POLA: Online Time Series Prediction by Adaptive Learning Rates

Wenyu Zhang

Abstract

Online prediction for streaming time series data has practical use for many real-world applications where downstream decisions depend on accurate forecasts for the future. Deployment in dynamic environments requires models to adapt quickly to changing data distributions without overfitting. We propose POLA (Predicting Online by Learning rate Adaptation) to automatically regulate the learning rate of recurrent neural network models to adapt to changing time series patterns across time. POLA meta-learns the learning rate of the stochastic gradient descent (SGD) algorithm by assimilating the prequential or interleaved-test-then-train evaluation scheme for online prediction. We evaluate POLA on two real-world datasets across three commonly-used recurrent neural network models. POLA demonstrates overall comparable or better predictive performance over other online prediction methods.

Introduction

In dynamic environments like financial markets, traffic monitoring, or energy grids, time series data is non-stationary, exhibiting "concept drift" where underlying patterns change over time. Traditional models trained on static datasets often fail as they cannot adapt to these shifts. Online learning aims to update models continuously, but faces a critical "Goldilocks" challenge: selecting an optimal learning rate. A rate too low results in sluggish adaptation, while a rate too high leads to overfitting on noise. Most existing adaptive methods are either heuristic or computationally expensive, requiring second-order derivatives (Hessians) that are impractical for deep neural networks.

The POLA Framework

POLA (Predicting Online by Learning rate Adaptation) addresses these challenges through a first-order meta-learning approach. Instead of a fixed or heuristically scheduled learning rate, POLA treats each incoming data batch as a miniature "test-then-train" scenario to optimize the learning rate in real-time.

The POLA framework: Splitting the update batch into meta-train and meta-validation sets to optimize the learning rate.

As illustrated in Figure 1, when a new batch $I_t$ arrives, it is split into a meta-training set and a meta-validation set. The model simulates an update on the training portion and evaluates the resulting performance on the validation portion. This allows the system to "learn how to learn" by selecting the learning rate factor ($\beta$) that minimizes immediate future error. By using a first-order Taylor expansion, POLA avoids the computational bottleneck of Hessian matrices.

The authors propose two implementations:

  • POLA-FS (Finite Search): Selects the best learning rate from a discrete set of candidates.
  • POLA-GD (Gradient Descent): Continuously optimizes the learning rate using its own gradient descent step, offering finer adaptation.

Experimental Evaluation

The framework was tested on two challenging real-world datasets known for concept drift:

  1. Sunspot Dataset: Monthly sunspot counts (1749–2020), characterized by complex, long-term cyclical patterns.
  2. Household Power Consumption: High-frequency, noisy multivariate data tracking electricity usage.

POLA was integrated with standard RNN, LSTM, and GRU architectures and compared against baselines like MAML, RMSprop, and "Follow The Leader" (FTL) strategies.

Results and Discussion

POLA consistently outperformed baseline methods across different architectures. On the Sunspot dataset, POLA-GD achieved a Normalized RMSE of 0.500, a significant improvement over the pre-trained static model (0.572) and standard Online-SGD (0.552). The method proved architecture-agnostic, improving performance regardless of whether an LSTM or GRU was employed.

Visualization of POLA-FS and POLA-GD adapting learning rates in response to time series patterns over time.

A key insight from Figure 3 is how the learning rate adapts to the data stream. POLA acts as a "gatekeeper": when data is noisy or inconsistent with the immediate future, it automatically reduces the learning rate to prevent erratic model updates. Conversely, when a clear new pattern emerges, it increases the rate to accelerate adaptation.

Conclusion

POLA provides a robust framework for online time series forecasting by automating learning rate regulation. By leveraging first-order approximations, it brings the benefits of meta-learning to deep RNNs without the prohibitive computational burden of second-order methods. While the meta-learning step introduces some overhead, the gains in predictive accuracy and resilience to concept drift make it a compelling choice for real-world streaming applications.

Summary updated

Abstract

Online prediction for streaming time series data has practical use for many real-world applications where downstream decisions depend on accurate forecasts for the future. Deployment in dynamic environments requires models to adapt quickly to changing data distributions without overfitting. We propose POLA (Predicting Online by Learning rate Adaptation) to automatically regulate the learning rate of recurrent neural network models to adapt to changing time series patterns across time. POLA meta-learns the learning rate of the stochastic gradient descent (SGD) algorithm by assimilating the prequential or interleaved-test-then-train evaluation scheme for online prediction. We evaluate POLA on two real-world datasets across three commonly-used recurrent neural network models. POLA demonstrates overall comparable or better predictive performance over other online prediction methods.
BibTeX
@inproceedings{zhang2021pola,
  author = {Wenyu Zhang},
  title = {{POLA: Online Time Series Prediction by Adaptive Learning Rates}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2021},
  url = {https://openreview.net/forum?id=31q2QJo-Nsp},
}
Figure