Work

Recolorized financial time series imaging: a multi-dimensional analysis of MTF, GAF, and recurrence plots for imbalanced stock trend classification

Year: 2026
Type: article
Abstract: Stock price prediction is a difficult task. The work proposed in this article takes on this issue and proposes a recolorized image encoding approach to classify the directional trend in a stock’s price. To this end, raw prices are complemented via technical and statistical features and projected by UMAP into a compact five dimensional space. This is then converted into three different image sets: 1) the Gramian Angular Field; 2) Markov Transition Field; 3) Recurrence Plot. We experiment with these methods to compare which method produces the best performance while preserving the temporal structure of the markets. For each image, the values are then recolorized, and a pretrained vision transformer is used to extract spatial features. Subsequently to classify the images, we use four layer bidirectional LSTM to model how these features evolve across time. In this paper, and for our experiments, we have also used SMOTE to correct the dominance of one particular class in the dataset. Finally, we experiment with seven downstream classifiers. Experimentation is performed on twenty companies of the National stock exchange spanning seven sectors. We have found that recolorized RP generalizes best at sector level. To be specific, sector wise number for accuracy is 0.7708 for RP, MTF at 0.7702 and GAF at 0.7535. Furthermore, without sector consideration, the numbers stand at: 0.7626 for RP, 0.7645 for MTF, 0.7410 for GAF. Lastly using statistical testing, we have found that the overall performance for RP and MTF is near identical (0.7626 vs. 0.7645; Cohen’s d = −0.0184), at the same time, both outperform GAF in rank based comparison (0.7410; Cohen’s d = −0.8028). These differences are confirmed by the Friedman test (p = 0.0211). Lastly, through the work presented in this article, we also try to show that the choice of recolorization, rather than the choice of classifier, could serve as a new benchmark for visual-temporal stock trend classification.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7226007323
FWCI:
Cites: 16
Cited by:
Related to: 10