Back to the Future: Look-ahead Augmentation and Parallel Self-Refinement for Time Series Forecasting

Official implementation of Back to the Future (BTTF), a simple yet effective framework for long-term time series forecasting (LTSF) via look-ahead augmentation, parallel self-refinement, and ensembled Forecasting.

BTTF leverages segments of model-generated future predictions as future-aware context, enabling DMS-style parallelism while implicitly preserving IMS-like temporal dependencies.

Method Overview

BTTF is a two-stage framework that refines a base forecasting model through future-aware augmentation and parallel self-refinement, and further stabilizes predictions via a step-wise top-K ensemble.

1) Look-ahead Augmentation

Given a first-stage forecast, the predicted horizon is split into N segments, each of which is appended to the original input window to form augmented inputs.

2) Parallel Self-Refinement

N independent second-stage predictors are trained on the augmented inputs, each learning a distinct refinement pattern.

3) Ensembled Forecasting

Second-stage models are ranked by validation performance, and a step-wise top-K ensemble is constructed by selecting the optimal $K^*$ based on variance and covariance analysis.

Dataset

The /dataset directory contains the datasets used for model training and evaluation.

Quick Start

# Single test run
python run_parallel.py --gpus 0 --dataset etth1 --pred_len 96

# Full experiments (All Datasets & Horizon)
python run_parallel.py --gpus 0,1,2,3

=> For more detailed experiment configurations, please refer to ExpDetails.md.

Citation

If our work was helpful in your research, please kindly cite this work:

@inproceedings{kim2026back,
  title={Back to the Future: Look-ahead Augmentation and Parallel Self-Refinement for Time Series Forecasting},
  author={Kim, Sunho and Yoon, Susik},
  booktitle={Proceedings of the ACM Web Conference 2026},
  pages={8581--8584},
  year={2026}
}