[FEATURE] Concatenation Layer for Per-Ticker Data

[GongJr0]

Feature Details

Implement a concatenation module that merges numeric AR(n) features with categorical embeddings (ticker, sector, period, etc.) into a unified torch.Tensor input. This will act as the final pre-model feature aggregation stage, ensuring consistent input formatting for the downstream neural network.

The module should:

• Accept heterogeneous feature groups (lag features, embeddings, derived stats)
• Handle variable-length embeddings per feature while ensuring consistent output shape
• Be modular so additional features can be appended later without breaking existing pipelines

Affected Modules

As stated in the parent issue.

Implementation Checklist

• Define input spec for numeric and embedding features
• Implement concatenation logic in FeatureGen (likely torch.cat along feature dimension)
• Add shape validation to ensure outputs are consistent across tickers and time steps
• Support batch-wise concatenation for multiple tickers simultaneously
• Unit tests:
  • Verify numeric + embedding features combine to the expected final dimension
  • Check correct handling of missing embeddings or features
  • Test batched input scenarios

Limitations

As stated in the parent issue.