Using TabNet to Rank Stocks from Price and Volume Factors

Summary

This article outlines a machine-learning stock-selection experiment using TabNet and 98 price-and-volume factors. It trains the model on data from 2010 through 2018, then applies its predictions in a backtest covering 2018 through September 2021. The write-up lists key TabNet settings, including input feature count, decision steps, dimensions for prediction and attention, attention-update scaling, and batch-normalization momentum.

The supplied text gives the broad training and evaluation periods but does not state the target variable, portfolio construction rules, transaction-cost assumptions, benchmark, or numerical backtest results. It labels the implementation as an older version intended for study, so its setup should not be taken as a current or fully specified investment process. The material is useful as a compact example of applying an attention-based tabular model to equity ranking, but it does not provide enough detail to reproduce or assess the reported experiment independently.

Key ideas

• The example applies TabNet to quantitative stock selection using price-and-volume inputs.
• It uses 98 factors and separates the training period from a later prediction and backtest period.
• The article identifies model settings for feature input, decision steps, attention, and normalization.
• The implementation is explicitly presented as an older educational version.
• The text omits portfolio rules and detailed performance evidence needed for independent evaluation.

Tags

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