I used an ensemble of GBTs with a single online component. Published my training code here.

Details

Ensemble:

• 3 x LGB, offline
• 4 x XGB
• 1 x LGB, online (re-trained every 5 days with a window of 60 days)

Training:

• Used n-fold cross validation by date_id with 5-day gaps between the folds.
• Hyperparameter selection: used public notebooks as a reference, optuna + manual tweaks as the final pick.

Features that I haven't seen in public notebooks:

• Revealed_target

• Intraday revealed_target using the wap from previous time steps

• Features based on revealed stocks:

  • sector_id
  • embeddings based on historical open/close/high/low data

• Group features

  • performance against the mean
  • performance against the sector

• At-the-money call price estimate with "expiry" at the end of auction

• Inferred price based on tick size

Post processing:

• Replaced "zero-sum" with subtraction by index-weighted mean targets.

Other:

• Dropped stock_id and relied on the embeddings based on historical performance and sector to reduce the effects of delisting and other drastic changes.

Stuff that didn't work:

• Most rolling features I introduced caused a drop in LB, so I omitted those such as:
  • Rolling cross-correlation
  • Rolling z-score
• Clustering stock_ids based on correlated wap, target
• Stock embeddings from neural networks

I got the biggest public LB boost from online learning, post-processing and hyperparameter tuning.

Acknowledgements

I re-used feature engineering code from:
https://www.kaggle.com/code/meli19/lgb-kf-baseline
https://www.kaggle.com/code/zulqarnainali/explained-singel-model-optiver/notebook
https://www.kaggle.com/code/judith007/lb-5-3405-rapids-gpu-speeds-up-feature-engineer
https://www.kaggle.com/code/verracodeguacas/fold-cv

Thanks to Kaggle and Optiver for the interesting competition. This is the first time I saw a competition through and learned a lot during the process.