Multi-task Meta Label Correction for Time Series Prediction

TL;DR

The paper addresses two core problems in financial time-series classification: partial feature information and uncertain label quality. It proposes Multi-task Meta Label Correction (MMLC), a framework combining a two-branch label-correction meta-model with multi-task inner-loop classifiers, optimized via bi-level gradient updates. It couples novel data visualizations SGAF, SRP for historical data and RRP for horizon labels to enable adaptive labeling across multiple prediction horizons. Empirical results on XOM, S&P500, and SZ50 show that MMLC achieves higher accuracy and F1 than existing label correction baselines and is model-agnostic across classifier backbones. These findings suggest practical improvements for stock movement prediction by delivering cleaner labels and horizon-aware, shared meta-knowledge.

Abstract

Time series classification faces two unavoidable problems. One is partial feature information and the other is poor label quality, which may affect model performance. To address the above issues, we create a label correction method to time series data with meta-learning under a multi-task framework. There are three main contributions. First, we train the label correction model with a two-branch neural network in the outer loop. While in the model-agnostic inner loop, we use pre-existing classification models in a multi-task way and jointly update the meta-knowledge so as to help us achieve adaptive labeling on complex time series. Second, we devise new data visualization methods for both image patterns of the historical data and data in the prediction horizon. Finally, we test our method with various financial datasets, including XOM, S\&P500, and SZ50. Results show that our method is more effective and accurate than some existing label correction techniques.

Figures (9)

• Figure 1: The triple barrier method with wrong label examples. (a). Manual labeling with triple barrier method, where the label of price trends is determined by the amount of change from beginning to the first stopping time that hits boundaries of the rectangular. (b). The stock goes down as a whole but the manual label is up. (c). The stock trend is uncertain in the whole picture but the manual label is down.
• Figure 2: Historical data is used for generating samples of training features denoted as "X" in the following algorithm. Prediction horizon is used for generating labels, whose image patterns are denoted by "Y" in the following.
• Figure 3: Tendency confusion problem. The upper left is $y=2x+N(0,0.5)$. $N(0,0.5)$ is a Gaussian noise. The upper middle is the corresponding GAF of $y=2x+N(0,0.5)$. The lower left is $y=2x+N(0,0.5)$ in reverse order. The lower middle is the corresponding GAF plot of reverse data. The right column with our proposed method SGAF, shows different image patterns of the ascending and descending trends respectively, while the middle column (original GAF) has tendency confusion with the same patterns for different trends.
• Figure 4: An example of the SRP method. (The upper left is $y=2x+N(0,0.5)$ and lower left is $y=-2x-N(0,0.5)$, where $N(0,0.5)$ is the Gaussian noise. The middle plots are using corresponding RP method on the left column. The right column are images with our SRP method, with a sign function added into RP.
• Figure 5: A particular case with RRP. The left figure shows that $x_{n+k+h}-x_{n+k} <0$ with a "fall" label. The middle figure shows that the fluctuation of $\|x_{n+k+h}-x_{n+k}\|$ is small that implies the sequence is stationary. The right figure is the opposite of the left figure with $x_{n+k+h}-x_{n+k} >0$ indicating a rise.