InjectTST: Adding Global Context to Independent Time-Series Channels

Summary

The article explains InjectTST, a Transformer method for forecasting multivariate time series that combines independent channel processing with selectively shared global context. Separate channel pathways use a shared encoder, positional information, and learnable channel identifiers, while a parallel mixing route extracts cross-channel context. The global representation is supplied as keys and values to cross-attention, with each channel’s representation as the query. Two mixing designs are described: Channel as Token and Patch as Token; the article says the patch-based option is more stable overall, while the channel-based design can suit particular datasets.

The method also uses masked pretraining followed by two fine-tuning stages. The author implements an interpretation in MQL5 and applies it to an environmental-state model. The reported trading test shows a small profit factor and a drawdown exceeding one third, with no clear rising balance trend. The article acknowledges that results fell short and suggests that not following the original three-stage training process may have contributed. Its implementation and trading results should be viewed as a limited experiment, not evidence of a reliable forecasting edge.

Key ideas

• InjectTST keeps channel-wise Transformer processing as its base and adds global context through cross-attention.
• Learnable channel identifiers help a shared model retain channel-specific information.
• Global mixing can represent channels as tokens or patches, with different stability characteristics reported.
• The proposed training process combines masked pretraining with staged forecasting fine-tuning.
• The article’s MQL5 trading experiment reports weak results and substantial drawdown, leaving performance unresolved.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.