Quantitative Modeling Differences and GAT Lookback Tradeoffs
Summary
The discussion compares quantitative research practices in China and overseas. It says Chinese investors more often try to predict stock prices directly, while overseas researchers may estimate missing observations, factor values, company revenue, or margins. The speakers attribute these tendencies partly to differences in data standardization and market complexity, and mention causal analysis and generated alternative histories as areas of overseas research.
A second exchange addresses the lookback period for a graph attention network factor. The historical window can be varied, but searching many candidate lengths increases both overfitting risk and computation time. As an example, the speaker describes training costs for a model using 200 factors and 300 stocks on a GTX1060, with annual rolling calculations taking several hours. The discussion does not report predictive performance or establish an optimal window, and the hardware and timing figures are illustrative rather than a general benchmark.
Key ideas
- Quantitative research practices may differ with data standardization and market complexity.
- Overseas examples include imputing missing data and estimating business or factor variables instead of predicting prices directly.
- Causal analysis and generated alternative histories are discussed as research directions for complex financial data.
- Graph attention network lookback length can be varied, but broader parameter searches raise overfitting and computation concerns.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.