acceptodds
Under review as a conference paper at ICLR 2027

Hidden States Are Better Keys for Retrieval-Augmented Time Series Forecasting

Abstract

Time series foundation models (TSFMs) deliver high-quality zero-shot forecasts on a target dataset. Unlike a conventional dataset-specific model, however, a TSFM can use only the part of the target series' history that fits in its input window, while the history outside the window is rich in periodic structure and statistical regularities. Retrieval-augmented forecasting can exploit that history without training the model, but existing methods compute similarity on raw series or with a separate embedding model: they measure how closely a historical segment's context matches in shape, whereas what enters the forecast is the segment's continuation, so the quantity the key measures is misaligned with the quantity the forecast needs. We argue that a retrieval key should measure predictive similarity, and that the hidden state of a frozen TSFM has exactly this property: used as the key, it retrieves continuations whose error is 14% lower than with the raw series and 17% lower than with a separate embedding model. Building on this, we propose HiRAF (Hidden-state Retrieval-Augmented Forecasting), which reads the hidden state from the forward pass that produces the forecast, retrieves the most similar segments from the target series' history with it, and uses their continuations to correct the backbone's forecast. No model is trained; the key is a by-product of the forward pass that produces the forecast, so no separate embedding model is needed, and with a bucketed index each retrieval scans only a small fraction of the history. On six benchmark datasets HiRAF achieves better zero-shot accuracy than existing TSFMs and 7% lower MSE than the retrieval-augmented method TS-RAG; with an index of five million entries, its prediction time is 1.05× that of the backbone alone, against 8.17× for TS-RAG.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Time Series Forecasting, Retrieval-Augmented

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_u8eBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HiRAF, a retrieval-augmented forecasting method that uses a frozen time-series foundation model’s hidden states to retrieve historical segments with useful future continuations. It averages the retrieved continuations after adjusting their level and scale, then blends them with the backbone forecast using a validation-fitted scalar. Across six datasets, HiRAF improves both evaluated backbones, while a bucketed index reduces online retrieval overhead; the central claim is that forecasting-model representations provide more predictively relevant retrieval keys than raw-series similarity or separate embeddings.

Critical strengths
  • The design is simple and easy to understand. The method reuses the forecasting model’s hidden states without updating model weights or requiring a separate embedding model in the self-retrieval configuration.
  • The authors report consistent forecasting improvements. Both Chronos-Bolt and Timer-S1 improve across all six datasets, with approximately 6–7% reductions in their respective mean MSEs.
  • Layer-wise results show increasing alignment with future compatibility in deeper layers, providing empirical support for the proposed retrieval key.
  • The paper does a good job on scaling to larger databases for retrieval. It evaluates approximate retrieval against exact search and reports end-to-end timing on a five-million-entry database.
Critical weaknesses
  • The reported 7% MSE improvement over TS-RAG compares different backbones, Timer-S1 for HiRAF and Chronos-Bolt for TS-RAG. With the same backbone, the gain is approximately 3.6%, so the headline partly reflects backbone differences.
  • I think there are important relevant baselines. The controlled comparison includes instance-normalized raw contexts and one embedding model, but omits strong alternatives such as endpoint-anchored L2 or retrieval using predicted futures.
  • No confidence intervals or significance tests are reported.
Questions for authors
  • Can you emphasize the same-backbone improvement over TS-RAG and compare runtime with both methods using the same approximate-search implementation?
  • I am curious about isolating the effect. Can you separately report individual-neighbor future error, averaged retrieval-forecast error, and complementarity with backbone errors to establish why hidden-state retrieval helps?
  • Do the improvements remain statistically reliable under temporal block-bootstrap evaluation, different index seeds, and matched baseline comparisons at longer horizons?
7 comments

We thank you for these valuable comments. We have run the additional experiments you asked for, and respond to each point below. We hope these resolve your concerns, and we look forward to your reply; if you have any further questions, we will answer them promptly.

W1 — The 7% also involves a backbone change

We accept this. Retrieval-augmented forecasting methods of this kind are a correction applied to a backbone forecast — which may differ from RAG in this respect — so a more appropriate measure is the improvement each method achieves over its own backbone. Taking the three rows of Table 1 that share the Chronos-Bolt backbone:

ETTh1ETTh2ETTm1ETTm2WeatherElectricityMean
Chronos-Bolt (backbone).3615.2516.3107.1486.2006.1126.2309
TS-RAG.3556.2451.2904.1465.1905.1114.2232
HiRAF (Chronos-Bolt).3450.2465.2739.1405.1743.1104.2151

On the same backbone, TS-RAG improves on the backbone by 3.3% and HiRAF by 6.8%, a factor of 2.05×. On the Timer-S1 backbone (Table 2) HiRAF improves on the backbone by 6.2% and TS-RAG's retrieval key, run through the same pipeline, by 2.8%, a factor of 2.25× — the ratio is close on the two backbones, which makes it a more stable measure than a direct comparison of the absolute values. The 3.6% you compute is the same three rows read as HiRAF against TS-RAG (.2232 → .2151); the two readings are consistent.

We will revise the abstract and introduction so that the same-backbone numbers are stated first and the cross-backbone 7% is reported as a secondary remark.

W2 — Stronger baselines: endpoint-anchored L2, and retrieval by predicted futures

We added both. In each case the only thing that changes is the vector handed to the search: same backbone, same retrieval database, same re-alignment and aggregation, same scalar fusion.

(a) Retrieval by the backbone's predicted futures. Six datasets, Timer-S1 self-retrieval, k = 50, three normalisations of the forecast tried and the strongest reported per dataset. The hidden-state key wins on all six, and the best forecast key retains only 13%–56% of HiRAF's gain over the backbone. Working well takes two things: the retrieved forecast has to be accurate on its own, and it has to offer a correction where the backbone is wrong. This key has only the first: it is the backbone's forecast itself, so retrieving with it groups together the windows the backbone predicts alike, and the mean true future of that group is what training already aligns the forecast to, leaving almost no systematic error to remove. On electricity it retrieves a more accurate forecast than ours (.1407 vs .1445) and keeps only 56% of the gain.

(b) Endpoint-anchored L2. Six datasets × two backbones = 12 cells, k = 10, exact search; the hidden-state key is Timer-S1's in both halves.

retrieval keymean fused MSEvs backbone
backbone (no retrieval).2259—
instance-normalised raw context (the control used in the paper).2191−3.03%
Chronos-T5 embedding (TS-RAG's key).2178−3.60%
endpoint-anchored L2.2157−4.51%
hidden state (HiRAF).2141−5.22%

Endpoint-anchored L2 is one common form of prediction-oriented retrieval: it matches on where the context ends rather than on the shape of the context as a whole. It is a strong baseline, which agrees with our hypothesis that a retrieval key has to carry what is relevant to the continuation. It also beats the separate embedding model TS-RAG retrieves with (−4.51% against −3.60%), although that embedding comes from a pretrained time-series model and this one is a hand-written distance. But the continuation is not always governed by proximity to the end of the context, and where it is not, the hidden-state key is the better one, since the backbone has already computed whatever is predictively relevant for that series. Overall the hidden-state key is ahead: −5.22% against −4.51%. We will add endpoint-anchored L2 to the key comparison in the revision.

W3 — No confidence intervals or significance tests

You are right, and we have replaced the procedure. The test in the submission resampled (channel, window) rows independently, but test windows slide with stride 1, so adjacent windows share L + H − 1 = 575 steps; against a moving-block bootstrap this understated the width of the confidence interval by a factor of 5–12, and a Newey–West HAC estimator (bandwidth 576) agrees with the block bootstrap. The revision reports moving-block bootstrap intervals (10,000 resamples), with a channel-clustered variant and Politis–Romano subsampling as block-length-independent cross-checks. Intervals are reported at the widest block length b satisfying b/n ≤ 0.1, since beyond that ratio each resample covers nearly the whole series and the interval narrows spuriously — in the direction that would favour us.

HiRAF against its own backbone, H = 64, all six datasets and both backbones:

datasetbackboneHiRAFbackboneΔMSE%95% CI of the differencep
ETTh1Timer-S1.3172.3242−2.17%[−1.12e−2, −2.38e−3]0.0036
ETTh1Chronos-Bolt.3441.3615−4.82%[−2.44e−2, −9.81e−3]<1e−4
ETTh2Timer-S1.2410.2483−2.94%[−1.18e−2, −3.66e−3]<1e−4
ETTh2Chronos-Bolt.2449.2516−2.65%[−1.23e−2, −2.61e−3]<1e−4
ETTm1Timer-S1.2646.2948−10.27%[−4.12e−2, −2.19e−2]<1e−4
ETTm1Chronos-Bolt.2794.3107−10.08%[−4.53e−2, −2.08e−2]<1e−4
ETTm2Timer-S1.1505.1609−6.44%[−1.42e−2, −7.22e−3]<1e−4
ETTm2Chronos-Bolt.1436.1486−3.35%[−6.62e−3, −3.69e−3]<1e−4
weatherTimer-S1.1749.1799−2.82%[−8.25e−3, −1.79e−3]0.0016
weatherChronos-Bolt.1860.2006−7.29%[−2.01e−2, −8.75e−3]<1e−4
electricityTimer-S1.1136.1170−2.95%[−4.23e−3, −2.76e−3]<1e−4
electricityChronos-Bolt.1098.1126−2.52%[−3.78e−3, −2.12e−3]<1e−4

All 12 cells are significant.

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

This paper addresses a limitation of TSFMs that the history they can use is restricted by the input context length, through retrieval. Existing works retrieve based on the raw time series or on a separate embedding model, whereas this work uses the TSFM’s own embedding as the retrieval key, so that retrieval is performed at the hidden-state level.

Critical strengths
  1. The main idea makes sense. With the raw time series, small but important features may be omitted, while using the hidden state as the key can mitigate this, since the hidden state is a vector that contains the information the TSFM has processed for forecasting. Using the TSFM itself instead of an additional model also seems efficient.

  2. The implementation details are also good. Time-series retrieval usually has the concern that the search space grows as the history becomes longer. The use of buckets of keys addresses this concern well. Although the notion of bucketing keys is not novel, this work seems to apply it in the right place.

  3. Most of the necessary experimental results are provided. The main results against TSFMs (including a TSFM with retrieval), and the ablation studies on the use of retrieval history and on fusion, support the main claim.

Critical weaknesses
  1. The paper does not discuss or verify why using the TSFM’s own hidden state is better than using the embedding of a separate model. The empirical results are there, but a more theoretical or intuitive hypothesis, together with an analysis that supports it, is necessary.

  2. The writing of some sections must be improved. Although the core idea is simple, straightforward and intuitive, I had to spend a significant amount of time understanding the sentences in some sections. This is not a matter of vocabulary or grammar, but rather of the structure of sentences and paragraphs. Specifically, Sections 2, 3.3 and 4.2 were hard to understand.

  3. Fusing the retrieval results, rather than “utilizing” them, is another concern. Naive fusion of the retrieved results with the model prediction can cause conflicts between the two and lead to worse performance.

  4. The experiments are limited to TSFMs. This is understandable since the paper’s main subject is TSFMs, but it would also be interesting to see whether HiRAF can catch up with models trained specifically for a given dataset. A comparison with RAFT, which trains on the dataset together with retrieval, would also be interesting.

Questions for authors
  1. What would be the main reason that the hidden-state key outperforms the T5 embedding? Is it simply because the backbone is a stronger model than Chronos-T5? If so, would the benefit decrease if the hidden state of another strong TSFM, rather than the backbone’s own, were used as the key?

  2. Is there a reason for fusing? Why not put the retrieved results into the TSFM’s input context? If this is possible, could the performance of that variant be compared with HiRAF?

  3. What would be a good direction for using the core idea of HiRAF in the pretraining of TSFMs?

  4. In the introduction, the paper asserts that time-series retrieval differs from retrieval in NLP, but kNN-LM, which is mentioned in the same paragraph, has the same retrieval structure as the time-series setting: the key comes from the context, but the value comes from the future tokens. In this case, can we still say that the value mismatch in retrieval is a problem unique to time-series forecasting, or should we rather say that the core insight of kNN-LM also applies to time-series forecasting?

8 comments

We thank you for these valuable comments. We have run the additional experiments you asked for, and respond to each point below. We hope these resolve your concerns, and we look forward to your reply; if you have any further questions, we will answer them promptly.

W1 — Why the backbone's own hidden state beats a separate embedding model (answered together with Q1)

We compare several retrieval keys side by side.

The raw series. It compares the similarity of the input series themselves, which is not a prediction-oriented retrieval; the paper already makes this point.

A separate model's embedding. It is an encoding of the input, so it is still a retrieval oriented towards the input, with some weight given to signals that are useful for the forecast. The fact that the correlation between the hidden state and the forecast increases layer by layer (Figure 5) corroborates this.

The model's predicted values used directly as the key. What this retrieves is a good forecast when used directly as the prediction, but it works poorly as a retrieval augmentation, because it is too close to the model's own forecast to supply new information.

Taken together, a good retrieval key has to meet two requirements:

  1. it is prediction-oriented, and has extracted the many kinds of temporal information that are useful for forecasting — how good the retrieved result is when used directly as the prediction reflects this;
  2. it has to differ from the backbone's forecast, which is what allows it to supply additional information, and in particular to correct the backbone where the backbone departs sharply from the truth.

The hidden state meets both:

  1. It is prediction-oriented: layer by layer, its association with the future increases (Figure 5).
  2. It has been pretrained at scale and carries a large and varied set of temporal features relevant to forecasting, of which the model's forecast uses only a part — the hidden state still holds useful information that was extracted but not used. The forecast is only one point taken from the predictive distribution the hidden state supports; other outcomes may be equally reasonable, yet the model can emit only a single value. (A simplified intuition: given a multimodal distribution, the forecast can pick only one of its modes as the answer.)

Our experiment on specialised models makes the second point from the other side: with a specialised model's predicted values as the key, retrieval is accurate when used directly, but fails as a retrieval augmentation — the model has been trained on this one dataset only, without large-scale pretraining, and extracts only part of the temporal structure (see W4(b)).

Finally, the question you raise in Q1: whether another model's hidden state would do. A specialised model's would not, for the reason just given; another TSFM's may — a weaker model augmented with the hidden state of a stronger one meets both requirements, and may even be the stronger forecaster. (Fused MSE, k = 10, exact search, mean over six datasets):

backboneits own hidden stateanother TSFM's hidden state
Timer-S1 (.2209).2103.2105 (Chronos-Bolt's key)
Chronos-Bolt (.2309).2187.2179 (Timer-S1's key)

We use the backbone's own hidden state because it is a by-product of the forward pass that produces the forecast, and so costs nothing extra.

W2

We will improve the structure of the paper.

W3 — Fusing the retrieved results

We do not claim this is the best way to fuse. We chose a deliberately simple one — a single scalar fitted on validation — so that the gain is attributable to the hidden-state key rather than to a fusion module.

We are grateful for the comment nonetheless, and have run an ablation over fusers with the key, the library and the aggregation held fixed (Chronos-Bolt backbone, four ETT datasets, k = 10; every fuser fitted on validation):

fuserETTh1ETTh2ETTm1ETTm2meanvs backbone
backbone, no retrieval.3615.2516.3107.1486.2681—
one scalar (HiRAF).3441.2449.2794.1436.2530−5.63%
two scalars.3417.2448.2775.1430.2518−6.09%
two scalars per channel.3430.2493.2776.1423.2531−5.60%
MLP gate on (m, r, m−r).3405.2476.2810.1426.2529−5.66%
the same gate, fitted on test (upper bound).3210.2267.2477.1376.2333−12.99%
optimal α per window (upper bound).2598.1838.1578.1013.1757−34.47%
optimal α per step (upper bound).2147.1753.1577.0964.1610−39.93%

The four usable fusers land within half a percentage point of one another, so the method is not sensitive to the choice of fuser. The italic rows are fitted on the test set and are only an upper bound. We have added this ablation to the appendix.

Reviewer_w8XjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HiRAF, a retrieval-augmented forecasting method that uses the hidden states of pretrained time series foundation models to retrieve similar historical patterns and improve forecasting accuracy. The proposed method does not require additional model training and uses an efficient retrieval mechanism to reduce computational costs. The authors evaluate the method on six benchmark datasets and report improvements over existing foundation models and retrieval-based forecasting approaches.

Critical strengths
  1. The proposed method uses the backbone’s own hidden states for retrieval, without requiring a separate embedding model or updating the backbone parameters. This makes the method simple and potentially useful in practice, since it can use additional historical information with limited extra computation during inference.
  2. The controlled comparison in Table 2 shows that the hidden-state key improves forecasting accuracy over raw-series and T5 embedding keys while keeping the other components fixed. This is an important strength because it provides direct evidence for the benefit of the proposed retrieval key, beyond comparing the performance of complete systems.
Critical weaknesses
  1. The novelty over existing retrieval methods is not sufficiently established. The overall design is closely related to kNN-LM, while TimeRAF already considers forecasting quality when learning to retrieve useful candidates. Using the frozen backbone’s own hidden states is a useful design choice, but the paper needs to explain more clearly what new contribution it provides beyond these existing ideas.
  2. The main comparisons do not fully isolate the benefit of the proposed method. The reported 7% improvement over TS-RAG also involves changing the backbone; with the same backbone, the improvement is approximately 3.6%. Similarly, the efficiency comparison uses approximate search for HiRAF and exact search for TS-RAG. Although the controlled experiment in Table 2 supports the retrieval key, these differences make the overall accuracy and efficiency advantages less convincing and weaken the evidence supporting acceptance.
Questions for authors
  1. Clarification of novelty: TimeRAF already uses forecasting-quality feedback to guide retrieval, and kNN-LM uses pretrained hidden states to retrieve targets and combine them with model predictions. Could you clarify the main methodological contribution beyond these works, particularly the claim that previous methods do not consider predictive usefulness when designing retrieval?
  2. Clarification of the evaluation setting:How is the fusion weight α fitted: separately for each dataset, channel, or prediction horizon? Since this fitting uses target-domain validation labels, how much improvement remains when α is fixed or selected on separate source datasets? This would help determine how much the results depend on target-specific calibration.
  3. Evidence needed for the main comparisons:Could you provide accuracy and efficiency comparisons with TS-RAG using the same backbone, retrieval database, and comparable approximate-search settings? 4- The method assumes that cosine similarity between hidden states reflects similarity between future values. However, it is not clear why cosine similarity is the best choice. Have the authors compared it with retrieval based on the backbone's predicted values?
7 comments

We thank you for these valuable comments. We have run the additional experiments you asked for, and respond to each point below. We hope these resolve your concerns, and we look forward to your reply; if you have any further questions, we will answer them promptly.

W1 — Novelty relative to kNN-LM and TimeRAF

You are right that our statement is inaccurate, and we will revise it to the following: most prior methods do not consider predictive usefulness, in particular those that retrieve in the history-prediction manner, while TimeRAF, which retrieves history only, does consider it when training its retriever.

TimeRAF. Beyond the question of training, two differences remain. First, what is retrieved differs: TimeRAF retrieves an additional context segment and splices it into the input embedding, whereas our retrieved value is itself a forecast and can be fused directly. Second, TimeRAF's retriever is trained jointly with its Channel Prompting fusion module, so the contribution of usefulness-aware retrieval cannot be separated from that of the learned fusion; their own ablation also reports that the learned retriever is only marginally better than plain cosine retrieval (0.359 vs 0.360 on ETTh1, 0.152 vs 0.153 on Weather). HiRAF obtains a prediction-relevant retrieval key without any training.

kNN-LM. kNN-LM operates at the token level, where orienting the key towards next-token prediction is immediate; in that form it acts as a training-free adaptor rather than as retrieval-augmented generation. Our setting is document-level: under the tokenisation of a time series foundation model, both the key (a 512-step context) and the value (a 64-step future) span many tokens, and the retrieved value is a complete forecast rather than a single next-step distribution. Our observation that context similarity does not imply future similarity is likewise made with respect to document-level retrieval.

W2(a) — The 7% also involves a backbone change

We accept this. Retrieval-augmented forecasting methods of this kind are a correction applied to a backbone forecast — which may differ from RAG in this respect — so a more appropriate measure is the improvement each method achieves over its own backbone. Taking the three rows of Table 1 that share the Chronos-Bolt backbone:

ETTh1ETTh2ETTm1ETTm2WeatherElectricityMean
Chronos-Bolt (backbone).3615.2516.3107.1486.2006.1126.2309
TS-RAG.3556.2451.2904.1465.1905.1114.2232
HiRAF (Chronos-Bolt).3450.2465.2739.1405.1743.1104.2151

On the same backbone, TS-RAG improves on the backbone by 3.3% and HiRAF by 6.8%, a factor of 2.05×. On the Timer-S1 backbone (Table 2) HiRAF improves on the backbone by 6.2% and TS-RAG's retrieval key, run through the same pipeline, by 2.8%, a factor of 2.25× — the ratio is close on the two backbones, which makes it a more stable measure than a direct comparison of the absolute values.

W2(b) — The efficiency comparison used approximate search on one side only

We have installed our own inverted-file index, unchanged, on TS-RAG's key space: same backbone, same library, same batch, same nlist/nprobe, with their own L2 metric retained. electricity, L=512, H=64, k=10, batch 256, ms per query, inference only, pooled library N = 5,725,677.

key extractionsearchtotalvs backbone
HiRAF, exact0.00010.58401.03942.28×
HiRAF, IVF0.00010.02190.47731.05×
TS-RAG, exact2.60750.65473.71758.17×
TS-RAG, same IVF2.60750.02043.08326.77×

The correction is small — 8.17× to 6.77× — and TS-RAG remains considerably slower than HiRAF. As shown in the paper, the cost lies in the second encoder rather than in the index.

Q1 — Clarification of novelty

See W1.

Q2 — How α is fitted, and how much survives without target-domain labels

α is a single scalar per (dataset, backbone), fitted by least squares on the validation split only — a small set; no training data is used — and it is fitted neither per channel nor per horizon.

We also ran the variant you ask for: leave-one-dataset-out, in which the α applied to a target dataset is selected entirely from the other five datasets and the target's own validation labels are never used. Fraction of the gain retained, where gain = MSE(α=0) − MSE(α):

datasetTimer-S1Chronos-Bolt
ETTh1102.0%105.4%
ETTh2107.2%158.8%
ETTm184.4%88.6%
ETTm2104.0%126.3%
weather126.1%90.5%
electricity98.6%98.2%

Across all 12 cells the transferred α retains at least 84.4% of the gain, and in 7 of 12 it is better than the α fitted on the target's own validation split. The results therefore do not depend on target-specific calibration.

Q3 — Same backbone, same retrieval database, comparable approximate search

Accuracy is reported in W2(a) and runtime in W2(b); both use the same backbone, the same retrieval database and the same index.

Q4(a) — Why cosine

We compared metrics directly: the representation is fixed and only the metric handed to the search changes — cosine, L2, inner product, and whitened cosine. Averaged over six datasets and two hidden-state representations (this ablation uses k=10 and exact search):

metricTimer-S1, MSETimer-S1, MAEChronos-Bolt, MSEChronos-Bolt, MAE
cosine0.21040.26640.21830.2688
L20.21080.26650.21850.2689
inner product0.21450.26790.22420.2722
whitened cosine0.21290.26750.22010.2701

Cosine is the best of the four on both backbones, in MSE and in MAE, but the method is not sensitive to this choice: all four metrics lie within 2.7% of each other. The comparison between retrieval keys in Table 2 therefore does not depend on the metric.

Q4(b) — Retrieval based on the backbone's predicted values

We ran this on all six datasets, in one pipeline where the only difference between columns is the vector handed to the search (Timer-S1 self-retrieval, k=50). Three normalisations of the forecast were tried, and we report the strongest of the three per dataset.

datasetbackboneHiRAFbest forecast keyfraction of HiRAF's gain retained
ETTh10.32420.31620.320744%
ETTh20.24830.23850.244538%
ETTm10.29480.25610.282931%
ETTm20.16090.14560.158913%
weather0.17990.17170.178320%
electricity0.11700.11450.115656%

The hidden-state key wins on all six datasets, and the best forecast key retains only 13%–56% of HiRAF's gain. The reason is that the hidden state carries many kinds of information useful for prediction, of which the final forecast uses only a part. Retrieving by the predicted values therefore introduces no new information — it carries exactly the information that the backbone's own forecast already provides — and it returns the windows that the model itself already considers similar.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.