acceptodds
Under review as a conference paper at ICLR 2027

Beyond Numerical Similarity: Resolution-Variant Visual Retriever for Time Series Forecasting

Abstract

Retrieval-augmented time series forecasting relies on identifying historical segments that provide relevant predictive context. However, a single numerical similarity metric is constrained in capturing global and local temporal patterns. Viewing time series as visual trajectories makes trends, fluctuations, and turning points explicit, but exploiting this information requires both a comprehensive representation and a quantifiable similarity measurement for retrieval. In this work, we propose , a visual retrieval method that represents temporal patterns through triangle regions and measures their geometric agreement. Specifically, Visual Sampling selects visual anchors, and the triangle regions induced by these anchors form a region map for each time series. For retrieval similarity, we introduce Approximate Triangle-IoU, which measures region overlap on a shared time-value grid, providing an interpretable visual similarity without a pretrained vision encoder. Historical time series selected by the visual similarity are aggregated and fused with a lightweight forecaster. To control the overhead of visual retrieval, configurable sampling density and resolution balance representation content against cost, while packed binary masks enable efficient overlap computation. Experiments on seven real-world benchmarks demonstrate that ReViR achieves state-of-the-art forecasting performance, while substantially reducing retrieval-library storage in the evaluated settings.

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

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

The paper proposes ReViR, a visual retriever for time series forecasting. It first retrieves historical windows, then reranks them by the overlap of triangle regions constructed from representative points. The retrieved continuations are aggregated and combined with a lightweight forecaster. Experiments cover seven real-world forecasting benchmarks.

Critical strengths
  1. Comparing time series through geometric overlap is an interesting approach to historical retrieval. The similarity score can be inspected directly through the triangle regions, without relying on a pretrained vision encoder.
  2. Sampling density and raster resolution are separately configurable, giving users control over the representation size and computational cost.
Critical weaknesses
  1. The evaluation needs improvement. First, it is unclear whether the forecasting gains come from retrieval. The paper presents only one retrieval example and does not quantitatively evaluate retrieval quality. Second, the non-retrieval baselines miss several of the latest and relevant methods, such as MixLinear, TimePro, etc. Third, the numbers used in the results are different from the numbers in the original papers for many methods. The authors need to show more details if they reran the experiments. To summarize, it is difficult to support the claim of SOTA with these gaps.

  2. The paper evaluates the visual retriever by removing the visual module in the ablation study. This shows that the module is useful, but does not prove whether the gains come from the complex design of triangle-region overlap. The authors may compare it with simpler alternatives, such as using the same representative points with a distance-based similarity measure, to determine whether the gains come from representative point selection or the triangle shape designed.

Questions for authors

See the questions in the Weaknesses.

More:

  1. Many concepts remain unexplained, such as how the Raw Future MSE is calculated. It only appears in Fig. 1.
  2. Why does performance decrease despite an increase in the number of retrieved candidates?
  3. The hundredfold storage saving needs a clearer explanation, ensuring it's due to the new design and not other implementation skills.
  4. Reporting more experiment results with different seeds would be better, especially when the experiments are not heavy.
5 comments
AuthorsOpenReview ↗

Response to Reviewer UkEM

Dear Reviewer UkEM,

We greatly appreciate your time and the insightful comments on our manuscript. Your suggestions are very helpful for improving this work, and we have carefully revised the paper to address all the issues you raised. Our point-by-point responses are provided below:


W1: It is unclear whether the forecasting gains come from retrieval. The paper presents only one retrieval example and does not quantitatively evaluate retrieval quality.

We have expanded the case studies in the Appendix D to cover datasets. Each case contains two panels. Panel (a) shows the histories and continuations of the top-five retrieved candidates, allowing readers to examine how similar historical patterns relate to subsequent trajectories. Panel (b) overlays the query’s Visual Sampling triangles with those of the top-ranked candidate after reranking, and compares the forecasting continuation with the query’s ground-truth future. These examples illustrate the geometric correspondence captured by Triangle-IoU and the information that retrieved continuations provide for forecasting.


W1: The numbers used in the results are different from the numbers in the original papers for many methods. The authors need to show more details if they reran the experiments. To summarize, it is difficult to support the claim of SOTA with these gaps.

The main results table compares forecasting accuracy under a unified experimental protocol. We reran all baselines because their original evaluations used different settings, including input lengths and hardware environments, and some papers did not report all the metrics needed for our comparison. To ensure a fair comparison across all datasets and metrics, we evaluated every method using 720 historical time steps, forecast horizons of 96, 192, 336, and 720 steps, and identical chronological training, validation, and test splits. All experiments were conducted on an NVIDIA GeForce RTX 4090 GPU with 24 GB of memory, using PyTorch 2.4.0 and CUDA 12.1.

We will release the exact code, commands, and configurations for every baseline run, organized by model, dataset, and forecast horizon. These materials will specify the learning rate, batch size, number of training epochs, random seed, and model-specific settings.

W1: The non-retrieval baselines miss several of the latest and relevant methods, such as MixLinear, TimePro, etc.

We have added TimePro and MixLinear, and the table below reports their results alongside ReViR’s. Since the original TimePro paper uses a lookback length of 96, we reran its official implementation with a lookback length of 720 to match ReViR's evaluation setting, keeping all other configurations unchanged. The MixLinear paper reports MSE only, so we reran its official scripts as provided to obtain both MSE and MAE.

As shown in the table, ReViR outperformed TimePro. MixLinear performed slightly better than ReViR on ETTh1 and ETTh2, while ReViR performed better on the remaining datasets. As stated in the MixLinear paper, the authors performed hyperparameter searches for each dataset and setting. Therefore, its superior prediction performance on the ETTh1 dataset can be regarded as an implementation skill.

ModelHorizonETTh1 MSEETTh1 MAEETTh2 MSEETTh2 MAEETTm1 MSEETTm1 MAEETTm2 MSEETTm2 MAEExchange MSEExchange MAEWeather MSEWeather MAEElectricity MSEElectricity MAE
ReViR960.3620.3850.2700.3300.2900.3380.1600.2460.0830.2000.1620.2050.1300.226
ReViR1920.4030.4110.3360.3760.3240.3570.2160.2860.1690.2910.2020.2450.1420.238
ReViR3360.4340.4290.3710.4060.3530.3760.2670.3210.3070.4000.2470.2820.1510.249
ReViR7200.4420.4570.4110.4460.4040.4060.3480.3750.7520.6480.3150.3330.1670.264
ReViRAverage0.4100.4200.3470.3900.3430.3690.2480.3070.3280.3850.2320.2660.1480.244
MixLinear960.3510.3830.2830.3400.3330.3750.1700.2610.0850.2040.1720.2280.1390.233
MixLinear1920.3950.4230.3370.3760.3530.3820.2230.2980.1740.2950.2140.2620.1540.248
MixLinear3360.4110.4340.3560.3950.3910.4070.2750.3320.3190.4070.2580.2950.1710.265
MixLinear7200.4230.4560.3800.4230.4390.4350.3600.3841.2420.8370.3250.3450.2090.297
MixLinearAverage0.3950.4240.3390.3830.3790.4000.2570.3190.4550.4360.2420.2830.1680.261
TimePro (Rerun)960.4230.4360.3280.3760.3110.3610.1860.2760.1000.2290.1910.2520.1870.295
TimePro (Rerun)1920.4620.4620.4010.4180.3560.3890.2580.3140.1960.3230.2420.2930.2220.327
TimePro (Rerun)3360.5270.5030.4340.4460.3810.4030.3110.3520.3500.4390.2740.3130.2150.319
TimePro (Rerun)7200.5050.5040.5060.5000.4430.4360.3670.3951.7180.9650.3350.3560.2830.373
TimePro (Rerun)Average0.4790.4760.4170.4350.3730.3970.2810.3340.5910.4890.2610.3030.2270.329
TimePro (Original)960.3750.3980.2930.3450.3260.3640.1780.2600.0850.2040.1660.2070.1390.234
TimePro (Original)1920.4270.4290.3670.3940.3670.3830.2420.3030.1780.2990.2160.2540.1560.249
TimePro (Original)3360.4720.4500.4190.4310.4020.4090.3030.3420.3280.4140.2730.2960.1720.267
TimePro (Original)7200.4760.4740.4270.4450.4690.4460.4000.3990.8170.6790.3510.3460.2090.299
TimePro (Original)Average0.4380.4380.3770.4030.3910.4000.2810.3260.3520.3990.2510.2760.1690.262

W2: The authors may compare it with simpler alternatives, such as using the same representative points with a distance-based similarity measure, to determine whether the gains come from representative point selection or the triangle shape designed.

We compared Triangle-IoU with two alternatives that use the same sampled anchors. We kept the candidate pool and forecasting setup fixed and varied only the reranking measure.

Piecewise-linear L2L_2. We connect the anchors of each series at their original time coordinates to obtain two piecewise-linear curves, f(t)f(t) and g(t)g(t). We measure their root-mean-square separation over the history window of length T=L−1T=L-1 in time coordinates:

dL2=1T∫0T[f(t)−g(t)]2 dt,T=L−1.d_{\mathrm{L2}} =\sqrt{\frac{1}{T}\int_0^T [f(t)-g(t)]^2\,dt}, \qquad T=L-1.

To evaluate the integral exactly, we merge and sort the anchor timestamps from both curves as u0,…,uRu_0,\ldots,u_R. With δr=f(ur)−g(ur)\delta_r=f(u_r)-g(u_r), we obtain dL2=1T∑r=0R−1ur+1−ur3(δr2+δrδr+1+δr+12).d_{\mathrm{L2}} =\sqrt{ \frac{1}{T}\sum_{r=0}^{R-1} \frac{u_{r+1}-u_r}{3} \left(\delta_r^2+\delta_r\delta_{r+1}+\delta_{r+1}^2\right) }.

The factor ur+1−uru_{r+1}-u_r weights each interval by its actual duration.

DTW. We extract the anchor values in temporal order as a1,…,ana_1,\ldots,a_n and b1,…,bmb_1,\ldots,b_m, counting the window start once if it is also selected by the first sampling bucket. We use the absolute difference between anchor values as the local matching cost:

Di,j=∣ai−bj∣+min⁡(Di−1,j, Di,j−1, Di−1,j−1).D_{i,j} = |a_i-b_j|+ \min\left(D_{i-1,j},\,D_{i,j-1},\,D_{i-1,j-1}\right).

We set D0,0=0D_{0,0}=0 and initialize the remaining entries in the zeroth row and column to +∞+\infty. The final distance is: dDTW=Dn,m(n+m)/2.d_{\mathrm{DTW}}=\frac{D_{n,m}}{(n+m)/2}.

The table below compares forecasting accuracy across the three measures. Because all three use the same anchors, the comparison tests whether filled triangle-region overlap adds value beyond what can be achieved through pointwise comparison of the interpolated curves or elastic matching of the anchor values.

DatasetHorizonReViR MSEReViR MAEPiecewise-linear L2 MSEPiecewise-linear L2 MAEDTW MSEDTW MAE
ETTh1960.3620.3850.3650.3900.3630.390
ETTh13360.4340.4290.4410.4520.4380.451
Exchange960.0830.2000.0830.2010.0830.201
Exchange3360.3070.4000.3110.4020.3070.400

Q1: Many concepts remain unexplained, such as how the Raw Future MSE is calculated. It only appears in Fig. 1.

We have added definitions of both Raw Future MSE and Aligned Future MSE to the main text.

Let LL be the history length and HH the forecast horizon. Let qtq_t and xtx_t denote the observed values of the query and the top-ranked candidate returned by the retrieval method, respectively.

Raw Future MSE measures the error between the retrieved candidate’s original continuation and the query’s ground-truth future.

MSEraw=1H∑h=1H(xL+h−qL+h)2MSE_{raw} =\frac{1}{H}\sum_{h=1}^{H} \bigl(x_{L+h}-q_{L+h}\bigr)^2

Aligned Future MSE uses the same candidate and measures the smallest future MSE attainable by applying a constant vertical shift from the set D\mathcal D. The set D\mathcal D includes zero and the shift that aligns the candidate with the query at each historical time step. We select the shift retrospectively using the query’s ground-truth future, so Aligned Future MSE is an oracle diagnostic of how closely the retrieved continuation can match the query’s future after a vertical shift.

MSEaligned=min⁡δ∈D1H∑h=1H(xL+h+δ−qL+h)2MSE_{aligned} =\min_{\delta\in\mathcal D}\frac{1}{H}\sum_{h=1}^{H} \bigl(x_{L+h}+\delta-q_{L+h}\bigr)^2,

D={0}∪{qt−xt: t=1,…,L}\mathcal D =\{0\}\cup\{q_t-x_t:\ t=1,\ldots,L\}


Q2: Why does performance decrease despite an increase in the number of retrieved candidates?

Increasing KK incorporates lower-ranked candidates whose less relevant continuations may introduce noise into the weighted aggregation. Its optimal value depends on the recurrence of relevant patterns and retrieval-library coverage: broader aggregation benefits from abundant relevant matches but may introduce mismatched continuations when such matches are scarce. Among the evaluated settings, K=20K=20 achieves the best trade-off between forecasting accuracy and retrieval efficiency.

Reviewer_mwD3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ReViR is a two-stage retriever for retrieval-augmented forecasting:

  1. It recalls the top-M training windows using correlation, recent-weighted distance, and time-aware scores.
  2. It reranks them with an "Approximate Triangle-IoU": anchors from iterative largest-triangle sampling define triangle regions, which are rasterized into bit-packed masks and compared by IoU.

The top-K continuations are kernel-aggregated and fused with an affine base forecaster. The method reports sota results on 7 benchmarks and large retrieval-storage savings.

Critical strengths
  1. The idea is simple, training-free, and interpretable. The shape similarity needs no pretrained vision encoder, and bit-packed AND/OR popcount comparison is an elegant implementation.

  2. The main results are strong, especially on non-stationary data: Exchange averages 0.328 vs. 0.407 for SARAF, with H = 720 at 0.752 vs. 0.985. The best result is reached in 54/70 entries under a shared L = 720 protocol.

Critical weaknesses
  1. In Table 2, w/o RVVR vs. full gives 0.363 -> 0.362 on ETTh1-96, ties on Exchange-96, and is better on ETTm1-336. Table 8 shows that raster resolutions from 16^2 to 256^2 give identical MSE/MAE to 3 decimals. Taken together, this suggests the Triangle-IoU signal barely changes the selected top-K. The large Exchange gain more plausibly comes from endpoint alignment, the recall scores, or the forecaster, and none of these is isolated against SARAF/RAFT.

  2. ReViR stores 8.79 MiB on ETTm2, yet it must still hold the raw training series to recall with correlation and to aggregate continuations. A 99.7% reduction vs. RAFT therefore reflects storing the series once and indexing windows, rather than materializing 33k windows as RAFT does. It does not come from the bit-packed visual representation. ReViR's library build time is also more than 100× higher (129 s vs. 1.1 s on ETTm2). On ETTh2 (Table 7), ReViR is slower than RAFT in mean latency (161 vs. 113 ms) and throughput (197 vs. 275 q/s), but the main text reports only the favorable ETTm2/Weather settings.

  3. In Fig. 1, "Ours" has a raw future MSE of 244, vs. 28 (Euclidean) and 26 (DTW). It only wins after an oracle best-alignment that uses the future, which is not available at inference.

Questions for authors
  1. For each test query, what fraction of the final top-K changes when the visual score is removed (e.g., Jaccard overlap of top-K with vs. without RVVR)? Please report this for all 7 datasets.

  2. Can you run SARAF and RAFT with your recall stage plus endpoint alignment, but without Triangle-IoU, to isolate which component drives the Exchange improvement?

  3. How does Triangle-IoU compare against cheap shape baselines on the same candidate pool, such as shape-DTW, SAX/PAA-distance, or IoU of the filled area under the curve without ILTS anchors?

  4. Please report a storage comparison where RAFT also stores the series once and indexes windows. What is the residual saving attributable to the visual representation?

6 comments
AuthorsOpenReview ↗

Response to Reviewer mwD3

Dear Reviewer mwD3,

We greatly appreciate your time and the insightful comments on our manuscript. Your suggestions are very helpful for improving this work, and we have carefully revised the paper to address all the issues you raised. Our point-by-point responses are provided below:


W1&Q1: For each test query, what fraction of the final top-K changes when the visual score is removed (e.g., Jaccard overlap of top-K with vs. without RVVR)? Please report this for all 7 datasets.

For each forecast horizon hh and test query qq, we compared the final top-20 candidate sets Ah,qA_{h,q} and Bh,qB_{h,q} selected with and without the visual score, respectively. We computed the fraction of candidates replaced, 1−∣Ah,q∩Bh,q∣/201-|A_{h,q}\cap B_{h,q}|/20, and the Jaccard overlap, ∣Ah,q∩Bh,q∣/∣Ah,q∪Bh,q∣|A_{h,q}\cap B_{h,q}|/|A_{h,q}\cup B_{h,q}|. We also report the proportion of queries for which reranking replaces at least one candidate. The result demonstrates that numerical recall can indeed retrieve valid candidates, and visual reranking can reasonably adjust the ranking results.

DatasetMean Jaccard overlapMean candidate replacement rateQueries with a changed top-20 set
ETTh10.27957.46%100.00%
ETTh20.46137.52%100.00%
ETTm10.29056.93%100.00%
ETTm20.35549.20%100.00%
Exchange0.60925.25%99.53%
Weather0.48636.78%99.49%
Electricity0.70917.52%99.13%

W1&Q2: Can you run SARAF and RAFT with your recall stage plus endpoint alignment, but without Triangle-IoU, to isolate which component drives the Exchange improvement?

We designed endpoint alignment specifically for ReViR to compare the shapes of query and candidate histories relative to their values at the forecast origin. It is thus a deliberate methodological design choice rather than a mere implementation trick.

To assess the effectiveness of endpoint alignment independently of ReViR, we incorporated this operation into SARAF and RAFT while retaining each baseline's original candidate retrieval stage. As shown in the table below, endpoint alignment improved some forecasting metrics for the baselines, and ReViR still achieved better overall forecasting accuracy on Exchange. These results indicate that endpoint alignment provides complementary benefits to retrieval-based forecasting, while ReViR maintains an additional performance advantage over both baselines even when they adopt the same alignment strategy.

DatasetHorizonReViR MSEReViR MAESARAF MSESARAF MAESARAF + EA MSESARAF + EA MAERAFT MSERAFT MAERAFT + EA MSERAFT + EA MAE
Exchange960.0830.2000.0860.2030.0870.2040.0890.2060.0890.206
Exchange1920.1690.2910.1900.3070.1910.3070.1920.3090.1910.310
Exchange3360.3070.4000.3650.4360.3450.4220.3700.4420.3740.443
Exchange7200.7520.6480.9850.7420.9680.7341.1330.8071.1040.796
ExchangeAverage0.3280.3850.4070.4220.3980.4170.4460.4410.4390.439

W3&Q3: How does Triangle-IoU compare against cheap shape baselines on the same candidate pool, such as shape-DTW, SAX/PAA-distance, or IoU of the filled area under the curve without ILTS anchors?

We evaluated four variants of ReViR using the same top-M candidate pool for each query and the same forecasting setup, replacing only the Triangle-IoU score used for reranking with the following similarity measurements:

  1. Shape-DTW aligns sequences using distances between descriptors of local neighborhoods. (Zhao & Itti, 2016)

  2. SAX-distance converts each series into a sequence of symbols and computes distance using a lookup table for aligned symbol pairs. (Lin et al., 2003)

  3. PAA-distance divides each series into segments and computes the Euclidean distance between the vectors of segment means. (Keogh et al., 2001)

  4. Raw-series IoU computes IoU directly on the original series, providing a comparison with the combination of visual sampling and filled triangle regions in Triangle-IoU.

The tables below report forecasting accuracy and retrieval efficiency. ReViR with Triangle-IoU achieved the best forecasting results in most settings, while shapeDTW performed better in some.

MethodHorizonETTh1 MSEETTh1 MAEETTh1 TimeETTm1 MSEETTm1 MAEETTm1 TimeExchange MSEExchange MAEExchange TimeWeather MSEWeather MAEWeather Time
ReViR960.3620.3851.6250.2900.3384.6170.0830.2001.9880.1620.20515.314
ReViR3360.4340.4291.9240.3530.3765.0060.3070.4002.0440.2470.28215.293
Shape-DTW960.3630.388279.6680.2900.337204.9160.0830.201202.1450.1600.202218.558
Shape-DTW3360.4370.450202.3980.3530.375204.7160.3090.401203.6540.2470.283220.025
SAX-distance960.3640.3891.5070.2930.3394.1170.0830.2011.4340.1660.21013.786
SAX-distance3360.4390.4501.3580.3540.3764.4380.3020.3981.5850.2520.28813.575
PAA-distance960.3630.3891.3700.2920.3384.4400.0830.2011.4400.1650.21013.802
PAA-distance3360.4380.4491.2560.3540.3754.4270.3070.4011.4530.2520.28813.822
Raw-series IoU960.3630.3881.4230.2920.3394.4320.0830.2011.6160.1650.20914.221
Raw-series IoU3360.4360.4481.5170.3520.3764.4710.3080.4001.7270.2530.28814.231

The table also reports retrieval latency in milliseconds per query (milliseconds/query). For each dataset and forecast horizon, all methods were evaluated using the same batch of 32 queries to ensure a consistent timing protocol. The reported latency encompasses the entire retrieval pipeline, including candidate recall, reranking, and the fusion of retrieved continuations. ReViR with Triangle-IoU achieves substantially lower retrieval latency than the Shape-DTW variant, while incurring only a marginal overhead compared with the SAX-, PAA-, and Raw-series IoU variants. These findings demonstrate that Triangle-IoU provides a favorable trade-off between forecasting performance and retrieval efficiency among the evaluated approaches.


W2&Q4: Please report a storage comparison where RAFT also stores the series once and indexes windows. What is the residual saving attributable to the visual representation?

The indexed RAFT comparison disentangles storage savings from eliminating redundant windows and those achieved through bit-packing. We recalculated storage across six settings, assuming both indexed RAFT and ReViR store a single copy of the training series and a 64-bit start index per candidate window, from which histories and continuations are retrieved. To the best of our knowledge, this storage layout has not been documented in prior retrieval-augmented time-series forecasting work and was independently designed for ReViR. All storage costs in both tables are reported in megabytes (MB).

DatasetHorizonRAFTIndexed RAFTReViRBit-packed masksReViR excluding bit-packed masks
ETTh296682.010.292.031.670.36
ETTh2336855.530.281.981.620.36
ETTm2962941.151.188.797.211.58
ETTm23363779.131.178.747.161.58
Weather969431.913.2326.7823.123.66
Weather33612124.793.2226.6222.963.66

The first table shows that storing the training series once and indexing candidate windows substantially reduces RAFT’s storage footprint. Excluding the packed visual masks, ReViR incurs slightly higher storage overhead due to additional metadata for visual reranking, including raster dimensions and candidate value ranges.

To isolate bit-packing savings, we compared three representations of identical binary visual regions while fixing candidate windows, channels, and raster resolution (16 \times 1). Each format stores one raster per candidate window and channel, using 1, 8, and 24 bits per pixel for bit-packed, grayscale, and RGB representations, respectively. Across all six settings, bit-packing reduces storage to approximately 1/8 of grayscale and 1/24 of RGB, substantially mitigating the storage overhead of visual reranking.

DatasetHorizonBit-packed masksUncompressed grayscaleUncompressed RGB
ETTh2961.6713.3740.12
ETTh23361.6212.9638.89
ETTm2967.2157.67173.01
ETTm23367.1657.26171.78
Weather9623.12184.94554.82
Weather33622.96183.71551.13
Reviewer_By8FOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a retrieval method called ReViR for time series forecasting. This method first combines numerical similarity and temporal information to recall historical fragments, then constructs triangular regions through visual sampling, and uses the approximate IoU to measure the similarity of trajectory structures and the candidates for rearrangement. Ultimately, the model aggregates the subsequent observations of the retrieved fragments and fuses them with the basic prediction results. The paper evaluated its prediction performance and retrieval efficiency on seven benchmark datasets.

Critical strengths
  1. ReViR carefully design the retrieval similarity for the temporal patterns and the visual representation.
  2. The design of ReViR is clear and showing the effectiveness compared to the up-to-date baselines.
Critical weaknesses

The main weakness is that the experiments do not sufficiently isolate the contribution of the proposed triangle-region representation from the effects of sampling, alternative shape similarities, and changes in retrieval aggregation. Although the ablation results suggest that the visual branch can be useful, it remains unclear why Triangle-IoU identifies more predictive historical continuations and when it offers an advantage over simpler alternatives.

Questions for authors
  1. Additional datasets like traffic and solar. Could authors add the experiment on the traffic dataset to show the effectiveness of ReViR?
  2. Could the authors randomly permute the visual similarity scores among candidates? This would help determine whether the gains depend on meaningful visual matching rather than merely adding another score
  3. Abaltion study may have at least the four datasets to show the effectiveness of ReViR (the ablation study in the appendix should also have the same datasets). This is to show your method's components are generally effective across different datasets.
  4. Could the authors compare Triangle-IoU with simpler shape similarities under the same recall pool and forecasting setup? Such comparisons would help isolate the value of the filled triangle regions from that of sampling or generic shape matching.
  5. Some possible typos in the paper in table 1: Electricity in DLinear has the same results as Weather in iTransformer
9 comments
AuthorsOpenReview ↗

Response to Reviewer By8F

Dear Reviewer By8F,

We greatly appreciate your time and the insightful comments on our manuscript. Your suggestions are very helpful for improving this work, and we have carefully revised the paper to address all the issues you raised. Our point-by-point responses are provided below:


Weakness: "It remains unclear why Triangle-IoU identifies more predictive historical continuations and when it offers an advantage over simpler alternatives."

We aim to summarize local temporal patterns using fewer time steps and reduce the influence of intermediate noise. Uniform sampling selects points at fixed intervals and can miss short-lived peaks, dips, or changes in slope between sampled points. We therefore propose Visual Sampling based on maximizing triangle area. Given two fixed anchor points, the largest triangle criterion selects the intermediate point that maximizes the triangular area formed with the anchors. This preserves local variations in the temporal slope, enabling the retention of subtle trend changes even within monotonically increasing or decreasing segments.

Triangle-IoU computes the intersection-over-union of the query and candidate triangle regions in shared time–value coordinates, capturing how closely local changes match in both timing and magnitude. A small pointwise error can coexist with differences when salient changes occur relative to the forecast origin or how large those changes are. Such differences can be associated with different future trajectories. By comparing triangle regions, Triangle-IoU captures these local temporal features and helps identify historical segments with more similar patterns and potentially more predictive continuations.


Q1: Additional datasets like traffic and solar. Could authors add the experiment to the traffic dataset to show the effectiveness of ReViR?

We additionally evaluated ReViR on two public benchmarks: Solar-Energy (Lai et al., 2018), which contains simulated photovoltaic power output at 10-minute intervals for 137 sites in Alabama, and Traffic (Wu et al., 2021), which contains hourly road occupancy rates from 862 freeway sensors in the San Francisco Bay Area. Solar and Traffic provide a test of ReViR on datasets with large size of samples or channels, where constructing visual representations, retrieving historical candidates, and training forecasting models can be computationally demanding. As shown in the table, ReViR are broadly comparable to those of the strongest baselines in our comparison, showing that ReViR can remain competitive on these larger datasets under limited computational resources.

ModelHorizonSolar MSESolar MAETraffic MSETraffic MAE
ReViR960.1820.2120.3750.258
ReViR1920.2050.2240.3860.261
ReViR3360.2240.2350.3960.265
ReViR7200.2310.2380.4230.281
ReViRAvg0.2100.2270.3950.266

Q2: Could the authors randomly permute the visual similarity scores among candidates? This would help determine whether the gains depend on meaningful visual matching rather than merely adding another score

For each query, we first computed the Triangle-IoU scores of the top-M retrieved candidates and then randomly permuted these scores before assigning them back to the corresponding candidates for reranking. This procedure preserves the candidate set and the empirical distribution of IoU scores for each query, while deliberately disrupting the alignment between individual candidates and their actual visual similarity to the query. The results across the evaluated datasets are reported below. The Shuffled-IoU variant consistently underperforms ReViR with the original Triangle-IoU scores. This result supports the interpretation that meaningful visual matching contributes to ReViR’s forecasting gains.

DatasetHorizonReViR MSEReViR MAEw/o Retr MSEw/o Retr MAEw/o RVVR MSEw/o RVVR MAEw/ Shuffled-IoU MSEw/ Shuffled-IoU MAE
ETTh1960.3620.3850.3670.3910.3630.3890.3640.389
ETTh13360.4340.4290.4510.4550.4380.4480.4380.449
ETTm1960.2900.3380.2980.3400.2920.3390.2930.340
ETTm13360.3530.3760.3620.3780.3520.3760.3520.377
Exchange960.0830.2000.0860.2070.0830.2010.0830.201
Exchange3360.3070.4000.3190.4050.3090.4010.3090.401
Weather960.1620.2050.1700.2090.1640.2090.1630.209
Weather3360.2470.2820.2580.2880.2530.2880.2520.289

Q3: Abaltion study may have at least the four datasets to show the effectiveness of ReViR (the ablation study in the appendix should also have the same datasets). This is to show your method's components are generally effective across different datasets.

We have expanded the ablation study in the main passage and Appendix B to cover more datasets. For each dataset, we report the results of all ablation variants alongside the full ReViR model, allowing readers to see how each component contributes across datasets.


Q4: Could the authors compare Triangle-IoU with simpler shape similarities under the same recall pool and forecasting setup? Such comparisons would help isolate the value of the filled triangle regions from that of sampling or generic shape matching.

We evaluated four variants of ReViR using the same top-M candidate pool for each query and the same forecasting setup, replacing only the Triangle-IoU score used for reranking with the following similarity measurements:

  1. Shape-DTW aligns sequences using distances between descriptors of local neighborhoods. (Zhao & Itti, 2016)

  2. SAX-distance converts each series into a sequence of symbols and computes distance using a lookup table for aligned symbol pairs. (Lin et al., 2003)

  3. PAA-distance divides each series into segments and computes the Euclidean distance between the vectors of segment means. (Keogh et al., 2001)

  4. Raw-series IoU computes IoU directly on the original series, providing a comparison with the combination of visual sampling and filled triangle regions in Triangle-IoU.

The tables below report forecasting accuracy and retrieval efficiency. ReViR with Triangle-IoU achieved the best forecasting results in most settings, while shapeDTW performed better in some.

MethodHorizonETTh1 MSEETTh1 MAEETTh1 TimeETTm1 MSEETTm1 MAEETTm1 TimeExchange MSEExchange MAEExchange TimeWeather MSEWeather MAEWeather Time
ReViR960.3620.3851.6250.2900.3384.6170.0830.2001.9880.1620.20515.314
ReViR3360.4340.4291.9240.3530.3765.0060.3070.4002.0440.2470.28215.293
Shape-DTW960.3630.388279.6680.2900.337204.9160.0830.201202.1450.1600.202218.558
Shape-DTW3360.4370.450202.3980.3530.375204.7160.3090.401203.6540.2470.283220.025
SAX-distance960.3640.3891.5070.2930.3394.1170.0830.2011.4340.1660.21013.786
SAX-distance3360.4390.4501.3580.3540.3764.4380.3020.3981.5850.2520.28813.575
PAA-distance960.3630.3891.3700.2920.3384.4400.0830.2011.4400.1650.21013.802
PAA-distance3360.4380.4491.2560.3540.3754.4270.3070.4011.4530.2520.28813.822
Raw-series IoU960.3630.3881.4230.2920.3394.4320.0830.2011.6160.1650.20914.221
Raw-series IoU3360.4360.4481.5170.3520.3764.4710.3080.4001.7270.2530.28814.231

The table also reports retrieval latency in milliseconds per query (milliseconds/query). For each dataset and forecast horizon, all methods were evaluated using the same batch of 32 queries to ensure a consistent timing protocol. The reported latency encompasses the entire retrieval pipeline, including candidate recall, reranking, and the fusion of retrieved continuations. ReViR with Triangle-IoU achieves substantially lower retrieval latency than the Shape-DTW variant, while incurring only a marginal overhead compared with the SAX-, PAA-, and Raw-series IoU variants. These findings demonstrate that Triangle-IoU provides a favorable trade-off between forecasting performance and retrieval efficiency among the evaluated approaches.


Q5: Some possible typos in the paper in table 1: Electricity in DLinear has the same results as Weather in iTransformer

We confirmed that the DLinear results for Electricity were entered incorrectly and duplicated the iTransformer results for Weather. We have corrected the DLinear entries for Electricity in Table 1 of the revised manuscript. The corrected values are provided below.

DatasetHorizonDLinear MSEDLinear MAE
Electricity960.1370.237
Electricity1920.1510.251
Electricity3360.1660.268
Electricity7200.2010.300
ElectricityAverage0.1640.264

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.