LG Aimers Demand Forecasting Hackathon: Optimizing an Ensemble Through Systematic Experiments

Summary

• Original article title, translated: "LG Aimers Demand Forecasting Hackathon: Optimizing an Ensemble Through Systematic Experiments." Only the filename's unsupported colon was replaced. Published 2025-08-27 · slug lg-aimers-demand-forecasting-hackathon · category "Projects, hackathons, and competitions / ML and data competitions" · source article recorded in Korean and English.
• A retrospective on an LG Aimers hackathon about forecasting menu demand at food-and-beverage venues. The final private score was 0.48718, ranked 61st among 1,500 participants (top 4.2%); I did not reach the final round.
• After an initial plan combining complex feature engineering with tree and deep-learning models stalled, I switched to the validated V44 baseline: a simple average of three predictions, V19 with feature-based LGBM, XGBoost, and CatBoost; V31 with per-menu PatchTST time series; and V34.2 with global N-BEATS time series. The experimental rule was to hold the individual models fixed and change only how their predictions were combined.
• Failed hypothesis: adding features to V19 worsened the public score from 0.5030 to 0.5092. Predictions became more similar across models, reducing ensemble diversity. Successful hypotheses: increasing V19's weight to 0.4 yielded 0.5017; replacing zero predictions with one, motivated by SMAPE's heavy penalty on zero predictions, yielded 0.4921. The note reports a 15% RMSE reduction from the baseline with weighted ensembling and postprocessing.
• Conclusion: when several models already work well, refining how they are combined can be more effective than changing their internals. The final submitted code used weight 0.5 for V19 and 0.25 for each other prediction, followed by the zero-to-one postprocessing.

Mentioned

• LG Aimers 7th cohort: the hackathon discussed in the article. Its completion note records a discrepancy over whether the ranking denominator refers to people or teams.
• LG AI Research: the organizer of LG Aimers.

Connections

• Writing · Credentials
• Atomic notes: Inspect ensemble diversity when a single large model fails; extra features can make models too similar · Read the evaluation metric's penalty structure first; SMAPE heavily penalizes zero predictions · Refine the combination of good models before changing their internals
• Concepts: CatBoost · LightGBM · N-BEATS · PatchTST · SMAPE · TSB-AD and TAB benchmarks · XGBoost · i18n and multilingual support · Running a technical blog · Demand forecasting · Time series · Ensembles · Evaluation design · Feature engineering