Combining Persisted StockRanker Predictions into a Meta-Model

Summary

This discussion describes a model-stacking problem in a Chinese quantitative research platform. A user has trained and persisted twenty StockRanker models using the same factor set but differing in date ranges, data filters, or prediction horizons. The goal is to obtain each model’s score for the same observations, combine those scores into a single feature record, and use the resulting predictions as inputs to another model. Attempts to join many prediction blocks in a visual workflow reportedly caused data-column conflicts.

The post points readers toward using a custom Python module to load persisted models and generate predictions, and references related material on persisting deep-learning and StockRanker models. However, the supplied text does not include the implementation steps or resolve the reported conflict. It gives no evaluation of the ensemble or meta-model, so it offers a framing of the workflow rather than evidence that combining the scores improves ranking or returns.

Key ideas

• The proposed workflow uses predictions from multiple persisted StockRanker models as inputs to a further model.
• The models share a factor set but differ in training dates, filters, or target horizons.
• The user reports column conflicts when combining prediction blocks in a visual workflow.
• A custom Python module is identified as the intended way to load models and generate scores.
• The post itself supplies no implementation details or evidence of predictive improvement.

Tags

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