Research map / Time series / Foundation models for time series

Apollo-Forecast: Overcoming Aliasing and Inference Speed Challenges in Language Models for Time Series Forecasting

AAAI 2025 · Time series › Foundation models for time series · approach: LLM-based time series forecasting · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Apollo-Forecast tokenizes time series with a Butterworth-filtered anti-aliasing quantizer and speeds up Chronos-T5 inference via Race Decoding, which races a small draft model against the main model with a tolerance check.

Problem

Language-model-based time series forecasters suffer from aliasing distortion introduced by fixed-length quantization/tokenization and slow autoregressive inference for long horizons with large models.

Approach

Pre-filter the signal with a Butterworth low-pass filter before quantization (AAQM) to suppress aliasing, and run a small draft model in parallel with the large Chronos model, returning the draft's tail frames when the two agree within a tolerance (Race Decoding).

Key results

• Agg. relative WQL reduced by up to 35.41% and MASE by up to 17.01% on UCR under zero-shot settings versus Chronos baselines.
• Race Decoding gives 1.9X-2.7X inference speedup over baseline methods on UCR and public datasets.
• On LBS crowd flow, Apollo-S reaches WQL 0.10 and MASE 0.24 versus Chronos-S 0.30 and 0.28 (scaled values).

Measured effects reported in the paper

Surprising findings

• Disabling Race Decoding (w/o-RD) slightly improves WQL on the UCR ablation, while the authors attribute most error reduction to AAQM.

Limitations

• Tolerance threshold gamma and its effect are not systematically analyzed.
• Evaluation relies on Chronos-family baselines with limited details on hardware and training.
• Reported improvements are inconsistent across tables (e.g., 31.44% vs 35.41% WQL, 17.01% vs 18.99% MASE).

Experimental setup

Datasets: UCR time series archive ETTh1 ETTh2 ETTm1 ETTm2 Weather ECL TrafficL LBS Shanghai crowd flow

Benchmarks: UCR dataset ETTh1 ETTh2 ETTm1 ETTm2 Weather ECL TrafficL LBS

Metrics: WQL MASE MAE MSE inference time

Baselines: Chronos-S Chronos-B Chronos-L Moirai N-HiTS N-BEATS DLinear TFT LSTM AutoARIMA AutoETS

Builds on: Chronos speculative decoding Butterworth filter

Closest papers by meaning

• AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.88)
• TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.88)
• Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.88)
• Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.87)
• A decoder-only foundation model for time-series forecasting (ICML 2024 · similarity 0.87)
• TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.87)