acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Interpreting Time-series Foundation Models via Zero-shot Basis Decoding

Abstract

Time-series foundation models (TSFMs) generalize across diverse domains, enabling zero-shot forecasting. However, their forecasts do not explicitly identify the contribution of trends and periodic patterns, even though interpretation is a central objective of classical time-series analysis. We introduce PRISM, which decodes a frozen TSFM’s representation of the observed context into an additive forecast, fitting nothing to the target series. PRISM predicts a piecewise-linear trend, gates and weights a Fourier dictionary restricted to the periods the sampling interval and horizon admit, and adds a residual contribution, so that the forecast is a sum of components that can be inspected one at a time. On TimesFM 2.5, the components reproduce the calendar and physical structure that domain knowledge predicts in three domains, and where a cycle is present the seasonal components agree with a consensus of classical decompositions of the realized future. Where no cycle exists, as in exchange rates, the seasonal component stays near zero while the decomposers still report one. On strongly seasonal series under input noise, it revises less than the same bases fitted to the backbone’s forecast. Among component-wise forecasters, Ours has the lowest error on three of five dataset groups, with dataset-averaged MAE/σ 7.15% above the frozen backbone.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series foundation model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3hv1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PRISM makes forecasts from a frozen time-series foundation model more interpretable by replacing TimesFM 2.5’s forecasting head with three decoders whose outputs sum to the prediction: a two-knot piecewise-linear trend, a gated combination of predefined daily-to-yearly Fourier components, and an autoregressive residual. The authors validate PRISM through domain examples, agreement with classical decompositions, forecasting benchmarks, and noise tests.

Overall, the results support PRISM as a useful structured forecasting interface, but do not prove that it fully represent TimesFM’s internal reasoning due to bias in training and framework design.

Critical strengths

The paper propose a practically meaningful bridge between foundation-model forecasting and structured interpretability. PRISM converts a frozen TSFM representation into an exactly additive trend, seasonal, and residual forecast without fitting parameters to each evaluation series. While limiting the average accuracy degradation relative to TimesFM, and PRISM still outperforms the evaluated post-hoc explanable forecasters on three of five dataset groups.

Critical weaknesses
  • It is unclear whether PRISM interprets information genuinely encoded by the TSFM or primarily learns a new, author-designed forecasting model on top of it. The trend and seasonal meanings are imposed through a particular synthetic-data generator, a fixed calendar-period dictionary, sparsity and overlap penalties, and a sequential training schedule. These choices introduce substantial design bias into the resulting decomposition.

  • If PRISM is intended only as an intrinsically structured forecasting framework, the paper should not use strong language such as "reading components from the model itself" and "interpreting time-series foundation models".The authors should either provide faithfulness or causal evidence, or narrow the framing to structured forecasting from pretrained TSFM features.

  • Not sure if R is interpretable as autoregressive model is hard to understand.

  • The authors do not provide comprehensive verification against common sense R, T, and S across datasets, only several selected plausibility checks. Thus, we are not sure if the proposed framework is more interpretable than others.

Questions for authors
  • How stable are the decoded (T/S/R) components across decoder-training seeds, synthetic-data distributions, regularization strengths, and training orders?
  • Can PRISM be extended to other TSFMs such as Chronos or Morair?
  • What are the results of removing (T) or (S), analogous to the reported removal of (R)? How do (T+R), (S+R), (R)-only, and jointly trained variants perform?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3hv1 (1/2)

Dear Reviewer 3hv1,

We thank Reviewer 3hv1 for the valuable comments and suggestions. We appreciate your positive assessment of PRISM as a practical bridge between foundation-model forecasting and structured interpretability, without fitting parameters to each target series. Our responses below clarify the scope of our claims and address your questions about component validation, stability, and applicability to other backbones.


We have updated only the references to sections, tables, and line numbers to match the latest revision; the content of this response is unchanged.


[W1] Contribution of the pretrained backbone

PRISM's forecasting accuracy and decomposition depend on the pretrained backbone, not only on the designed decoders and their training (Appendix H). To separate its contribution from that of the designed heads and their training, we replace only the frozen backbone and train the same PRISM decoders on each backbone in the same way. Besides the pretrained TimesFM 2.5, we use the TimesFM 2.5 architecture trained from scratch, the same architecture with random weights, and two MLP backbones.

With the pretrained backbone, PRISM's error is closest to that of TimesFM 2.5 with its original forecasting head. Its MAE increase over this reference is 3.1%, 8.5–19.3 percentage points smaller than with the other backbones. In synthetic experiments that decompose T and S along three axes (trend, number of seasonal components, and amplitude ratio), the pretrained backbone consistently yields the highest decomposition accuracy.


[W2] Scope of the claims

PRISM adapts pretrained TSFMs for interpretable forecasting through basis decoding, and we narrowed the framing accordingly (revised in Section 1, lines 51–53, 69–70, and 74–75). The revision no longer uses the phrase "reading components from the model itself." Interpretability here refers to PRISM's own forecast components. This adaptation requires only decoder training, without backbone retraining or parameter fitting to individual evaluation series.


[W3] Role and interpretability of the residual decoder

R makes the forecast easier to interpret mainly by keeping T and S clean. The trend and seasonal decoders capture the overall structure of the series, leaving R to predict the remaining local variation. Unlike a forecasting head that outputs each horizon time point directly, R thus isolates this local variation. Where R's output is large, as with periodic fluctuations at periods outside the Fourier dictionary or predictable spikes, this narrows down what the pattern can be attributed to and supports further analysis. Figure 5 shows such a case on ILINet, where R recovers the outbreaks that T and S smooth over.


[W4] Interpretability compared with other decomposition models

PRISM is more interpretable than other decomposition models in terms of sparsity, regularity, and concurvity (Appendix F). We have added a quantitative analysis of these three criteria on the evaluation data, comparing PRISM with the decomposition models to which each criterion applies. Plots of PRISM's decomposition across the evaluation datasets and on synthetic data are included for qualitative inspection (Appendix F.4–F.6), and the decoder checkpoints we provide allow further checks on other series.

First, sparse bases make the periods behind a seasonal pattern explicit, so the forecast can be read in terms of specific daily or weekly cycles (Appendix F.1). Compared with N-BEATS-I, which also outputs its seasonal component as an explicit Fourier basis expansion, PRISM sets unnecessary coefficients to zero and explains its seasonal component with fewer bases at a similar MAE, concentrating it on a few of them (2.45 vs. 5.11 effective atoms).

Second, regular component curves are easier to read, because T and S are meant to summarize the forecast at the level of the whole trajectory (Appendix F.2). Even a component output as T or S must be followed point by point if its local fluctuations are large. PRISM keeps sample entropy low in both T and S, whereas the other models place short-interval fluctuations in T or S.

Third, low concurvity keeps each component responsible for its own part of the forecast (Appendix F.3). When a component can be approximated by the others, part of the forecast can move between components without changing their sum, which makes the split into T, S, and R ambiguous. PRISM shows the lowest correlation-based concurvity on every component pair. In addition, the variance of R projected onto the span of the T and S bases is only 5.2% of the variance of T+S, so R has only a minor influence on the component-level interpretation.

AuthorsOpenReview ↗

Response to Reviewer 3hv1 (2/2)

We have updated only the references to sections, tables, and line numbers to match the latest revision; the content of this response is unchanged.


[Q1] Stability of the decoded components

The decoded T, S, and R are stable across seeds, synthetic data distributions, and regularization coefficients. In contrast, the training order substantially increases their variation across runs (Appendix G.4). We retrain the decoders while varying one factor and fixing the others. The across-run standard deviation of each component's output is at most 0.044 context standard deviations for the first three factors and 0.139–0.408 for the training order. Changing the synthetic data distribution also leaves accuracy, sparsity, and output overlap nearly unchanged (Appendix G.3).


[Q2] Extension to other TSFMs

PRISM can be extended to Chronos-2 and Moirai 2.0 (Appendix E). We train PRISM on each frozen backbone in a setting similar to that described for TimesFM 2.5. On both backbones, the metrics of Sections 4.1–4.3 generally support the same claims as on TimesFM 2.5. These include PRISM's MSE relative to each model's own backbone (−4.1% on Chronos-2 and −3.9% on Moirai 2.0) and the change in MSE when R is removed (+4.3% and −0.2%).

TimesFM 2.5Chronos-2Moirai 2.0
Backbone MSE0.3090.3090.323
PRISM MSE0.3120.2960.311
vs. backbone+0.8%−4.1%−3.9%
MSE increase without R+1.8%+4.3%−0.2%

[Q3] Component removal and decoder subsets

Removing T or S from the forecast raises MAE by 7.4% and 75.5%, respectively, compared with 3.3% for removing R (Appendix Table 16).

The decoder subsets and joint training let R absorb structure that T and S can represent. Models that train only some of the decoders, with R last, stay close to PRISM in accuracy (MAE 0.335 for T+R, 0.340 for S+R, and 0.341 for R only; PRISM 0.332). Training all three decoders jointly from random initialization increases MAE to 0.350. R's variance share is 13.2% in PRISM. It rises to 23.5% without T, 94.9% without S, and 52.8% under joint training. The T and S decoders therefore limit how much of the forecast is assigned to the unconstrained R.


Please let us know if you have any further concerns.

Sincerely,
Authors

AuthorsOpenReview ↗

Respectful Inquiry Before Discussion Deadline

Dear Reviewer 3hv1,

Thank you for your time and thoughtful review of our work.

As the discussion period is approaching its end, we would like to kindly follow up on our previous response. We hope that our clarifications have adequately addressed your concerns, and we are willing to engage in further discussion if needed.

Sincerely,
Authors

Reviewer_HjG9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PRISM builds an additive forecasting model on a frozen TimesFM 2.5 encoder. Separate decoders predict a piecewise-linear trend, gated Fourier seasonal components, and an autoregressive residual; their sum forms PRISM’s forecast. The decoders are trained sequentially using real forecasting data and synthetic component supervision, with a separate decoder set for each horizon and no fitting to the target series at inference. Evaluation includes agreement with domain priors and classical decompositions, forecasting accuracy, and component stability under input noise. PRISM achieves the lowest forecasting error among the compared component-wise models on three of five dataset groups. Its seasonal components also show competitive agreement with statistical references and lower revision under noise than the evaluated post-hoc decompositions.

Critical strengths
  1. This work addresses a practical gap between general-purpose pretrained forecasting and explicit trend/seasonal outputs. The combination of structured basis decoders, synthetic component supervision, and staged residual training is coherent: the explicit components get a chance to explain the forecast before the unrestricted residual is trained. Producing these components directly from a frozen encoder, without optimizing parameters for each target series, could be useful when inspecting forecasts across heterogeneous series.

  2. PRISM is competitive with component-wise forecasting baselines and achieves the lowest error on three of five dataset groups. The evaluation goes beyond aggregate forecasting error to examine seasonal agreement, domain-specific patterns, and sensitivity to input perturbations. In particular, the building analysis extends beyond the two illustrated examples to all 70 eligible labelled buildings at the evaluated site. The detailed appendix also makes several qualifications explicit, including selected qualitative examples, training exposure in the trend-dominated analysis, and the distinction between correlation in shape and amplitude accuracy.

Critical weaknesses
  1. The evidence does not yet establish the semantic validity of the decomposition or the stronger interpretation claim. The authors explicitly state (in Section 3.1) that PRISM replaces the original forecasting head and is not constrained to reproduce its output. The readers therefore view the method primarily as a structured forecaster using pretrained representations, rather than a faithful explanation of the original TSFM’s predictions. Even under this narrower interpretation, component validity remains insufficiently established. Agreement with classical decompositions is useful evidence, but the references share calendar-period assumptions, and the phase-averaged Pearson metric does not assess amplitude recovery or contamination between components. The average correlation advantage over N-BEATS-I is small (0.841 versus 0.833), with N-BEATS-I obtaining higher correlations on five of seven datasets. Moreover, the unrestricted residual can absorb or counteract trend and seasonal structure; sequential training encourages a particular allocation but does not demonstrate that this allocation is semantically correct.

  2. The experiments and released artifacts do not sufficiently isolate and reproduce the contribution of representation-based decoding. The manuscript does not isolate the contributions of backbone pretraining, synthetic supervision, gating, and sequential training. Some qualitative comparisons also confound the decoding approach with basis capacity: PRISM uses 44 weekly Fourier pairs, including high-order harmonics, whereas the domain examples use default Prophet. There is also a concrete gap between the reported experiments and the accessible code release: the reproduction documentation states that decoder weights, paper results, and diagnostic experiments are not bundled, and that the current GIFT-Eval-only loader differs from the original training pipeline and is not promised to reproduce the reported scores.

Questions for authors
  1. Does the advantage in weekend-amplitude separation persist when Prophet or a regularized context-fit baseline receives a comparable Fourier dictionary or conditional seasonalities?

  2. Are the original decoder checkpoints and diagnostic scripts available in supplementary materials outside the repository?

3 comments
AuthorsOpenReview ↗

Response to Reviewer HjG9 (1/3)

Dear Reviewer HjG9,

We are grateful to Reviewer HjG9 for the careful reading and detailed feedback. We are pleased that you valued the combination of TSFM generalization with analysis of individual forecast components, together with analysis regarding quality of decomposition. We respond to your concerns and questions below.


[W1] Scope of the claims

PRISM adapts pretrained TSFMs for interpretable forecasting through basis decoding, as the revised framing now states (lines 49–53 and 72–73). Interpretability here refers to PRISM's own forecast components. This adaptation requires only decoder training, without backbone retraining or parameter fitting to individual evaluation series.


[W1] Amplitude recovery

PRISM generally recovers the seasonal amplitudes over the forecast horizon (Appendix F.7). Real data do not identify the true seasonal coefficients. We therefore test whether PRISM recovers the amplitudes of known periodic components under noise. We use the periodic signals of SynTSBench and place them on the harmonics of PRISM's hourly basis. We then add white Gaussian noise at a given signal-to-noise ratio (SNR), defined as the ratio of signal variance to noise variance. At 10 dB SNR, the noise variance is one tenth of the signal variance. At this SNR or higher, the relative amplitude error stays at or below 0.110 with up to 8 components. The coefficient R², which also reflects phase, is 0.976 or higher over the same range.

ComponentsNoise-free30 dB20 dB10 dB0 dB
10.0370.0410.0460.0530.096
20.0190.0230.0230.0340.086
30.0290.0350.0390.0530.149
50.0900.0920.1020.1090.168
80.0810.0800.0850.1100.264

[W1] Absorption and counteraction by the residual

In PRISM, R absorbs or counteracts the trend and seasonal structure only to a limited extent. The part of R that does so amounts to 5.2% of the variance of T+S (Appendix F.3). Absorbed or counteracted structure lies in the span of the T and S bases, so we measure this part by projecting R onto that span. R therefore has a minor influence on the component-level interpretation.

PRISM has the lowest correlation-based concurvity on every component pair, including the two pairs with R. High concurvity means that one component can be approximated by the others, so part of the forecast can shift between them without changing the sum, and the split into T, S, and R becomes ambiguous. Low concurvity thus limits this ambiguity.


[W1] Semantic validity

In the revision, PRISM's mean seasonal correlation is 0.874, compared with 0.849 for N-BEATS-I (Section 4.2). N-BEATS-I remains higher on four of the seven datasets, by at most 0.015, whereas PRISM leads by up to 0.158 (ETTh2).

Additional quantitative analyses show that PRISM's components are more interpretable than those of other decomposition models in terms of sparsity and regularity (Appendix F.1 and F.2). A sparse basis makes explicit which periods constitute the seasonal component, so the forecast can be read in terms of specific daily or weekly cycles. PRISM explains its seasonal component with fewer Fourier atoms than N-BEATS-I at a similar MAE (2.45 vs. 5.11 effective atoms).

We measure the regularity of T and S with sample entropy, which is lower when patterns in a curve repeat more consistently. Regular T and S summarize the direction and repeated shape of the forecast. An irregular curve must instead be read point by point. PRISM keeps sample entropy low in both T and S, while the other models place short-interval fluctuations in T or S.

Decomposition plots across the evaluation datasets and against synthetic ground truth are included for qualitative inspection (Appendix F.4–F.6), and the decoder checkpoints we provide allow further checks on other series.

AuthorsOpenReview ↗

Response to Reviewer HjG9 (2/3)

[W2] Method ablations

Each of the four factors serves its intended role in PRISM, as shown by the new ablations in Appendix H (backbone pretraining) and Appendix G (synthetic supervision, gating, and sequential training).

The pretrained backbone contributes to both PRISM's forecasting accuracy and its decomposition (Appendix H). To separate its contribution from that of the designed heads and their training, we replace only the frozen backbone and train the same PRISM decoders on each backbone in the same way. Besides the pretrained TimesFM 2.5, we use the TimesFM 2.5 architecture trained from scratch, the same architecture with random weights, and two MLP backbones. With the pretrained backbone, PRISM's error is closest to that of TimesFM 2.5 with its original forecasting head. Its MAE increase over this reference is 3.1%, 8.5–19.3 percentage points smaller than with the other backbones. In synthetic experiments that decompose T and S along three axes (trend, number of seasonal components, and amplitude ratio), the pretrained backbone consistently yields the highest decomposition accuracy. The designed decoders thus fix the form of each component, while the pretrained backbone contributes to how accurately PRISM forecasts and recovers these components.

Synthetic supervision affects the sparsity of the seasonal component on trend-dominated datasets (Appendix G.2). Without it, sparsity changes little on the seasonal-dominated datasets, whereas on the trend-dominated datasets of Table 3 the numbers of active and effective atoms both increase. This result indicates that synthetic supervision acts on the division between the trend and seasonal components. Synthetic data themselves are also needed: removing them raises the R-in-span ratio from 5.8–6.1% to 22.7%, so R overlaps more with T and S in the output, whereas changing their distribution has little effect (Appendix G.3).

Gating determines how many Fourier atoms the seasonal component uses (Appendix G.2). Once gating is removed, every admissible atom becomes active on the seasonal- and trend-dominated datasets alike. With gating, PRISM explains the seasonal component with a sparse basis while maintaining its forecasting accuracy (MAE 0.332 vs. 0.335).

The training order affects whether the unconstrained R absorbs structure that T and S can represent (Appendix G.1). When R is trained before S or jointly with the other decoders, the variance share of R rises from 13.2% to 84.2–85.9% and 52.8%, respectively. In the same settings, more of R falls within the span of the T and S bases, and the concurvity between components increases. After R is trained first, the remainder is too irregular for T and S to learn, and S then degrades rather than improves the forecast. Training R last, as PRISM does, therefore confines R to what T and S leave unexplained.


[W2] Reproducibility of the reported experiments

We have added the revised decoder checkpoints and the diagnostic scripts to the anonymized repository linked in the Reproducibility statement. We will release more detailed implementation after publication.

AuthorsOpenReview ↗

Response to Reviewer HjG9 (3/3)

[W2, Q1] Prophet comparison and seasonal basis capacity

PRISM keeps the smallest worst-case gap to the observed weekday/weekend separation when Prophet or a regularized context fit receives PRISM's Fourier dictionary or conditional seasonality. All four strengthened baselines recover the weekday/weekend pattern far better than Prophet with library defaults, but PRISM is the only one that stays close to the observed separation whether the pattern is present or absent.

We add four baselines: Prophet with PRISM's dictionary, Prophet with conditional seasonality, Prophet with both, and a ridge fit on the same dictionary with oracle lambda per series. Separation is measured by η2\eta^2, the share of variation among the seven predicted day-of-week profiles explained by grouping the days into weekdays and weekends; the table reports its median (%) per group on the same 512+336 window, against the same quantity from each series' whole observed record.

PRISM's largest difference from the observation is 11.9 points, less than half that of any other method. Prophet with conditional seasonality is closest on traffic but separates lodging as strongly as education (0.8 against an observed 52.1), imposing the split whether or not one exists.

MethodWeekly pattern present: Traffic (862)Weekly pattern present: Electricity, education (66)Weekly pattern absent: Electricity, lodging (4)education − lodginglargest difference
Observation87.6%95.8%43.7%52.1
PRISM81.5%94.5%32.2%62.311.5
Prophet, library defaults11.2%3.6%8.0%−4.492.2
Prophet, conditional seasonality (default weekly order)91.7%88.1%87.3%0.843.6
Prophet, matched dictionary61.1%87.8%35.4%52.526.5
Prophet, matched dictionary + conditional seasonality59.0%85.4%29.9%55.528.6
Ridge context fit (oracle lambda)61.7%87.3%41.8%45.525.9

[Q2] Decoder checkpoints and diagnostic scripts

The decoder checkpoints and diagnostic scripts are in the anonymized repository, as described in our response to W2.


Please let us know if you have any further concerns.

Sincerely,
Authors

Reviewer_e3o9OpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The authors propose PRISM, a method where a time series model is augmented with 3 decoder heads: a trend head, a seasonal head, which is expressed through Fourier bases, and a residual head. These are linearly added back together to produce the final forecast.

Critical strengths

1. Components built into the forecast. I quite like the idea; it reminds me of N-BEATS and is really versatile. Especially forcing the model to use these components not only allows for readability but also makes sure that the heads use them in the computation, so it also allows for targeted steering.

2. Transfer. The authors showcase strong transfer results. One decoder set per horizon is applied zero-shot and still has the lowest error among the component-wise forecasters on three of the five dataset groups (Table 1), although N-BEATS-I, DLinear, ETSformer, and Autoformer are trained on each target dataset. The components also transfer to unseen data: across all 70 labelled buildings of a site not used in decoder training, the predicted weekend amplitude separates education from lodging buildings while Prophet's does not.

Critical weaknesses

1. Comparison and evaluation protocol. I think it is a bit suboptimal that the authors compare mostly against supervised methods, while they themselves use a foundation model. N-BEATS-I, DLinear, ETSformer, and Autoformer are trained on each target dataset, Prophet is fitted to each window, and the only other pretrained model, TEMPO, has Electricity and ETTm in its pretraining corpus. PRISM is only instantiated on TimesFM 2.5, and no other TSFM, such as Chronos or Moirai, which the introduction names, appears as a backbone or a baseline. An ablation that tries out their method on different models would strengthen their method quite drastically.

Moreover, I observed that PRISM is trained with MAE losses and its checkpoint is selected by MAE, while the trained baselines minimize the MSE and are selected by validation MSE, yet only the MAE is reported. To be honest, this sounds a bit strange to me: the baselines are trained and selected on a different metric than the one they are evaluated on, which favors PRISM, and the MSE is usually reported alongside the MAE on these benchmarks. Without this being clarified, I recommend rejection for this reason, but I am happy to raise my score if this gets clarified, although my concerns in Weakness 2 would remain.

2. No semantic alignment of the heads. Now, my major complaint is that there is no real semantic alignment for the heads: apart from the synthetic series, whose contribution is not ablated, nobody is really telling the model that it should really learn those concepts, or at least it is not shown. The main quantitative component evaluation measures agreement in seasonal shape with a consensus of classical decompositions. This is useful evidence, but correlation does not establish that seasonal amplitudes are correct, as the authors note themselves. Averaging across cycles also reduces sensitivity to changes in seasonal behavior. In addition, the margin over N-BEATS-I is small. Furthermore, the trend and residual get even less evidence: the trend is only evaluated through the total forecast error on trend-dominated datasets, two of which (M4 Daily and M4 Weekly) were used in decoder training, and the residual only qualitatively and through one ablation. While the authors try to circumvent this with sequential training, there is no real mathematical guarantee, and the residual decoder can still absorb variation represented by the trend and seasonal components.

The paper asks: "can the components of a TSFM forecast be read from the model itself, at inference time and without fitting to the target series, and at what cost in accuracy?". I highly disagree with this statement, and the authors do not answer it with PRISM. The only answer is that they can train heads based on a frozen representation, not that the original model is using any of this internally. PRISM replaces TimesFM's forecasting head, and its forecast "is not constrained to reproduce the original head's", so the components describe PRISM's forecast, not TimesFM's.

Questions for authors
  1. I am a bit confused by Table 1: does your method improve upon TimesFM or worsen it? Making your method bold, while TimesFM, which has a lower error in every row, is excluded from the ranking, does not make any sense to me; Table 1 also only reports the differences to TimesFM, not its own errors. If this reconstruction improves models in any way, it would be great to show this on another foundation model like Chronos.
  2. As most foundation models are probabilistic nowadays, would your framework also work there? TimesFM 2.5 itself provides quantile forecasts through an optional quantile head, whereas PRISM only produces point forecasts (Section 3).
  3. Is there a reason why you only report the MAE? Does the MSE penalize your method's outliers too strongly?
  4. How much does the pretrained representation contribute? Can you provide evidence separating the benefit of the frozen TSFM from that of the structured decoders and the synthetic component supervision? In particular, how would the same component-prediction approach perform with a simpler encoder trained directly on the observed history?

Suggestions for presentation

  • "Ours" is used as a model name in the abstract and the running text, but you could also use "PRISM"
  • Line 092: a space is missing in "exists.Where".
  • The average increase over TimesFM is reported as 7.15% and as 7.1%
5 comments
AuthorsOpenReview ↗

Response to Reviewer e3o9 (1/3)

Dear Reviewer e3o9,

We thank Reviewer e3o9 for the thoughtful review and constructive suggestions. We appreciate your recognition of the ability to inspect and steer forecasts at the level of basis functions and coefficients, as well as the transfer of these components to unseen data. Below, we address your concerns about the evaluation protocol and component validity, along with your questions.


[W1-1] Comparison with and extension to other TSFMs

PRISM can be extended to Chronos-2 and Moirai 2.0 (Appendix E, Table 10). We train PRISM on each frozen backbone in a setting similar to that described for TimesFM 2.5. On both backbones, the metrics of Sections 4.1–4.3 generally support the same claims as on TimesFM 2.5. These include PRISM's error relative to each model's own backbone (MSE −4.1% and −3.9%, MAE +2.9% and +4.3% on Chronos-2 and Moirai 2.0) and the change in MSE when R is removed (+4.3% and −0.2%).

TimesFM 2.5Chronos-2Moirai 2.0
Backbone MSE0.3090.3090.323
PRISM MSE0.3120.2960.311
vs. backbone+0.8%−4.1%−3.9%
MSE increase without R+1.8%+4.3%−0.2%

The 0.1 percentage-point difference between the TimesFM 2.5 column (+0.8% in MSE) and Section 4.3 (+0.9%) arises because Appendix E evaluates on a subset of the test windows, as stated in Appendix D.1 (lines 1186–1189); we will update Appendix Table 10 accordingly.

We add a comparison axis that clarifies PRISM's accuracy relative to pretrained TSFMs. Most forecasting models that provide a decomposition are trained or fine-tuned on the target data, so they offer only a limited comparison for zero-shot PRISM. The revised Appendix Table 10 adds this axis. It compares PRISM with three pretrained TSFMs, TimesFM 2.5, Chronos-2, and Moirai 2.0, by building PRISM on each of them as the backbone. For each dataset in Table 1, it reports the MAE and MSE of each TSFM and of PRISM on that TSFM.


[W1-2] Training and selection metric of the baselines

In the revision, the baselines are trained and selected on the metric used for evaluation, and we report both MAE and MSE (Appendix D.1). The MAE results (Table 1) use the MAE-trained models and the MSE results (Appendix Table 9) use the MSE-trained ones; Appendix Table 9 also includes the results of PRISM trained with an MSE forecast loss for reference. All errors are normalized by each channel's training-split standard deviation following the evaluation protocol of MOMENT.

AuthorsOpenReview ↗

Response to Reviewer e3o9 (2/3)

[W2-1] Semantic alignment of the decoders

Trend decoder

PRISM's trend output makes the global direction of the forecast easy to read (Appendix F.2). We evaluate this on T itself, not only through the total forecast error. T is meant to summarize the whole trajectory, so a T with large local fluctuations would have to be read point by point. We measure the regularity of T with sample entropy, which is lower when the curve evolves more predictably, with fewer short-interval fluctuations. PRISM keeps the sample entropy of T low, whereas TEMPO, DLinear, and Autoformer place short-interval fluctuations in T.

We train the decoders again without M4 Daily and M4 Weekly in the training data (Appendix A.2). PRISM still has lower error than TimesFM 2.5 on M4 Weekly and M4 Daily (−11.6% and −5.3%; Table 3 in Section 4.4).

Seasonal decoder

PRISM generally recovers the seasonal amplitudes over the forecast horizon (Appendix F.7). The correlation with classical decompositions in Section 4.2 reflects the seasonal shape but not its amplitude, and the true seasonal coefficients are generally unknown in real data. We therefore test whether PRISM recovers the amplitudes of known periodic components under noise. We place the periodic signals of SynTSBench on the harmonics of PRISM's hourly basis and add white Gaussian noise at a given signal-to-noise ratio (SNR). The SNR is the ratio of signal variance to noise variance; at 10 dB, the noise variance is one tenth of the signal variance. At this SNR or higher, the relative amplitude error stays at or below 0.110 with up to 8 components. Over the same range, the coefficient R², which also reflects phase, is 0.976 or higher.

ComponentsNoise-free30 dB20 dB10 dB0 dB
10.0370.0410.0460.0530.096
20.0190.0230.0230.0340.086
30.0290.0350.0390.0530.149
50.0900.0920.1020.1090.168
80.0810.0800.0850.1100.264

In the revision, PRISM's mean seasonal correlation with the classical references is 0.874, compared with 0.849 for N-BEATS-I (Section 4.2). N-BEATS-I remains higher on four of the seven datasets, by at most 0.015, whereas PRISM leads by up to 0.158 (ETTh2).

Residual decoder

In PRISM, R absorbs or counteracts the trend and seasonal structure only to a limited extent (Appendix F.3). Such structure lies in the span of the T and S bases, so we project R onto that span. The projected part amounts to 5.2% of the variance of T+S. R therefore has only a minor influence on the component-level interpretation.

PRISM also has the lowest correlation-based concurvity on every component pair, including the two pairs with R. High concurvity means that one component can be approximated by the others. Part of the forecast can then shift between components without changing their sum, which makes the split into T, S, and R ambiguous. Low concurvity limits this ambiguity.

Sequential training keeps R this small, as the training-order ablation shows empirically (Appendix G.1). When R is trained before S, its variance share rises from 13.2% to 84.2–85.9%. When the three decoders are trained jointly, it rises to 52.8%. In both settings, more of R falls within the span of the T and S bases, and the concurvity between components increases. When R is trained first, the remainder is too irregular for T and S to learn, and S then degrades rather than improves the forecast. Training R last, as PRISM does, confines R to what T and S leave unexplained.

Overall

Decomposition plots across the evaluation datasets and synthetic ground truth are included for qualitative inspection (Appendix F.4–F.6), and the decoder checkpoints we provide allow further checks on other series.

We address the contribution of the synthetic series, including its ablation, in [Q4].


[W2-2] Scope of the claims

PRISM adapts pretrained TSFMs for interpretable forecasting through basis decoding, and we revised the research question accordingly (lines 49–53 and 72–73): it now asks whether a frozen TSFM can be adapted to forecast through explicit components by training only decoders on its representation. Interpretability here refers to PRISM's own forecast components, rather than a faithful reproduction of TimesFM's forecast. This adaptation requires only decoder training, without backbone retraining or parameter fitting to individual evaluation series.

AuthorsOpenReview ↗

Response to Reviewer e3o9 (3/3)

[Q1] Accuracy relative to the backbone

On TimesFM 2.5, PRISM slightly worsens accuracy, by 3.1% in MAE and 0.9% in MSE (Section 4.3), but on other backbones it can improve MSE by global prediction. Applied to Chronos-2 and Moirai 2.0, it lowers the average MSE over the seven datasets by 4.1% and 3.9% relative to each backbone, while its MAE is 2.9% and 4.3% higher (Appendix Table 10). Through its bases, PRISM forecasts T and S over the entire horizon at once. On volatile series at long horizons, the backbones tend to overshoot the level, trend, or amplitude. The globally predicted T and S instead extract stable patterns. On ETTm, the MSE reduction reaches 9.4% on Chronos-2 and 12.9% on Moirai 2.0. This behavior resembles that on the trend-dominated series of Section 4.4, where PRISM reduces the error of TimesFM 2.5.

These gains may not carry over to other datasets or backbones. PRISM is designed to preserve the backbone's generalization performance under the strong structural constraint of its basis outputs, rather than to improve it. PRISM requires no large-scale pretraining and trains only lightweight decoders on the frozen backbone. The decoders inherit the backbone's generalization and steer forecasts from the unconstrained output space RH\mathbb{R}^H toward structured basis expansions.

Because the backbone stays frozen, PRISM's accuracy largely follows that of TimesFM 2.5. In Table 1, a trained baseline outperforms PRISM on four dataset groups. On three of them (Traffic, Solar-Energy, and ETTm), the same baseline also outperforms TimesFM 2.5 itself. We therefore organize Table 1 around the gap to the backbone. The revised table reports TimesFM 2.5's own errors in an unranked reference column.


[Q2] Probabilistic forecasting

PRISM is, in principle, applicable to probabilistic TSFMs. Basis decoding could be applied both to models that sample a different forecast at each generation and, as the reviewer notes, to models that provide quantile forecasts alongside the point forecast. We regard this extension as an interesting direction for future work.


[Q3] Reporting MSE

We report MSE alongside MAE, as described in [W1-2]. MSE does not penalize PRISM: its gap to TimesFM is smaller in MSE than in MAE. Averaged over the seven datasets, PRISM's error exceeds that of TimesFM by 0.9% in MSE (0.308 vs. 0.305) and by 3.1% in MAE (0.331 vs. 0.321; Section 4.3). Its average MSE is comparable to that of the best baseline, DLinear (0.308 vs. 0.309).


[Q4] Contributions of the pretrained backbone and synthetic supervision

The pretrained backbone contributes to both PRISM's forecasting accuracy and its decomposition (Appendix H). To separate its contribution from that of the designed heads and their training, we replace only the backbone and train the same PRISM decoders on each backbone in the same way. Besides the pretrained TimesFM 2.5, we use the TimesFM 2.5 architecture trained from scratch, the same architecture with random weights, and two MLP backbones.

With the pretrained backbone, PRISM's error is closest to that of TimesFM 2.5 with its original forecasting head. Its MAE increase over TimesFM 2.5 with its original head is 3.1%, compared with 11.6% when the same architecture is trained from scratch and 15.7–22.4% with the other backbones. In synthetic experiments that decompose T and S along three axes, the pretrained backbone consistently yields the highest decomposition accuracy. As a result, the pretrained backbone contributes to how accurately PRISM forecasts and recovers these components.

Synthetic supervision affects the sparsity of the seasonal component on trend-dominated datasets (Appendix G.2). Without it, sparsity changes little on the seasonal-dominated datasets, whereas on the trend-dominated datasets of Table 3 the numbers of active and effective atoms both increase. This result indicates that synthetic supervision acts on the division between the trend and seasonal components. Synthetic data themselves are also needed: removing them raises the R-in-span ratio from 5.8–6.1% across the tested synthetic distributions to 22.7%, so R overlaps more with T and S in the output, whereas changing their distribution has little effect (Appendix G.3).

We have also extended our analysis to the effects of the other factors used in decoder training. With the backbone fixed, Appendix G ablates these factors: the training order, gating, the synthetic data distribution, and stability across seeds, synthetic data distributions, and regularization strengths.


Suggestions for presentation

Thank you for these careful corrections; we have applied all three throughout the text of the revision.


Please let us know if you have any further concerns.

Sincerely,
Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We sincerely thank all reviewers for their valuable, professional, and constructive comments.

We respond to each reviewer individually under the corresponding review, and summarize below the main changes, several of which address concerns shared by reviewers. Given the feedback, we have made the following changes to the paper (colored in red for the first revision and in blue for the second revision):

  1. (e3o9, 3hv1) Extended PRISM to two further frozen TSFMs, Chronos-2 and Moirai 2.0, to show its applicability beyond TimesFM 2.5 (Appendix E‾\underline{\text{Appendix E}}).
  2. (e3o9, HjG9, 3hv1) Strengthened the interpretability analysis with quantitative comparisons to other decomposition models (Appendix F‾\underline{\text{Appendix F}}).
  3. (e3o9, HjG9, 3hv1) Clarified the contribution of the pretrained backbone to PRISM through an ablation study (Appendix H‾\underline{\text{Appendix H}}).
  4. (e3o9, HjG9, 3hv1) Added ablations of the training design, covering training order, gating, synthetic supervision, and stability, to clarify the role of each design choice (Appendix G‾\underline{\text{Appendix G}}).
  5. (e3o9) Aligned the accuracy evaluation with a standard protocol (Section 4.3‾\underline{\text{Section 4.3}}, Appendix D‾\underline{\text{Appendix D}}).
  6. (e3o9, HjG9, 3hv1) Refined the title and the research question to state more clearly that PRISM adapts frozen TSFMs for interpretable forecasting (Section 1‾\underline{\text{Section 1}}).
  7. (e3o9, HjG9) Improved reproducibility by releasing the checkpoints and diagnostic scripts, and updated all materials (Section 4‾\underline{\text{Section 4}}, Appendices A–D‾\underline{\text{Appendices A–D}}).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.