Mantis Time-Series Classification: Convolutional Patching and Token Building
Summary
The article explains how the Mantis time-series classification model turns financial data into a fixed sequence of tokens for transformer analysis. It describes convolutional feature extraction, division into patches, and pooling to summarize each patch. The broader framework also combines raw and differenced inputs with window-level means and standard deviations, then adds positional information and uses attention to capture local and long-range patterns.
The article outlines contrastive pretraining, a classification stage with temperature calibration, and channel adapters intended to reduce the cost of combining multiple indicators. Its practical focus is implementing model components, including temporal embeddings and a segmentation object. The text claims that calibration can make confidence scores meaningful and gives training-corpus and hardware details, but this section does not provide independent validation or trading results. The implementation discussion is also incomplete in the supplied text, so its pooling changes and full model performance cannot be assessed here.
Key ideas
- Convolution transforms input series into multiple feature channels before fixed-size patching.
- Patch aggregation creates compact tokens for transformer processing; the article’s implementation replaces mean pooling with convolution followed by max pooling.
- Parallel streams represent raw values, first differences, and window-level mean and standard deviation.
- Contrastive pretraining and temperature scaling are presented as ways to improve pattern representation and probability calibration.
- Adapters can reduce the cost of combining channels while retaining cross-channel information.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.