation metrics and repeated runs.

Evaluation metrics. We selected MASE as the primary metric because our evaluation covers time series from different domains, with varying sampling frequencies and numerical scales. Unlike directly aggregating MAE or MSE, MASE normalizes each series’ prediction error by its historical naïve or seasonal naïve forecasting error. This prevents larger-scale series from dominating the aggregate results and provides an interpretable reference relative to a simple forecasting baseline, making it suitable for comparisons across heterogeneous datasets.

To complement this point-forecast evaluation, we have added normalized CRPS to assess the quality of the predictive distributions, along with MAPE as an additional point-forecast metric. The complete results are provided in Appendix F, Tables 12–13 of the revised paper‾\underline{\text{Appendix F, Tables 12–13 of the revised paper}}. Across the nine dataset–horizon settings, MM-Time variants achieve the lowest CRPS in six settings and the lowest MAPE in seven among all evaluated methods. These results support the benefits of MM-Time beyond MASE, demonstrating improvements in both probabilistic and point-forecast accuracy.

Repeated runs. The foundation-model baselines in the main experiments use fixed, publicly available pretrained checkpoints for zero-shot inference, without additional training or fine-tuning. Given the computational cost of continued pretraining across the three backbones, the original results for MM-Time-C2, MM-Time-F2, and MM-Time-T2 were each obtained from a single continued-pretraining run.

To assess variability due to training randomness, we additionally conducted three independent continued-pretraining runs of MM-Time-C2 using different random seeds. All runs started from the same backbone checkpoint and used the same training data, training budget, and evaluation protocol. The table below reports the mean ± sample standard deviation across the three runs.

Evaluation scopeMASE ↓Normalized CRPS ↓
E-Commerce0.717 ± 0.0060.110 ± 0.001
Retail0.716 ± 0.0030.634 ± 0.005
Overall0.716 ± 0.0040.291 ± 0.002

Complete normalized CRPS results (↓), also reported in Appendix F, Table 13 of the revised paper‾\underline{\text{Appendix F, Table 13 of the revised paper}}.

ModelE-Commerce 1H (Long)E-Commerce 1H (Medium)E-Commerce 1H (Short)E-Commerce 1W (Short)Retail 15T (Long)Retail 15T (Medium)Retail 15T (Short)Retail 1D (Medium)Retail 1D (Short)
MM-Time-C20.0600.0800.0940.3170.6070.5870.5470.7820.668
MM-Time-F20.0650.0690.0860.3090.6650.5910.5580.8240.732
MM-Time-T20.0710.0820.0970.3140.6810.6430.5810.7500.686
TimesFM2.50.0740.0890.1120.3210.6670.6360.6010.7950.712
Toto 2.0-4m0.0880.1030.1290.3080.6670.6540.6360.7890.728
Timer-S10.0770.0900.1130.2960.6340.6090.5870.8540.667
Moirai-20.0830.0970.1180.3160.6330.6240.6030.7560.712
Falcon-20.0780.0850.1060.3850.6200.5910.5510.9280.913
Chronos-20.0820.0850.1080.3260.6210.5780.5670.7850.667
Sundial0.0910.1090.1350.3400.6890.6800.6590.7660.718
TiRex-20.0750.0840.0980.3340.6410.6240.6130.8390.719
Aurora0.0940.1060.1350.3290.9700.9661.0170.9860.987
AuthorsOpenReview ↗

Response to Reivewer 1Zeh (Part II)

Complete MAPE results (↓), also reported in Appendix F, Table 12 of the revised paper‾\underline{\text{Appendix F, Table 12 of the revised paper}}.

ModelE-Commerce 1H (Long)E-Commerce 1H (Medium)E-Commerce 1H (Short)E-Commerce 1W (Short)Retail 15T (Long)Retail 15T (Medium)Retail 15T (Short)Retail 1D (Medium)Retail 1D (Short)
MM-Time-C20.2550.2450.1750.2380.9190.8680.7490.1610.135
MM-Time-F20.2390.2160.1580.2310.8600.8250.7070.1690.141
MM-Time-T20.2710.2610.1940.2270.9920.8350.7510.1660.141
TimesFM2.50.2820.2550.1680.2531.1971.0280.8530.1650.141
Toto 2.0-4m0.3390.3110.2090.2441.2641.2131.1110.1730.148
Timer-S10.2780.2550.1780.2641.0971.0831.0290.1730.133
Moirai-20.2910.2550.1670.2381.2121.1180.9420.1620.144
Falcon-20.2520.2180.1650.3160.8940.8050.7150.1910.159
Chronos-20.3210.2760.1940.2461.0450.9580.8840.1680.136
Sundial0.3770.3350.2280.2821.8161.6541.1710.1780.138
TiRex-20.2910.2850.1820.3031.0850.9320.7660.1790.141
Aurora0.3790.3390.2560.2852.8292.9362.5750.2000.166
ChatTime1.0841.0781.2080.4456.6778.6947.4560.2720.283
VoT0.5560.4890.3200.2582.3921.9992.3640.1680.159
GPT4MTS0.7250.4340.3630.2612.4212.2041.8860.1910.170
TaTS0.9940.9600.8451.1440.9280.8860.9080.9570.939
TESS0.5660.6020.4930.2562.0752.4041.9980.2260.167
Time-LLM0.7310.6020.4070.2782.7583.1042.3030.1990.181
TimeCMA0.5270.4830.3020.3572.5543.0931.8150.2080.167

W2, Q2. Complete Results and Reproducible Baseline Configurations.

A2: Complete results. We have included the complete MASE results for all 19 methods across nine dataset–forecast-horizon combinations in Appendix F, Table 11 of the revised paper‾\underline{\text{Appendix F, Table 11 of the revised paper}}. These results show performance across different domains, sampling frequencies, and forecast horizons.

Complete forecasting results measured by MASE (↓), also reported in Appendix F, Table 11 of the revised paper.‾\underline{\text{Appendix F, Table 11 of the revised paper.}}

ModelE-Commerce 1H (Long)E-Commerce 1H (Medium)E-Commerce 1H (Short)E-Commerce 1W (Short)Retail 15T (Long)Retail 15T (Medium)Retail 15T (Short)Retail 1D (Medium)Retail 1D (Short)
MM-Time-C20.6650.6630.6040.9940.7320.6800.6650.7760.726
MM-Time-F20.6940.6430.6020.9700.7740.6980.6740.8940.776
MM-Time-T20.7950.7490.6550.9330.7900.7310.6880.7990.729
TimesFM2.50.7630.7220.6411.0970.8280.7490.7020.8620.754
Toto 2.0-4m0.8230.7890.7011.0550.8180.7660.7370.8460.780
Timer-S10.7640.7310.6381.1360.7780.7250.6980.8980.721
Moirai-20.7920.7370.6521.0380.8120.7560.7050.8070.776
Falcon-20.6800.6600.6041.3850.7480.6980.6661.0480.978
Chronos-20.8150.7310.7331.0870.8020.7170.6760.8190.709
Sundial0.8130.7760.6791.2390.8170.7570.7100.8270.739
TiRex-20.8160.7670.6501.2770.7640.7080.6910.8490.762
Aurora0.8120.7860.7511.2410.9620.9050.9030.9620.911
ChatTime2.5192.5912.8722.4802.3082.4022.4331.2101.555
VoT1.0220.9700.8321.0750.9410.8570.8940.8360.855
GPT4MTS1.1780.9710.9161.1120.9810.9250.8770.9190.912
TaTS1.0790.9290.8821.1161.0710.9640.8591.0210.929
TESS1.0891.1111.0871.1850.9470.9180.9011.0580.883
Time-LLM1.3161.2071.1621.2431.1191.1490.9780.9710.986
TimeCMA1.0791.1110.9271.5711.0771.1370.9000.9400.862
AuthorsOpenReview ↗

Response to Reivewer 1Zeh (Part III)

Baseline configurations. We have added the experimental configurations of all 16 main-comparison baselines to Appendix C.5, Tables 8–9 of the revised paper‾\underline{\text{Appendix C.5, Tables 8–9 of the revised paper}}. Table 8 describes the checkpoints, input history lengths, and inference settings of the fixed-checkpoint baselines; Table 9 reports the backbones and training settings of the downstream-trained baselines. All implementations and pretrained checkpoints are obtained from their official releases where applicable. All methods are evaluated using the same test splits, forecast origins, and forecast horizons.

Q3. Total Training Time and GPU Resources.

A3: Training time and computational resources. All continued-pretraining experiments use two NVIDIA B200 GPUs with Distributed Data Parallel (DDP). Each model is trained for 100,000 optimization steps, with a per-GPU batch size of 256 and a global batch size of 512, using FP32 precision. Text and event embeddings are generated offline by a frozen LLM text encoder and cached. During continued pretraining, we directly load these cached embeddings without updating the text encoder.

Under this configuration, a single 100,000-step run takes approximately 12 hours for MM-Time-C2, 14 hours for MM-Time-F2, and 20 hours for MM-Time-T2. Together, one run of each variant requires approximately 46 hours of combined training time, equivalent to 92 GPU-hours.

Code and data availability. Because the evaluation data contain proprietary business information, the code and evaluation data are currently undergoing our company’s internal disclosure review. We expect to release them within the next one to two days, subject to approval.

Thank you for your thoughtful reply; I’ve decided to raise the mark.

Reviewer_ewZvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MM-Time, a parameter-efficient framework that adapts pretrained time series foundation models to event-conditioned forecasting through multimodal continued pretraining. It maps event descriptions to sparse combinations of shared prototypes and fuses them with numerical states through interval-aware cross attention, and is evaluated on Chronos-2, Falcon-2 and TiRex-2 using e-commerce and retail data.

Critical strengths
  • The framework reuses pretrained numerical representations and trains only lightweight modules, and it consistently improves three architecturally different TSFMs.

  • The paper separates numerical continued pretraining from event conditioning and compares against covariate adaptation methods under the same training protocol, which is a good empirical practice.

Critical weaknesses
  • All experiments use 61 proprietary series from two business domains, including only 7 retail series. No public multimodal benchmark is used. The continued pretraining corpus also comes from the same business domain, so neither the zero-shot claim nor the generality of the results can be assessed.

  • The performance gains largely come from domain-matched continued pretraining. The comparison in Figure 2 pits the continually pretrained MM-Time against zero-shot baselines. Figure 3 shows that numerical continued pretraining alone accounts for more than half of the improvement on all three backbones. The event modules add only 1.7% to 4% in MASE. The headline gains therefore reflect additional in-domain training rather than event conditioning.

  • Missing comparison with TSFM adaptation methods: ChronosX is cited but not evaluated. Two simple alternatives are also absent: event indicators passed through the native covariate interface of Chronos-2 and TiRex-2, and LoRA adaptation on downstream data.

Questions for authors

see weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer ewZv (Part I)

We thank the reviewer for the thoughtful feedback on our evaluation protocol, performance attribution, and baseline comparisons.

W1. Zero-Shot Evaluation and Generalization Beyond E-Commerce and Retail.

A1: Thank you for raising this concern. We first clarify that “zero-shot” refers to forecasting on target datasets without dataset-specific training or fine-tuning after multimodal continued pretraining. The target evaluation datasets are excluded from our continued-pretraining corpus. This evaluation setting is clarified in Section 4 (Baselines) of the revised paper‾\underline{\text{Section 4 (Baselines) of the revised paper}}. It does not require the pretraining and evaluation data to come from entirely different domains. In the original experiments, the continued pretraining and evaluation data were drawn from different business scenarios, and the evaluation datasets were not used for continued pretraining.

To broaden the evaluation beyond proprietary e-commerce and retail data, we added three public evaluation datasets: Fidel-TS_Germany, CGTSF_LEU, and CiK_Transport, covering energy and transportation. For these additional experiments, we augmented the continued-pretraining corpus with other datasets from Fidel-TS and CiK, while keeping the three target evaluation datasets excluded from continued pretraining. The resulting models were evaluated directly on these held-out datasets without further training or fine-tuning.

Results: All three MM-Time variants outperform their corresponding backbones on all three public datasets, with relative MASE reductions of 2.057%–30.993%. These results support the applicability of MM-Time beyond e-commerce and retail to energy and transportation. We have included this evaluation in Section 4.1 and Table 1 of the revised paper‾\underline{\text{Section 4.1 and Table 1 of the revised paper}}.

DatasetMM-Time-C2MM-Time-F2MM-Time-T2Chronos-2Falcon-2TiRex-2TimesFM2.5Moirai-2Timer-S1
Fidel-TS_Germany0.7810.8510.7430.8100.8950.7870.8370.9050.819
CGTSF_LEU0.6790.7840.7020.7091.1360.7210.7270.7250.721
CiK_Transport0.5320.5950.5840.5460.6080.6280.6690.5960.699

W2. Separating the Gains from Continued Pretraining and Event Conditioning.

A2: Thank you for raising the question of performance attribution. We agree that numerical continued pretraining improves forecasting performance. MM-Time aims to equip existing TSFMs with event-conditioned forecasting capabilities through multimodal continued pretraining. To assess the incremental value of our event-conditioning design, we compare against Numerical CPT and Vanilla Multimodal CPT. All CPT variants use the same numerical training data and backbone adaptation strategy.

Compared with Numerical CPT, MM-Time further reduces MASE from 0.746 to 0.716 on Chronos-2, from 0.766 to 0.739 on Falcon-2, and from 0.773 to 0.760 on TiRex-2. These results support the additional value of the proposed event-conditioning design beyond numerical continued pretraining.

Vanilla Multimodal CPT provides a further comparison of event-fusion designs. Incorporating event text through standard cross-attention outperforms Numerical CPT only on Falcon-2, while performing worse on Chronos-2 and TiRex-2. MM-Time outperforms both controls across all three backbones. Under this experimental setting, simply introducing event text does not consistently improve forecasting, whereas the proposed event representation and temporal alignment design enables more effective use of event information.

AuthorsOpenReview ↗

Response to Reviewer ewZv (Part II)

W3. Comparisons with ChronosX, Event Covariates, and LoRA.

A3: Thank you for highlighting these missing comparisons. We have added ChronosX and LoRA-based adaptations of Chronos-2 and TiRex-2 as additional baselines. These models are trained on the same continued-pretraining corpus and evaluated using the same protocol as MM-Time-C2. For the LoRA baselines, adaptation is performed on the continued-pretraining corpus, followed by zero-shot downstream evaluation without target-specific training.

We also evaluate the publicly available Chronos-2 and TiRex-2 checkpoints with event indicators supplied through their native covariate interfaces, without additional training. Specifically, we construct a binary covariate indicating whether each timestamp falls within an available event interval and provide its historical and future values to the models.

Results: The table below reports MASE on E-Commerce and Retail. MM-Time-C2 achieves the lowest MASE in both domains, while MM-Time-T2 outperforms both the native event-indicator and LoRA baselines built on TiRex-2. These results provide additional evidence for the effectiveness of MM-Time under the evaluated adaptation and inference settings. The comparison is reported in Section 4.3 and Table 3 of the revised paper ‾\underline{\text{Section 4.3 and Table 3 of the revised paper }}.

Evaluation scopeMM-Time-C2MM-Time-T2ChronosX (CPT + zero-shot inference)Chronos-2 (event-indicator covariate)TiRex-2 (event-indicator covariate)Chronos-2 + LoRATiRex-2 + LoRA
E-Commerce0.7170.7760.8310.9100.8470.7840.812
Retail0.7150.7460.9381.0270.7660.7400.752
Overall0.7160.7600.8890.9730.8010.7590.778

Dear Reviewer ewZv,

Thank you again for your detailed and constructive feedback. Since the End of author/reviewer discussions is just one day away, may we know if our response addresses your main concerns? If so, we kindly ask for your reconsideration of the score. Should you have any further advice on the paper and/or our rebuttal, please let us know, and we will be more than happy to engage in more discussion and paper improvements. Thank you so much for devoting time to improving our work!

Best regards,

The Authors

Comments on the paper

1 comment

Dear reviewers,

The authors have now submitted their responses. Please review them and share your thoughts on whether your concerns have been addressed. I encourage you to engage with one another’s assessments, discuss any remaining disagreements, and update your reviews and scores as appropriate.

Thank you.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.