Hi everyone,
\nI’m still a beginner and actively learning, so I’d really appreciate any guidance or help from the community. I've been actively participating in this competition and have hit a bit of a wall. After trying out various models, advanced feature engineering, and multiple ensembling strategies, I still can't push my public score beyond 0.0965. Below is a detailed breakdown of my experiments so far. I'd love your feedback, ideas, or pointers if you've faced similar bottlenecks!
✅ Individual Model Scores (Public LB)
\nModel Score
\nLightGBM 0.10489
\nCatBoost 0.06139
\nXGBoost 0.10001
\nHistGBM 0.08649
\nRidge 0.08280
\nMLP 0.00996 (worst)
-> What I’ve Tried So Far
\nRealized volatility proxies
\nPressure × spread interactions
\nLiquidity imbalance
\nRolling stats, z-scores, etc.
\nFinal dataset size is ~13 GB+.
Weighted Ensembling with Optuna
\nUsed holdout predictions from best-performing models.
\nTried Optuna to find best weighted combination (maximizing Pearson).
\nResult: Score decreased to ~0.10176 🤯
\n(Worse than my best individual model — LGBM)
Holdout Strategy with Time Awareness
\nCreated time-aware holdout sets (15%)
\nMulti-seed training (3 seeds × 5 folds)
\nValidation scores were okay, but leaderboard didn’t improve.
LightGBM Focused Tuning
\nTried multiple configs — including dart, goss, and high-capacity GBDT with GPU.
\nBest LightGBM params:
\n'num_leaves': 2048, 'learning_rate': 0.005, 'reg_alpha': 0.1,
\n'reg_lambda': 0.1, 'bagging_fraction': 0.8, 'feature_fraction': 0.8,
\n'n_estimators': 5000, 'early_stopping_rounds': 300, 'device': 'gpu'
❌ Still Stuck — Main Issues:
\nWhy is ensembling degrading performance?
\nMy top models perform decently on their own — but Optuna-weighted predictions underperform even basic LightGBM!
Does meta-modeling (stacking) make sense here?
\nI was thinking of training a meta-learner (e.g., Ridge/XGB/LGB) on out-of-fold preds from my base models. Has anyone had success with that in this competition?
Pearson Optimization ≠ RMSE Optimization?
\nMost models are minimizing RMSE — but the leaderboard uses Pearson correlation. Could this mismatch be hurting performance?
Any tips for final push beyond 0.0965?
\nShould I reduce feature space using SHAP?
\nTry different ensembling (GBRank, voting, stacking)?
\nAny post-processing tips that helped you?
🧠 What I'm Looking For
\nIf you've:
\nManaged to boost your LB score beyond 0.10
\nGotten value from stacking or meta-modeling
\nFound Pearson-specific tricks that helped
\nUsed clever time-series validation
I’d love to hear how you approached it!
\nI’m happy to share more code or notebook logic if needed — thanks in advance for helping me.
Happy modeling!
", "messages": [ { "id": "3219350", "postDate": "06/07/2025 14:55:39", "content": "Hi everyone,
\nI’m still a beginner and actively learning, so I’d really appreciate any guidance or help from the community. I've been actively participating in this competition and have hit a bit of a wall. After trying out various models, advanced feature engineering, and multiple ensembling strategies, I still can't push my public score beyond 0.0965. Below is a detailed breakdown of my experiments so far. I'd love your feedback, ideas, or pointers if you've faced similar bottlenecks!
✅ Individual Model Scores (Public LB)
\nModel Score
\nLightGBM 0.10489
\nCatBoost 0.06139
\nXGBoost 0.10001
\nHistGBM 0.08649
\nRidge 0.08280
\nMLP 0.00996 (worst)
-> What I’ve Tried So Far
\nRealized volatility proxies
\nPressure × spread interactions
\nLiquidity imbalance
\nRolling stats, z-scores, etc.
\nFinal dataset size is ~13 GB+.
Weighted Ensembling with Optuna
\nUsed holdout predictions from best-performing models.
\nTried Optuna to find best weighted combination (maximizing Pearson).
\nResult: Score decreased to ~0.10176 🤯
\n(Worse than my best individual model — LGBM)
Holdout Strategy with Time Awareness
\nCreated time-aware holdout sets (15%)
\nMulti-seed training (3 seeds × 5 folds)
\nValidation scores were okay, but leaderboard didn’t improve.
LightGBM Focused Tuning
\nTried multiple configs — including dart, goss, and high-capacity GBDT with GPU.
\nBest LightGBM params:
\n'num_leaves': 2048, 'learning_rate': 0.005, 'reg_alpha': 0.1,
\n'reg_lambda': 0.1, 'bagging_fraction': 0.8, 'feature_fraction': 0.8,
\n'n_estimators': 5000, 'early_stopping_rounds': 300, 'device': 'gpu'
❌ Still Stuck — Main Issues:
\nWhy is ensembling degrading performance?
\nMy top models perform decently on their own — but Optuna-weighted predictions underperform even basic LightGBM!
Does meta-modeling (stacking) make sense here?
\nI was thinking of training a meta-learner (e.g., Ridge/XGB/LGB) on out-of-fold preds from my base models. Has anyone had success with that in this competition?
Pearson Optimization ≠ RMSE Optimization?
\nMost models are minimizing RMSE — but the leaderboard uses Pearson correlation. Could this mismatch be hurting performance?
Any tips for final push beyond 0.0965?
\nShould I reduce feature space using SHAP?
\nTry different ensembling (GBRank, voting, stacking)?
\nAny post-processing tips that helped you?
🧠 What I'm Looking For
\nIf you've:
\nManaged to boost your LB score beyond 0.10
\nGotten value from stacking or meta-modeling
\nFound Pearson-specific tricks that helped
\nUsed clever time-series validation
I’d love to hear how you approached it!
\nI’m happy to share more code or notebook logic if needed — thanks in advance for helping me.
Happy modeling!
", "rawMarkdown": "Hi everyone,\nI’m still a beginner and actively learning, so I’d really appreciate any guidance or help from the community. I've been actively participating in this competition and have hit a bit of a wall. After trying out various models, advanced feature engineering, and multiple ensembling strategies, I still can't push my public score beyond 0.0965. Below is a detailed breakdown of my experiments so far. I'd love your feedback, ideas, or pointers if you've faced similar bottlenecks!\n\n✅ Individual Model Scores (Public LB)\nModel\t Score\nLightGBM\t0.10489\nCatBoost\t 0.06139\nXGBoost\t 0.10001\nHistGBM\t 0.08649\nRidge\t 0.08280\nMLP\t 0.00996 (worst)\n\n-> **What I’ve Tried So Far**\n1. Advanced Feature Engineering\nBuilt a pipeline generating over 900+ features, including:\n\nRealized volatility proxies\nPressure × spread interactions\nLiquidity imbalance\nRolling stats, z-scores, etc.\nFinal dataset size is ~13 GB+.\n\n2. Weighted Ensembling with Optuna\nUsed holdout predictions from best-performing models.\nTried Optuna to find best weighted combination (maximizing Pearson).\nResult: Score decreased to ~0.10176 🤯\n(Worse than my best individual model — LGBM)\n\n3. Holdout Strategy with Time Awareness\nCreated time-aware holdout sets (15%)\nMulti-seed training (3 seeds × 5 folds)\nValidation scores were okay, but leaderboard didn’t improve.\n\n4. LightGBM Focused Tuning\nTried multiple configs — including dart, goss, and high-capacity GBDT with GPU.\nBest LightGBM params:\n'num_leaves': 2048, 'learning_rate': 0.005, 'reg_alpha': 0.1, \n'reg_lambda': 0.1, 'bagging_fraction': 0.8, 'feature_fraction': 0.8,\n'n_estimators': 5000, 'early_stopping_rounds': 300, 'device': 'gpu'\n\n\n\n❌ Still Stuck — Main Issues:\nWhy is ensembling degrading performance?\nMy top models perform decently on their own — but Optuna-weighted predictions underperform even basic LightGBM!\n\nDoes meta-modeling (stacking) make sense here?\nI was thinking of training a meta-learner (e.g., Ridge/XGB/LGB) on out-of-fold preds from my base models. Has anyone had success with that in this competition?\n\nPearson Optimization ≠ RMSE Optimization?\nMost models are minimizing RMSE — but the leaderboard uses Pearson correlation. Could this mismatch be hurting performance?\n\nAny tips for final push beyond 0.0965?\nShould I reduce feature space using SHAP?\nTry different ensembling (GBRank, voting, stacking)?\nAny post-processing tips that helped you?\n\n🧠 What I'm Looking For\nIf you've:\nManaged to boost your LB score beyond 0.10\nGotten value from stacking or meta-modeling\nFound Pearson-specific tricks that helped\nUsed clever time-series validation\n\nI’d love to hear how you approached it!\nI’m happy to share more code or notebook logic if needed — thanks in advance for helping me.\n\nHappy modeling!", "votes": null }, { "id": "3219498", "postDate": "06/07/2025 20:09:31", "content": "I guess why ensemble method here degrades final performance could be: there are too many sub-optimal models in the list: Since the data here is quite noisy, the performance metrics will be noisy as well, as a result the optimized ensemble weight may fit a lot of noise
", "rawMarkdown": "I guess why ensemble method here degrades final performance could be: there are too many sub-optimal models in the list: Since the data here is quite noisy, the performance metrics will be noisy as well, as a result the optimized ensemble weight may fit a lot of noise", "votes": null } ], "comments": [ { "id": 3219498, "author_name": "alexzhongs", "author_url": "", "post_date": "06/07/2025 20:09:31", "content": "I guess why ensemble method here degrades final performance could be: there are too many sub-optimal models in the list: Since the data here is quite noisy, the performance metrics will be noisy as well, as a result the optimized ensemble weight may fit a lot of noise
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3219350": "Hi everyone,\nI’m still a beginner and actively learning, so I’d really appreciate any guidance or help from the community. I've been actively participating in this competition and have hit a bit of a wall. After trying out various models, advanced feature engineering, and multiple ensembling strategies, I still can't push my public score beyond 0.0965. Below is a detailed breakdown of my experiments so far. I'd love your feedback, ideas, or pointers if you've faced similar bottlenecks!\n\n✅ Individual Model Scores (Public LB)\nModel\t Score\nLightGBM\t0.10489\nCatBoost\t 0.06139\nXGBoost\t 0.10001\nHistGBM\t 0.08649\nRidge\t 0.08280\nMLP\t 0.00996 (worst)\n\n-> **What I’ve Tried So Far**\n1. Advanced Feature Engineering\nBuilt a pipeline generating over 900+ features, including:\n\nRealized volatility proxies\nPressure × spread interactions\nLiquidity imbalance\nRolling stats, z-scores, etc.\nFinal dataset size is ~13 GB+.\n\n2. Weighted Ensembling with Optuna\nUsed holdout predictions from best-performing models.\nTried Optuna to find best weighted combination (maximizing Pearson).\nResult: Score decreased to ~0.10176 🤯\n(Worse than my best individual model — LGBM)\n\n3. Holdout Strategy with Time Awareness\nCreated time-aware holdout sets (15%)\nMulti-seed training (3 seeds × 5 folds)\nValidation scores were okay, but leaderboard didn’t improve.\n\n4. LightGBM Focused Tuning\nTried multiple configs — including dart, goss, and high-capacity GBDT with GPU.\nBest LightGBM params:\n'num_leaves': 2048, 'learning_rate': 0.005, 'reg_alpha': 0.1, \n'reg_lambda': 0.1, 'bagging_fraction': 0.8, 'feature_fraction': 0.8,\n'n_estimators': 5000, 'early_stopping_rounds': 300, 'device': 'gpu'\n\n\n\n❌ Still Stuck — Main Issues:\nWhy is ensembling degrading performance?\nMy top models perform decently on their own — but Optuna-weighted predictions underperform even basic LightGBM!\n\nDoes meta-modeling (stacking) make sense here?\nI was thinking of training a meta-learner (e.g., Ridge/XGB/LGB) on out-of-fold preds from my base models. Has anyone had success with that in this competition?\n\nPearson Optimization ≠ RMSE Optimization?\nMost models are minimizing RMSE — but the leaderboard uses Pearson correlation. Could this mismatch be hurting performance?\n\nAny tips for final push beyond 0.0965?\nShould I reduce feature space using SHAP?\nTry different ensembling (GBRank, voting, stacking)?\nAny post-processing tips that helped you?\n\n🧠 What I'm Looking For\nIf you've:\nManaged to boost your LB score beyond 0.10\nGotten value from stacking or meta-modeling\nFound Pearson-specific tricks that helped\nUsed clever time-series validation\n\nI’d love to hear how you approached it!\nI’m happy to share more code or notebook logic if needed — thanks in advance for helping me.\n\nHappy modeling!", "3219498": "I guess why ensemble method here degrades final performance could be: there are too many sub-optimal models in the list: Since the data here is quite noisy, the performance metrics will be noisy as well, as a result the optimized ensemble weight may fit a lot of noise" }, "source": "meta" }