acceptodds
Under review as a conference paper at ICLR 2027

Time-: A Time Series Action Model for Decision-Making

22 views

Abstract

Time series foundation models (TSFMs) have recently demonstrated a strong capability in modeling temporal dynamics, yet their potential for real-world operational decision-making remains largely unexplored. In this work, we formulate time series decision-making as the task of generating actions based on historical observations of an exogenous time series and the current state of a system. We introduce Time-, a simple yet effective framework that adapts pretrained TSFMs for direct action generation. Specifically, a TSFM encodes temporal history into context tokens, while a lightweight flow-matching action expert combines these representations with the system state to generate discrete action chunks. We further construct , a testbed comprising five real-world operational systems across energy, finance, telecommunication, supply chains and green industries. Each system couples an action-independent exogenous time series with a controllable system state. Experiments under single-chunk and rollout evaluations show that Time- achieves strong performance compared with forecast-then-optimize and direct offline policy baselines. Overall, our findings lay the groundwork for expanding the utility of TSFMs, shifting their paradigm from standard forecasting to actionable, real-world decision-making. The code and testbed are available.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Foundation Model, Decision-Making, Action Generation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_N2TLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework that adapts pretrained time series foundation models (TSFMs) for direct action generation. It combines a TSFM encoder with a flow-matching action expert to generate discrete action chunks conditioned on historical time series and current system states. The experimental results show that their model achieves competitive performance against forecasting-based optimization and offline policy learning baselines.

Critical strengths
  • Extending pretrained TSFMs from forecasting to direct decision-making is interesting.

  • Combining pretrained temporal representations with state-conditioned action generation is sound. The technical design and implementation is solid.

  • The authors design diverse evaluation tasks and a closed-loop evaluation. The proposed Cassandra covers five operational domains, including energy, finance, telecommunications, and supply chains, with comparisons against MPC, offline RL, and imitation learning methods.

  • This paper is well-written and easy to follow.

Critical weaknesses
  • My main concern lies in that the final outputs are discrete actions, with only 2–4 choices in several tasks. It is unclear whether flow matching offers meaningful advantages over simpler classification or autoregressive action prediction. I recognize the new setting and end-to-end decision design of time series action models, but it would be better to have a clearer motivation.

  • Are there some scenarios that same historical time series have different action labels? For hindsight supervision, the oracle generates optimal actions using realized future time series, while the learned policy only observes historical information. This information asymmetry may introduce ambiguous or conflicting action labels under future uncertainty.

  • The rollout evaluation is limited to three action chunks, which is insufficient to assess long-term decision quality, compounding errors, and robustness to distribution shifts.

Questions for authors
  1. How does the method handle conflicting hindsight-optimal actions arising from different future realizations under similar historical observations?

  2. Can the authors evaluate longer closed-loop rollouts and receding-horizon execution to investigate error accumulation and long-term decision performance?

  3. Can the authors compare the flow-matching action expert with simpler decision heads, including autoregressive classifiers or recent decision models, under the same TSFM backbone and training data?

Reviewer_TeoCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Time-(\pi_0), a framework for adapting a pre-trained TSFM)to operational decision-making. A LoRA adapted TSFM encodes historical observations into temporal context tokens, which condition a flow matching policy that generates an action chunk conditioned on the current system state. Training labels are action sequences computed by DP. The paper also introduces Cassandra, a benchmark of five semi-synthetic operational control environments based on real exogenous time series.

Critical strengths
  • The paper is well structured and generally easy to follow. The motivation is relevant and timely.

  • The proposed architecture is reasonably simple and modular. In principle, the action-generation component can be paired with any TSFM backbones, as supported by the backbone experiments.

  • Reporting both action accuracy and realized normalized reward is appropriate. As the paper notes, action agreement with an oracle does not necessarily imply better realized return. Evaluating both independent single-chunk predictions and multi-round rollouts is also useful.

Critical weaknesses
  • The conceptual difference of proposed work with the "forecast-then-optimize" framework remains somewhat unclear to me. Although Time-π0\pi_0 does not decode an explicit future trajectory or run a numerical optimizer at inference time, its TSFM representations are trained with forecasting supervision and may be interpreted as latent forecasts. The flow-matching policy then maps these predictive representations and the state to actions, plausibly acting as an amortized optimizer distilled from the hindsight DP oracle.

  • The ablations studies are informative but somewhat incomplete. The carbon dataset is missing. No details about horizon length, multiple seeds variability, and hyperparamers in general of ablated models are given. Moreover, to better assess proposed method gains, the "w/o TSFM" should use a better capacity model than a simple linear layer (with at least the same number of trainable parameters as Time-π0\pi_0); a version in which direct, fine-tuned, forecasts are given to the action policy would also empirically emphasized the differences with the "forecast-then-optimize" paradigm, pointed out in first comment. In general, the advantage of TSFM's does not clearly appear through current experiments.

  • Given public TSFM pre-training, how is possible contamination from the evaluation series handled?

  • Providing the raw rewards or some cumulative regret based on RDPR_{DP} could give interesting additional, interpretable, informations about models & baselines performances.

  • How is the accuracy metric displayed actually computed? In particular, is there any optimal actions ties cases within the benchmark? If so, how are those handled ?

Questions for authors

see weaknesses above

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

This paper introduces a method for adapting pretrained time-series foundation models (TSFMs) to operational decision-making. The problem is formulated as predicting a chunk of discrete actions from the history of an exogenous time series and the current state of a controllable system. A LoRA-adapted TSFM converts the temporal history into contextual tokens, while a conditional flow-matching Diffusion Transformer produces continuous action latents that are decoded into discrete actions. Training uses action chunks computed by a hindsight dynamic-programming (DP) oracle, together with flow-matching, action cross-entropy, and auxiliary forecasting losses. The authors also introduce Cassandra, a benchmark pairing real observed series with simulated operational dynamics for five settings: battery arbitrage, stock trading, base-station mode switching, inventory replenishment, and carbon-aware inference scheduling. Experiments compare Time-π₀ with heuristics, forecast-then-optimize/MPC, offline RL, behavior cloning, and Decision Transformer in single-chunk and three-chunk rollout evaluations. The method is competitive in several tasks and performs particularly well on battery arbitrage, but is less compelling in inventory replenishment.

Critical strengths
  1. The paper explicitly separates an action-independent exogenous sequence from an action-dependent operational state (Section 3.1), which captures an important pattern in energy, inventory, and scheduling problems. Cassandra implements this formulation across five substantially different discrete-action tasks with explicit transition/reward rules (Section 4 and Appendix B). This is a meaningful contribution because it facilitates comparisons that go beyond pure forecasting accuracy.

  2. The results are not limited to matching clairvoyant oracle actions: Tables 2–3 also evaluate reward under realized exogenous sequences. For example, in the three-round rollout table, Time-π₀ attains horizon-averaged normalized rewards of 78.8 on battery and 70.9 on telecom, compared with 69.4 and 67.1 for Decision Transformer, respectively (Table 3). The paper appropriately notes that action accuracy and reward need not track each other, and investigates different TSFM backbones (Figure 4). These experiments suggest that direct action prediction is a worthwhile research direction.

Critical weaknesses
  1. The proposed method benefits from a pretrained TSFM adapted with LoRA and an auxiliary future-forecasting loss, whereas the MPC-TSFM baseline uses zero-shot Chronos-2 forecasts (Sections 3.2, 5.1 and Appendix A.1). The MPC baselines use an ensemble of learned transition/reward models (Appendix A.1), even though the Cassandra environments define explicit dynamics that are used to construct the hindsight DP oracle (Section 4.2 and Appendix B). This introduces dynamics-model error into the forecast-then-optimize competitors and makes it difficult to attribute gains to direct actions rather than forecasts plus optimization. An MPC baseline using the known simulator dynamics, with a comparably adapted forecasting model, is a particularly important missing control. Furthermore, the Behavior Cloning baseline is a smaller MLP flow policy rather than an architecture/parameter-matched action expert, and the “w/o TSFM” ablation replaces the foundation model with a linear layer plus normalization (Table 4), not a comparable randomly initialized temporal backbone. Thus, the current evidence cannot convincingly distinguish benefits from TSFM pretraining, architecture capacity, auxiliary forecasting supervision, and baseline implementation choices. This is consequential because the main claim is about the distinctive value of pretrained TSFM representations for decision-making.

  2. The training target is an action sequence optimized using the realized future exogenous trajectory (Eq. 1 and Section 4.2), while the policy observes only past exogenous inputs and the current state. Minimizing disagreement with those oracle labels need not maximize expected deployment reward: a low-probability, high-payoff action can be optimal in expectation even when it is rarely the hindsight-optimal label. The authors acknowledge the distinction between accuracy and reward, but do not investigate the underlying statistical mismatch, its impact under uncertainty or distribution shift, or regret relative to an appropriate non-clairvoyant decision benchmark. The reported closed-loop evaluations span only three chunks (Section 5.1), and Appendix Table 6 shows as few as five three-round evaluation chains per seed for economy and supply, limiting evidence for long-horizon robustness. Moreover, these tasks pair real exogenous data with specified simulators rather than demonstrating performance in a deployed operational system. This matters because the paper motivates the method as actionable real-world decision-making, for which performance of a causal policy—not fidelity to a future-informed teacher—is the decisive claim.

  3. On the supply task, Time-π₀ obtains horizon-averaged rollout normalized reward 75.0, below the heuristic (89.7), MPC-native (87.5), and DiscreteCQL (91.8) in Table 3. In economy, its average rollout normalized reward 12.8 is below DiscreteCQL's 19.3. These counterexamples should receive more emphasis when stating the overall empirical conclusion.

Questions for authors
  1. Can you report MPC with the exact environment dynamics/rewards from Appendix B rather than learned transition models, using both the existing zero-shot forecast and a target-domain-adapted TSFM forecaster? How much of Time-π₀'s gain remains when the forecast model and system knowledge are comparable?

  2. Can you compare against (i) the same TSFM architecture trained from scratch rather than pretrained, (ii) an architecture/parameter-matched action expert with a simpler history encoder, and ideally (iii) the same action expert with/without the state-conditioned prior or flow matching? Please clarify Table 4's horizon/seed/evaluation protocol and why carbon is absent from that ablation. Such comparisons would more directly support the claimed source of improvement.

  3. What justification or empirical evidence supports using exact imitation of a future-informed DP oracle as a surrogate for maximizing expected reward given only historical observations? Could you evaluate performance against a causal policy/optimizer tuned to expected return (or report regret under synthetic distributions with controllable forecasting uncertainty), rather than relying mainly on hindsight-normalized reward?

  4. Can you evaluate longer rollouts and more independent test trajectories, or explain why only three chunks and as few as five evaluation chains (Table 6) suffice to support the robustness claim? Please report variation over episode selection and action-sampling randomness, not only training seeds, if feasible.

  5. Appendix C's Telecom description refers to segment routing and link utilization, whereas Appendix B.4 describes base-station modes and switching penalties. Which environment generated the results and Figure 5's Telecom panels, and can you reconcile the inconsistent descriptions? This clarification would help establish that the evaluated benchmark is correctly documented.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.