Patterns in time series to image phase and multi-step forecasting of financial indices

Jun Wu, Cheng Wang, Zelin Zhang, Lang He, Yaxiong Chen, Wenming ChengJ. King Saud Univ. Comput. Inf. Sci 2025 · DOI

Abstract

Predicting financial indices is a complex task due to the inherent volatility and non-stationarity of market data. This paper proposes a novel framework that transforms 1D financial time series into 2D images using a "relaxed recurrence plot" (RP) to capture multi-scale patterns. By integrating traditional econometric models with deep learning architectures—specifically CNNs and LSTMs—within hybrid encoder-decoder and boost schemes, the researchers demonstrate significantly improved multi-step forecasting accuracy for major indices like the S&P 500 and CSI300.

Introduction

Financial index prediction is plagued by high volatility and non-linearity. While traditional models like ARIMA or GARCH are limited by linear assumptions, modern deep learning approaches like LSTMs often struggle to identify localized temporal patterns hidden within raw 1D data. The authors identify a gap in current modeling: the failure to effectively combine data evolution (regression) with state representation (patterns). They propose treating time series as a "video" of evolving patterns, leveraging computer vision techniques to enhance predictive power.

Methodology: From Time Series to Imaging Networks

The core innovation is the Relaxed Recurrence Plot (RP). Unlike standard RPs that binarize data and lose granularity, the relaxed version preserves high-frequency components essential for prediction.

Imaging Pattern Transition Net

The sequence of generated images is treated as frames in a video. The authors utilize Optical Flow to track motion between frames and Markov Transition Fields to analyze pattern transitions over time. As shown in the transition network analysis, these patterns follow a power-law distribution, suggesting a structured "natural" memory in financial data.

The process of generating the Imaging Transfer Net, involving block decomposition, optic flow analysis, and pattern tracing.

Hybrid Architectures

Two primary schemes are proposed to bridge the gap between trend extraction and pattern recognition:

  1. Boost Scheme: Integrates non-parametric regression (NR) with LSTMs to manage both overall trends and residual fluctuations.
  2. Encoder-Decoder Scheme: Uses an "Encoder" (NR or LSTM) to extract the trend, while a "Decoder" (CNNs or GoogleNet) processes 2D image representations to capture spatial features.
The proposed hybrid architectures: (a) the Boost scheme combining LSTM with non-parametric regression, and (b) the Encoder-Decoder scheme utilizing 2D convolutional layers for pattern extraction.

Results and Key Insights

The models were evaluated using high-frequency data from the S&P 500 (1-minute intervals) and the CSI300 (240-minute intervals).

  • Superior Accuracy: The hybrid NR + LCNN (Non-parametric Regression + LSTM-CNN) model achieved an MSE of 0.0038 for 15-minute S&P 500 forecasts, consistently outperforming standalone models.
  • 2D Representation Advantages: Transforming data into 2D space allows for the simultaneous analysis of short-term autoregression and long-term relations, "unfolding" features invisible in linear formats.
  • CNNs and Volatility: Pattern-based CNN models showed faster convergence for volatility prediction, suggesting that volatility is better characterized by "shape" and "texture" than by previous numerical values alone.

Conclusion

While the hybrid models introduce increased complexity and potential information loss during the recurrence plot transformation, they offer a powerful new lens for financial forecasting. Future work aims to refine image reconstruction methods and explore the fundamental differences between natural and man-made time series to further optimize pattern recognition engines.

Summary updated

BibTeX
@article{wu2025patterns,
  author = {Jun Wu and Cheng Wang and Zelin Zhang and Lang He and Yaxiong Chen and Wenming Cheng},
  title = {{Patterns in time series to image phase and multi-step forecasting of financial indices}},
  journal = {J. King Saud Univ. Comput. Inf. Sci.},
  year = {2025},
  doi = {10.1007/s44443-025-00077-4},
}
Figure