Deep Non-Parametric Time Series Forecaster

Abstract: This paper presents non-parametric baseline models for time series forecasting. Unlike classical forecasting models, the proposed approach does not assume any parametric form for the predictive distribution and instead generates predictions by sampling from the empirical distribution according to a tunable strategy. By virtue of this, the model is always able to produce reasonable forecasts (i.e., predictions within the observed data range) without fail unlike classical models that suffer from numerical stability on some data distributions. Moreover, we develop a global version of the proposed method that automatically learns the sampling strategy by exploiting the information across multiple related time series. The empirical evaluation shows that the proposed methods have reasonable and consistent performance across all datasets, proving them to be strong baselines to be considered in one's forecasting toolbox.

• The paper introduces a non-parametric forecasting method that generates predictive distributions by strategically sampling past data without assuming a fixed distribution.
• It extends the local NPTS model globally by employing a feed-forward neural network to learn effective sampling strategies across multiple time series.
• Empirical results demonstrate that DeepNPTS achieves competitive accuracy, high calibration, and fast computation on diverse datasets.

Introduction

Time series forecasting is an essential tool in various fields such as finance, meteorology, and supply chain management. Classical forecasting approaches often rely on parametric models that assume a specific statistical distribution of data. A new non-parametric method is introduced which, unlike its predecessors, does not assume a parametric form for the predictive distribution and instead generates predictive distributions by strategically sampling from empirical data distributions.

Non-Parametric Time Series Forecasting

The paper introduces the Non-Parametric Time Series Forecaster (NPTS), a method that predicts future values by sampling past time indices from a given context window of the data. Given the absence of assumptions about the data distribution, NPTS can adapt to any dataset and thus provide reliable baseline forecasts. The model showcases robustness, particularly in the presence of numerical instability across various data distributions. Seasonality and trends in the data can be managed by adapting the sampling strategy using existing techniques.

Advanced Global Forecasting Approach

Building on the local NPTS model, the research extends the approach globally via DeepNPTS, which can learn sampling strategies from a broader array of related time series. A feed-forward neural network is utilized to take into account past information and covariates, determining the sampling probabilities. The global approach offers advantages in its ability to learn from an entire dataset collectively. Therefore, DeepNPTS stands out for its robustness as it circumvents numerical instability issues thanks to a straightforward feed-forward pass in the neural network after model training.

Empirical Evaluation and Conclusion

The article's empirical evaluation section showcases how the proposed forecasting methods perform comparably well with state-of-the-art models. Presented results indicate them as strong candidates for robust, fall-back forecasting methods across a diverse range of datasets. Notably, NPTS and DeepNPTS are competitive in terms of accuracy and offer the added benefits of high calibration and fast computation. The paper concludes by affirming the efficacy of the proposed non-parametric models in providing accurate and robust probabilistic forecasts without the need for tuning to specific data distributions.