A predictive approach to enhance time-series forecasting

Skye Gunasekaran, Assel Kembay, Hugo Ladret, Ruijie Zhu, Laurent Perrinet, Omid Kavehei, Jason K. Eshraghian

Abstract

Accurate time-series forecasting is crucial in various scientific and industrial domains, yet deep learning models often struggle to capture long-term dependencies and adapt to data distribution shifts over time. We introduce Future-Guided Learning, an approach that enhances time-series event forecasting through a dynamic feedback mechanism inspired by predictive coding. Our method involves two models: a detection model that analyzes future data to identify critical events and a forecasting model that predicts these events based on current data. When discrepancies occur between the forecasting and detection models, a more significant update is applied to the forecasting model, effectively minimizing surprise, allowing the forecasting model to dynamically adjust its parameters. We validate our approach on a variety of tasks, demonstrating a 44.8% increase in AUC-ROC for seizure prediction using EEG data, and a 23.4% reduction in MSE for forecasting in nonlinear dynamical systems (outlier excluded). By incorporating a predictive feedback mechanism, Future-Guided Learning advances how deep learning is applied to time-series forecasting.

Introduction

Time-series data is notoriously noisy, complex, and subject to abrupt distribution shifts, often referred to as non-stationarity. While deep learning models excel at identifying complex nonlinear relationships, they frequently fail to capture long-term dependencies or adapt to sudden changes in the environment. Traditional drift-detection methods, such as the Page-Hinkley test, typically handle these shifts by triggering complete model retraining, which abruptly discards valuable long-term knowledge. Conversely, modern self-supervised approaches often decouple pretraining from online correction, failing to incorporate continuous feedback from new observations. This research addresses this gap by proposing a method to dynamically adjust a forecasting model's parameters to handle long-term uncertainty without forgetting the past.

Methodology: Future-Guided Learning (FGL)

The authors introduce Future-Guided Learning (FGL), a framework inspired by the neuroscience theory of predictive coding. This theory suggests that the brain constantly refines its internal models by minimizing the discrepancy, or "surprise," between expected and actual sensory inputs.

FGL adapts the concept of Knowledge Distillation (KD) by placing the "teacher" model in the relative future of the "student" model. In this setup:

  • The Teacher: A model pretrained to perform a short-term task, such as next-step prediction.
  • The Student: A model tasked with long-term forecasting (e.g., $n+k$ steps ahead).
  • The Loss Function: During training, the student learns from a combined loss function consisting of standard cross-entropy against ground truth and Kullback-Leibler (KL) divergence against the teacher's softened probability distribution from the near future.

To apply this probability-based distillation to continuous signals, the authors discretize continuous values into $B$ equal-width bins, effectively transforming regression into a categorical classification task. This discretization allows the model to capture the teacher's uncertainty patterns, yielding a smoother and more informative loss surface than standard Mean Squared Error (MSE).

Experimental Setup

The FGL framework was evaluated across several diverse datasets to test its robustness and generalization capabilities:

  • CHB-MIT Scalp EEG Database: Used for patient-specific epileptic seizure prediction.
  • Kaggle American Epilepsy Society (AES) Seizure Prediction Challenge: Used to test generalization on dogs and humans using "universal" teachers trained on broad populations.
  • Mackey-Glass Time Series: A synthetic, chaotic delay differential equation dataset used to evaluate continuous regression forecasting.

Results and Analysis

FGL demonstrated significant performance improvements in both event prediction and regression tasks. In seizure prediction on the CHBMIT dataset, FGL achieved a massive 44.8% average increase in AUC-ROC compared to strong baselines like MViT and CNN-LSTM. On the AES dataset, using a universal teacher on unseen patients, FGL still improved performance by 8.9%.

Average AUC-ROC performance across CHBMIT and AES datasets.

In regression forecasting on the Mackey-Glass dataset, FGL showed extreme robustness in chaotic environments. While the baseline model suffered from extreme outliers when the task was made harder (50 bins), FGL cut the MSE by over 50% (from 28.34 to 11.68). This highlights how the predictive feedback mechanism prevents the student from making catastrophic errors on long-horizon predictions.

MSE Loss across forecasting horizons for the Mackey-Glass dataset (Bins=50).

Discussion and Conclusion

The results indicate that temporal offset is a powerful supervisory signal. By shifting the teacher's inputs into the relative future, the student model gains a "sneak peek" at near-future uncertainty, which is crucial for long-horizon forecasting. However, the study also notes a trade-off between tailored patient-specific guidance and the stability of universal teachers, particularly when target events are infrequent in the training data.

Future work will explore hybrid schemes that adaptively weight patient-specific and universal teachers, as well as extending FGL to multi-horizon and multi-modal forecasting. By integrating FGL with classical online drift-adaptation methods, the authors aim to further boost resilience to non-stationary environments without discarding long-term knowledge. FGL successfully translates biological predictive coding into a practical, high-performing deep learning architecture for complex time-series analysis.

Summary updated

Abstract

Accurate time-series forecasting is crucial in various scientific and industrial domains, yet deep learning models often struggle to capture long-term dependencies and adapt to data distribution shifts over time. We introduce Future-Guided Learning, an approach that enhances time-series event forecasting through a dynamic feedback mechanism inspired by predictive coding. Our method involves two models: a detection model that analyzes future data to identify critical events and a forecasting model that predicts these events based on current data. When discrepancies occur between the forecasting and detection models, a more significant update is applied to the forecasting model, effectively minimizing surprise, allowing the forecasting model to dynamically adjust its parameters. We validate our approach on a variety of tasks, demonstrating a 44.8% increase in AUC-ROC for seizure prediction using EEG data, and a 23.4% reduction in MSE for forecasting in nonlinear dynamical systems (outlier excluded). By incorporating a predictive feedback mechanism, Future-Guided Learning advances how deep learning is applied to time-series forecasting.
BibTeX
@inproceedings{gunasekaran2025predictive,
  author = {Skye Gunasekaran and Assel Kembay and Hugo Ladret and Ruijie Zhu and Laurent Perrinet and Omid Kavehei and Jason K. Eshraghian},
  title = {{A predictive approach to enhance time-series forecasting}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-63786-4},
}
Figure