PPGF: Probability Pattern-Guided Time Series Forecasting

TL;DR

PPGF addresses time series forecasting in settings where data exhibit multiple internal patterns by reframing the problem as probabilistic pattern classification guided forecasting. It introduces a quantile-based Grouping Strategy to form balanced pattern intervals, a TCP-enabled Confidence-aware Pattern Classifier, and a Relative Prediction module that forecasts offsets within a target interval, enforcing consistency between classification and forecasting. The framework combines a Temporal Information Extractor (1-D convolution + Transformer + GRN) with a ConfidNet-calibrated classifier and a regression head, trained under a joint loss ${\mathcal L}=\sum_i(\lambda_1\mathcal{L}^{conf}+\lambda_2\mathcal{L}^{cls}+\lambda_3\mathcal{L}^{reg})}$ where $\lambda_1=\lambda_2=1$, $\lambda_3=5$. Experiments on eight real-world TSF datasets show substantial improvements over state-of-the-art baselines, validating the benefits of end-to-end pattern-guided forecasting and the importance of classification-consistent predictions. The work offers practical impact by improving forecast accuracy across domains with diverse temporal regimes and provides a path toward data-driven pattern grouping and more reliable uncertainty estimation.

Abstract

Time series forecasting (TSF) is an essential branch of machine learning with various applications. Most methods for TSF focus on constructing different networks to extract better information and improve performance. However, practical application data contain different internal mechanisms, resulting in a mixture of multiple patterns. That is, the model's ability to fit different patterns is different and generates different errors. In order to solve this problem, we propose an end-to-end framework, namely probability pattern-guided time series forecasting (PPGF). PPGF reformulates the TSF problem as a forecasting task guided by probabilistic pattern classification. Firstly, we propose the grouping strategy to approach forecasting problems as classification and alleviate the impact of data imbalance on classification. Secondly, we predict in the corresponding class interval to guarantee the consistency of classification and forecasting. In addition, True Class Probability (TCP) is introduced to pay more attention to the difficult samples to improve the classification accuracy. Detailedly, PPGF classifies the different patterns to determine which one the target value may belong to and estimates it accurately in the corresponding interval. To demonstrate the effectiveness of the proposed framework, we conduct extensive experiments on real-world datasets, and PPGF achieves significant performance improvements over several baseline methods. Furthermore, the effectiveness of TCP and the necessity of consistency between classification and forecasting are proved in the experiments. All data and codes are available online: https://github.com/syrGitHub/PPGF.

Figures (10)

• Figure 1: Relationship between model accuracy and different patterns. We divided the data into five groups representing five patterns. (a) LSTM has different abilities to fit data in various intervals. (b) Time series data are imbalanced and have a long-tailed distribution. (c) The performance of LSTM is disparate for various patterns. The prediction performance is worse for the patterns with few samples, while it is better for the patterns with more samples.
• Figure 2: The distribution of the solar wind in the training set under different partition strategies. (a) Uniform partition. We can observe a large variety of frequencies among different groups. (b) The proposed grouping strategy is in Equation \ref{['equ:group']}. The data belonging to each group are balanced.
• Figure 3: Pipeline of our proposed Probability Pattern-Guided Forecasting method. Firstly, Temporal Information Extractor captures general temporal features at the shallow layer of the network. Then, these features are fed into the Probability Pattern Classifier and relative prediction strategy, respectively, to learn discriminant features and more specific forecasting features. Finally, the probability Pattern-Guided Forecasting strategy is used to get the final prediction results.
• Figure 4: Probability pattern-guided time series forecasting (PPGF) strategy. We present the classifier's probability output, the TCP calibration process, and the relative prediction of the predictor.
• Figure 5: Autocorrelation graphs of all datasets.