Retrieval for forecasting stops matching raw numbers: filtered trend charts cut long-horizon error
Synopsis
VRATF moves the comparison in retrieval-augmented forecasting from raw time series to denoised trend line charts: FFT low-pass filtering extracts the low-frequency trend, which is rendered as a 224×224 image and encoded by a frozen CLIP ViT-B/32 to build the retrieval database, with top-K neighbors selected by cosine similarity and aggregated by Gaussian-kernel weights. The retrieval forecast is fused at the prediction level with a DLinear base forecast through one learnable scalar. Across six multivariate benchmarks, VRATF reaches the lowest average MSE on four, with 0.340 on ETTh2 against 0.358 for the next-best RAFT, a 5.0% improvement.
Interpretation
Similarity for retrieval-augmented forecasting is no longer computed on raw values but on line-chart images of denoised trends, so matching rests on overall trend morphology rather than point-wise numerical correspondence. Earlier retrieval-augmented methods such as RAFT, DUET and SERAF retrieve over raw numerical sequences, learned temporal embeddings or generated text descriptions, where high-frequency fluctuations, sensor jitter and anomalous spikes weigh as much as informative signal. Across six multivariate benchmarks, VRATF attains the lowest average MSE on ETTh1, ETTh2, ETTm2 and Exchange; on ETTh2 it is lowest at every horizon among retrieval-augmented baselines, with an average MSE of 0.340 versus 0.358 for the second-best RAFT, a 5.0% improvement.
Each historical series is first shifted by subtracting its last observation to align endpoints, then passed through FFT low-pass filtering that keeps only the lowest frequency components and reconstructs a trend curve. The filtered trend is Min-Max normalized and rendered as a 224×224 RGB line chart, encoded by a frozen CLIP ViT-B/32 into a 512-dimensional L2-normalized vector, with a separate database per channel and training samples whose prediction windows overlap the query excluded. At query time the top-K entries by cosine similarity are weighted through a Gaussian kernel and their future trajectories summed into a retrieval forecast; the evaluation covers ETTh1, ETTh2, ETTm1, ETTm2, Weather and Exchange with look-back length 720 and horizons of 96, 192, 336 and 720.
The retrieval forecast and the DLinear base forecast are combined at the prediction level by a single learnable scalar rather than by fixed weights or gating in latent space. The scalar is mapped through a sigmoid into the interval from 0 to 1 and directly sets the relative contribution of the two prediction streams, after which a channel-wise linear projection calibrates the result and the removed offset is added back. On ETTh2 the learnable fusion beats a fixed coefficient of 0.5 at all four horizons, with the gap widening from 1.1% at 96 steps to 11.9% at 720 steps; removing the retrieval branch raises MSE at every horizon on ETTh2 and Weather.
Perspective
The results rest on six multivariate benchmarks, ETTh1, ETTh2, ETTm1, ETTm2, Weather and Exchange, with look-back length fixed at 720, horizons of 96, 192, 336 and 720, and a frozen CLIP ViT-B/32 as the vision encoder. The work suits readers interested in retrieval-augmented forecasting and multimodal representations. The authors propose future directions: scalable retrieval indexes for large time series databases, adaptive selection of frequency components across domains, and stronger vision encoders or time-series-specific visual pretraining objectives.
The reported advantage is competitive and often superior rather than uniformly leading: RAFT is best overall on ETTm1, DUET is best on Weather, and DLinear has a slightly lower average MSE than VRATF on Exchange (0.386 versus 0.392). The benefit of FFT filtering depends on periodicity, with the raw numerical variant marginally better at horizons 336 and 720 on ETTh1 and the two numerical variants close on Exchange. Each experiment is run once under a unified setting.