Skip to main content
Search

# Time Series Forecasting

15 related articles

Subscribe to updates
  1. Retrieval for forecasting stops matching raw numbers: filtered trend charts cut long-horizon error

    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.

  2. aclanthology.org

    A million dollars per fusion shot: a diffusion language model cuts neutron-peak timing error from 11.6 to 9.2 steps

    Each inertial confinement fusion shot costs on the order of $1M, which makes fast AI surrogates valuable. ICF-DLM splits the 512-step neutron-rate waveform into total yield, peak timing and a short local waveform around the peak, generates them jointly with a bidirectional diffusion language model, and refines it with physics-driven reinforcement learning. On the 50-shot held-out experimental test set of ICFBench its peak-timing error is 9.2 steps, against 11.6 steps for an autoregressive LLaMA-3-8B trained with the same SFT plus PPO recipe.

  3. Coastal buoys get a three-level wave-height quality flag: one temporal model transfers across stations at a mean binary F1 of 0.856

    Significant wave-height records from a Brazilian coastal buoy network can now receive an hourly three-level quality flag, GOOD, SUSPECT or BAD, instead of relying only on fixed physical-limit and rate-of-change thresholds. The model decides one-sidedly from the current value and the preceding 168 hours, reaching a binary F1 of 0.853 on a pooled controlled-failure test set across six buoys and a mean binary F1 of 0.856 when transferred to unseen stations. In hourly streaming simulation, abrupt, multiplicative and sentinel failures were detected at onset, while drift and stuck-sensor episodes required accumulated evidence.

  4. Hourly air temperature forecasts get the humidity-wind-pressure coupling early: six Chinese cities cut average error 6.2%

    Near-surface air temperature forecasts 24 hours ahead must capture the daily cycle, synoptic evolution and hourly fluctuations at once, while common decomposition transformers defer or hide cross-variable interactions. A cross-channel temporal correlation transformer exchanges information among meteorological channels at the embedding stage via 1×1 convolutions, then fuses temporal, trend and raw branches with gated attention and multi-scale trend attention. On six Chinese city datasets it lowers MSE by 6.2% and MAE by 7.5% on average against mainstream models, with MSE reduction above 9% in Chongqing's extreme heat.

  5. arXiv

    Frozen time-series models no longer ship as-is: a coding agent rewrites the deployment pipeline and tops six metrics across 198 tasks

    Time-series foundation models forecast well zero-shot, yet deploying them still means hand-building input and output pipelines. TimeRSI has a coding agent repeatedly rewrite preprocessing, features, context selection and post-processing around a frozen forecaster, using errors on historical forecast windows to guide each revision, then picks among the resulting pipelines at deployment by similar past conditions. Across 198 tasks on TIME and FEV-bench it beats all published submissions on six point and probabilistic metrics; against direct inference with the same frozen TimesFM 3.0, TIME MASE falls 1.52% and CRPS 1.61%.

  6. arXiv

    Calendar encoding is not a detail: swapping it shifts holiday error by up to 83% in one model

    Time series foundation models now accept covariates, yet how calendar features are encoded has been treated as an implementation detail. Across four models and nine datasets shaped by weekends and holidays, seven calendar encodings were compared in a zero-shot rolling evaluation: encoding choice changed accuracy in every model, with within-model MASE gaps up to 9.9% on non-holidays and 83% on holidays, and the best encoding differed by model. SHAP attribution locates the gains around holidays, where signals cannot be inferred from target history.

  7. arXiv

    Time-series foundation models can skip whole Transformer blocks: A100 latency drops 11% to 13% at moderate sparsity, with dataset-level variation

    Two pretrained time-series foundation models skip whole Transformer blocks at inference, leaving mean forecasting quality nearly unchanged across 1,088 held-out windows while cutting median A100 batch-1 latency by 11.2% and 13.4% at moderate sparsity.

  8. arXiv

    Linear DLinear forecasts PXP-chain observables stably across the whole initial-state family, while the nonlinear Transformer degrades progressively toward the Néel (scarred) limit

    The work casts quantum many-body dynamics as time-series forecasting on a few-qubit PXP chain realizable with Rydberg-atom arrays, training a nonlinear Transformer and a linear DLinear with the same composite loss to predict local magnetizations and connected two-point correlators from a history window; the Transformer predicts accurately in the more ergodic regime but deteriorates progressively as the initial state approaches the scarred limit, whereas DLinear stays accurate across the entire family of initial states, its main deviations being small high-frequency oscillations with little effect on overall prediction error.

  9. arXiv

    Computing loss on scaled targets cuts MASE in all 24 architecture-benchmark comparisons of time series foundation model pretraining, by 18.8% on GIFT-Eval on average

    The work proves that reversing an affine scaling (e.g., ReVIN) before computing the loss multiplies each series' gradient by a power of the scaling denominator, letting high-scale series dominate training (ScaleCon); computing a homogeneous residual loss directly on scaled targets (ScaleIn) instead leaves every mini-batch gradient and the full optimization trajectory invariant to arbitrary independent positive rescaling of training series, and in pretraining across four TSFM architectures it lowers MASE in all 24 architecture-benchmark comparisons, by 18.8% on GIFT-Eval and 21.9% on the M-competitions on average.

  10. arXiv

    A unified scaling law predicts held-out Toto 2.0 capacity curves at 1.09% and 1.50% MAPE, and activation interventions show frozen time series foundation models reuse historical rules across queries

    Using 18,768 experimental cells from 21 checkpoints on 23 dataset-frequency tasks across six domains, the work fits a five-parameter unified scaling law linking capacity, input length, and forecast horizon, and predicts the fully held-out Toto 2.0 series (about 4M-2.5B parameters) with MAPEs of 1.09% at input length 2048 and 1.50% at 4096 on horizon-averaged capacity curves; it also proposes a unified theory of time series learning in which Gaussian regression, matched-history comparisons, parameter exchanges, and activation interventions support the account that full-shot models accumulate rule information in weights while frozen time series foundation models extract and reuse historical rules through activations.