TimeRAF: Retrieval-Augmented Foundation model for Zero-shot Time Series Forecasting

Source

Status And Credibility

TimeRAF was posted to arXiv on 2024-12-30. The TimeCraft README labels it as a TKDE 2025 paper, and the IEEE Computer Society page confirms a TKDE publication entry. No official code repository was verified from the paper or TimeCraft README snapshot.

Core Claim

TimeRAF adds retrieval augmentation to zero-shot time-series forecasting. It builds customized time-series knowledge bases, learns a retriever, and integrates retrieved series through Channel Prompting.

Key Contributions

  • Treats external time-series examples as a retrieval knowledge base for TSFMs.
  • Learns retrieval scores end-to-end rather than using only fixed similarity.
  • Introduces Channel Prompting to integrate retrieved data along the channel dimension.
  • Evaluates zero-shot forecasting on six long-sequence forecasting datasets.

Evidence And Results

The paper trains on subsets of LOTSA and UTSD and evaluates ETTh1, ETTh2, ETTm1, ETTm2, Weather, and Electricity. It reports substantial zero-shot improvements over foundation-model baselines and ablations showing that learned retrieval and Channel Prompting matter.

Limitations

  • TimeRAF is retrieval-augmented forecasting, not time-series generation in the synthetic-data sense.
  • Retrieval can leak benchmark-like patterns if the knowledge base overlaps evaluation distributions.
  • It is passive forecasting; retrieved examples are not actions, interventions, or counterfactual trajectories.
  • Serving cost depends on knowledge-base construction, retrieval, and channel integration.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Context interfacepartially closesRetrieved time-series examples become external context for zero-shot forecasting.Needs provenance, leakage controls, and richer context types beyond example series.
Benchmarks and evaluation protocolwarningRetrieval can blur pretraining, adaptation, and evaluation boundaries.Needs explicit knowledge-base overlap audits.
Control and counterfactualsinsufficient evidenceNo action or intervention channel.Needs action-conditioned retrieval and rollout evaluation.