Thanks to @ambrosm for this great baseline here
\nThis solution blends a lightweight statistical ensemble approach with time-series logic and handcrafted stability mechanisms. Despite its simplicity (no neural nets or heavy feature engineering), it performed competitively β placing 8th on the private leaderboard.
\nLocal Statistical Models\nEach sectorβs monthly transaction volume is modeled using log-space weighted averages of recent months β a robust alternative to exponential smoothing.
\nt1 months are aggregated using custom weight patterns (linear decay, exponential decay, Gaussian, etc.).Diverse Model Pool\nA large grid search explored combinations of:
\nt1 β {1, 3, 6, 9} (lookback window)t2 β {1, 2, 3} (minimum-activity threshold)CSI scale β {0.8, 0.9, 1.0, 1.1}Around 80 candidate models were ranked using a custom competition metric, and a subset of 20 strong + 22 diverse models was selected.\nDiversity was promoted via correlation-penalized ranking, ensuring complementary behavior across folds.
Time-Aware Cross-Validation
\nTimeSeriesSplit with 4 folds and 12-month test windows.Horizon-Specific Ensemble Optimization\nFinal blending was done per forecast horizon:
\nFor each bucket, SLSQP optimization was used to find non-negative weights summing to 1, maximizing out-of-fold score while lightly regularizing with L2 and correlation penalties.
Stable Inference\nEach model produced test predictions independently, then horizon-dependent ensemble weights were applied row-wise.\nThis avoided horizon leakage and improved stability on volatile sectors.
| Metric | \nScore | \n
|---|---|
| Public LB | \n0.59331 | \n
| Private LB | \n0.50689 | \n
XGB, LGBM, NN Models with lag features did not outperform this simple approach.
\nP.S. This was written with the help of ChatGPT
", "comments": [], "comment_count": 0, "messages": [ { "id": 3309765, "postDate": "2025-11-01T10:47:20.960Z", "content": "Thanks to @ambrosm for this great baseline here
\nThis solution blends a lightweight statistical ensemble approach with time-series logic and handcrafted stability mechanisms. Despite its simplicity (no neural nets or heavy feature engineering), it performed competitively β placing 8th on the private leaderboard.
\nLocal Statistical Models\nEach sectorβs monthly transaction volume is modeled using log-space weighted averages of recent months β a robust alternative to exponential smoothing.
\nt1 months are aggregated using custom weight patterns (linear decay, exponential decay, Gaussian, etc.).Diverse Model Pool\nA large grid search explored combinations of:
\nt1 β {1, 3, 6, 9} (lookback window)t2 β {1, 2, 3} (minimum-activity threshold)CSI scale β {0.8, 0.9, 1.0, 1.1}Around 80 candidate models were ranked using a custom competition metric, and a subset of 20 strong + 22 diverse models was selected.\nDiversity was promoted via correlation-penalized ranking, ensuring complementary behavior across folds.
Time-Aware Cross-Validation
\nTimeSeriesSplit with 4 folds and 12-month test windows.Horizon-Specific Ensemble Optimization\nFinal blending was done per forecast horizon:
\nFor each bucket, SLSQP optimization was used to find non-negative weights summing to 1, maximizing out-of-fold score while lightly regularizing with L2 and correlation penalties.
Stable Inference\nEach model produced test predictions independently, then horizon-dependent ensemble weights were applied row-wise.\nThis avoided horizon leakage and improved stability on volatile sectors.
| Metric | \nScore | \n
|---|---|
| Public LB | \n0.59331 | \n
| Private LB | \n0.50689 | \n
XGB, LGBM, NN Models with lag features did not outperform this simple approach.
\nP.S. This was written with the help of ChatGPT
", "rawMarkdown": "Thanks to @ambrosm for this great baseline [here](https://www.kaggle.com/code/ambrosm/red-explained-baseline)\n\n### Overview\n\nThis solution blends a lightweight statistical ensemble approach with time-series logic and handcrafted stability mechanisms. Despite its simplicity (no neural nets or heavy feature engineering), it performed competitively β placing **8th on the private leaderboard**.\n\n---\n\n### π‘ Key Ideas\n\n1. **Local Statistical Models**\n Each sectorβs monthly transaction volume is modeled using log-space weighted averages of recent months β a robust alternative to exponential smoothing.\n\n * The last `t1` months are aggregated using custom weight patterns (linear decay, exponential decay, Gaussian, etc.).\n * Predictions are adjusted with simple **seasonality multipliers** and a **scaled city search index (CSI)** trend factor.\n\n2. **Diverse Model Pool**\n A large grid search explored combinations of:\n\n * `t1 β {1, 3, 6, 9}` (lookback window)\n * `t2 β {1, 2, 3}` (minimum-activity threshold)\n * `CSI scale β {0.8, 0.9, 1.0, 1.1}`\n * Multiple weighting schemes (equal, decayed, Gaussian, etc.)\n\n Around **80 candidate models** were ranked using a **custom competition metric**, and a subset of **20 strong + 22 diverse** models was selected.\n Diversity was promoted via **correlation-penalized ranking**, ensuring complementary behavior across folds.\n\n3. **Time-Aware Cross-Validation**\n\n * Used `TimeSeriesSplit` with 4 folds and 12-month test windows.\n * All validation scores were computed using the exact competition scoring function.\n\n4. **Horizon-Specific Ensemble Optimization**\n Final blending was done **per forecast horizon**:\n\n * Short-term (1β4 months)\n * Mid-term (5β8 months)\n * Long-term (9β12 months)\n\n For each bucket, **SLSQP optimization** was used to find non-negative weights summing to 1, maximizing out-of-fold score while lightly regularizing with L2 and correlation penalties.\n\n5. **Stable Inference**\n Each model produced test predictions independently, then horizon-dependent ensemble weights were applied row-wise.\n This avoided horizon leakage and improved stability on volatile sectors.\n\n---\n\n### βοΈ Technical Notes\n\n* **Libraries**: Pure Python + NumPy/Pandas/SciPy + scikit-learn (for CV).\n* **No heavy ML models**: Entire pipeline executes in under a minute.\n* **Reproducibility**: Randomness controlled via deterministic NumPy ops and fixed splits.\n\n---\n\n### π Results\n\n| Metric | Score |\n| ---------- | ----- |\n| Public LB | 0.59331 |\n| Private LB | 0.50689 |\n\n---\n\n### What did not work\nXGB, LGBM, NN Models with lag features did not outperform this simple approach.\n\nP.S. This was written with the help of ChatGPT\n", "votes": 5 } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T00:52:29Z", "competition_id": 111876, "competition_slug": "china-real-estate-demand-prediction", "team_id": 14249175, "solution_url": "/competitions/china-real-estate-demand-prediction/writeups/8th-place-solution" }