{ "id": 666180, "title": "Stacked LightGBM Ensemble for Detecting Market Reversal Points", "content": "

1. Problem Overview

\n

The objective of this competition is to detect reversal points in U.S. equities using historical and engineered market features.
\nThe task is formulated as a multi-class classification problem, where each observation represents a market state that may indicate a potential trend reversal.

\n

Given the noisy nature of financial time-series data and the presence of class imbalance, the proposed solution focuses on robust target engineering, strong tree-based models, and ensemble learning.

\n
\n

2. Target Engineering

\n

The original target labels contain several closely related categories. To reduce label noise and improve class separability, the labels were consolidated into three semantically meaningful classes:

\n\n

This simplification leads to:

\n\n
\n

3. Data Preparation

\n\n
\n

4. Handling Class Imbalance

\n

The dataset exhibits significant class imbalance.
\nTo mitigate this issue, balanced class weights were computed and converted into sample weights, which were passed directly to the learning algorithms during training.

\n

This approach ensures that minority classes contribute proportionally to the loss function and leads to improved Macro F1 performance.

\n
\n

5. Cross-Validation Strategy

\n

A 5-fold Stratified K-Fold Cross-Validation strategy was employed to:

\n\n
\n

6. Base Models (Level-0 Learners)

\n

To maximize ensemble diversity, three different LightGBM models were trained:

\n

Model 1: LightGBM (GBDT – Default)

\n\n

Model 2: LightGBM (GBDT – Tuned)

\n\n

Model 3: LightGBM (DART)

\n\n

All base models:

\n\n
\n

7. Evaluation Metric

\n

Macro F1 Score was selected as the primary evaluation metric because:

\n\n
\n

8. Ensemble Method: Stacking

\n

Instead of simple averaging, a stacking ensemble approach was implemented.

\n

Level-0 Output

\n\n

Level-1 Meta Learner

\n\n
\n

9. Meta-Model Validation

\n

The meta-learner was evaluated using cross-validated OOF predictions and achieved a higher Macro F1 score than any individual base model.

\n

This confirms that the stacking framework successfully captures the complementary strengths of the base learners and improves overall performance.

\n
", "comments": [ { "id": 3386724, "author_name": "tomy4reel", "author_url": "", "post_date": "2026-01-05T18:42:44.487000", "content": "

Thanks for sharing Mohamed👍.

\n

What were your private and public leaderboard scores on the new test set

", "votes": 0, "replies": [ { "id": 3386746, "author_name": "Mohamed Zakaria", "author_url": "", "post_date": "2026-01-05T19:04:23.587000", "content": "

Hi @tomy4reel,

\n

For the new test set, my solution achieved:\nPublic Leaderboard (97%)\nPrivate Leaderboard (96%)

", "votes": 0, "replies": [] } ] } ], "comment_count": 2, "messages": [ { "id": 3386705, "postDate": "2026-01-05T17:59:41.307Z", "content": "

1. Problem Overview

\n

The objective of this competition is to detect reversal points in U.S. equities using historical and engineered market features.
\nThe task is formulated as a multi-class classification problem, where each observation represents a market state that may indicate a potential trend reversal.

\n

Given the noisy nature of financial time-series data and the presence of class imbalance, the proposed solution focuses on robust target engineering, strong tree-based models, and ensemble learning.

\n
\n

2. Target Engineering

\n

The original target labels contain several closely related categories. To reduce label noise and improve class separability, the labels were consolidated into three semantically meaningful classes:

\n\n

This simplification leads to:

\n\n
\n

3. Data Preparation

\n\n
\n

4. Handling Class Imbalance

\n

The dataset exhibits significant class imbalance.
\nTo mitigate this issue, balanced class weights were computed and converted into sample weights, which were passed directly to the learning algorithms during training.

\n

This approach ensures that minority classes contribute proportionally to the loss function and leads to improved Macro F1 performance.

\n
\n

5. Cross-Validation Strategy

\n

A 5-fold Stratified K-Fold Cross-Validation strategy was employed to:

\n\n
\n

6. Base Models (Level-0 Learners)

\n

To maximize ensemble diversity, three different LightGBM models were trained:

\n

Model 1: LightGBM (GBDT – Default)

\n\n

Model 2: LightGBM (GBDT – Tuned)

\n\n

Model 3: LightGBM (DART)

\n\n

All base models:

\n\n
\n

7. Evaluation Metric

\n

Macro F1 Score was selected as the primary evaluation metric because:

\n\n
\n

8. Ensemble Method: Stacking

\n

Instead of simple averaging, a stacking ensemble approach was implemented.

\n

Level-0 Output

\n\n

Level-1 Meta Learner

\n\n
\n

9. Meta-Model Validation

\n

The meta-learner was evaluated using cross-validated OOF predictions and achieved a higher Macro F1 score than any individual base model.

\n

This confirms that the stacking framework successfully captures the complementary strengths of the base learners and improves overall performance.

\n
", "rawMarkdown": "## **1. Problem Overview**\nThe objective of this competition is to detect **reversal points in U.S. equities** using historical and engineered market features. \nThe task is formulated as a **multi-class classification problem**, where each observation represents a market state that may indicate a potential trend reversal.\n\nGiven the noisy nature of financial time-series data and the presence of class imbalance, the proposed solution focuses on **robust target engineering, strong tree-based models, and ensemble learning**.\n\n---\n\n## **2. Target Engineering**\nThe original target labels contain several closely related categories. To reduce label noise and improve class separability, the labels were consolidated into **three semantically meaningful classes**:\n\n- **H (High-type reversal)** → `{HH, LH, H}`\n- **L (Low-type reversal)** → `{HL, LL, L}`\n- **None** → all remaining cases\n\nThis simplification leads to:\n- Improved model stability \n- Higher **Macro F1** performance \n- Better generalization on unseen data \n\n---\n\n## **3. Data Preparation**\n- Training data from both the **original and newly released datasets** were merged to increase data coverage.\n- Non-informative identifiers and time-related columns were removed.\n- Missing values were handled using **zero imputation**.\n- Feature alignment between training and test sets was strictly enforced to ensure consistency.\n\n---\n\n## **4. Handling Class Imbalance**\nThe dataset exhibits significant class imbalance. \nTo mitigate this issue, **balanced class weights** were computed and converted into **sample weights**, which were passed directly to the learning algorithms during training.\n\nThis approach ensures that minority classes contribute proportionally to the loss function and leads to improved **Macro F1** performance.\n\n---\n\n## **5. Cross-Validation Strategy**\nA **5-fold Stratified K-Fold Cross-Validation** strategy was employed to:\n- Preserve class distribution across folds \n- Provide reliable **out-of-fold (OOF)** performance estimates \n- Reduce overfitting and variance \n\n---\n\n## **6. Base Models (Level-0 Learners)**\nTo maximize ensemble diversity, **three different LightGBM models** were trained:\n\n### **Model 1: LightGBM (GBDT – Default)**\n- Strong baseline model \n- Balanced bias–variance tradeoff \n\n### **Model 2: LightGBM (GBDT – Tuned)**\n- Larger tree capacity \n- L1/L2 regularization \n- Optimized feature and bagging fractions \n\n### **Model 3: LightGBM (DART)**\n- Dropout-based boosting strategy \n- Reduces correlation between trees \n- Improves ensemble robustness \n\nAll base models:\n- Used the **`multiclassova`** objective \n- Applied **early stopping** \n- Were evaluated using **Macro F1 Score**\n\n---\n\n## **7. Evaluation Metric**\n**Macro F1 Score** was selected as the primary evaluation metric because:\n- It treats all classes equally \n- It is robust to class imbalance \n- It better reflects real-world classification performance than accuracy \n\n---\n\n## **8. Ensemble Method: Stacking**\nInstead of simple averaging, a **stacking ensemble** approach was implemented.\n\n### **Level-0 Output**\n- Out-of-fold probability predictions from all three LightGBM base models \n\n### **Level-1 Meta Learner**\n- **Multinomial Logistic Regression** \n- Chosen for its simplicity, stability, and effectiveness in combining probabilistic outputs \n- Regularization helps prevent overfitting \n\n---\n\n## **9. Meta-Model Validation**\nThe meta-learner was evaluated using **cross-validated OOF predictions** and achieved a **higher Macro F1 score than any individual base model**.\n\nThis confirms that the stacking framework successfully captures the **complementary strengths** of the base learners and improves overall performance.\n\n---\n", "votes": 9 }, { "id": 3386724, "postDate": "2026-01-05T18:42:44.487Z", "content": "

Thanks for sharing Mohamed👍.

\n

What were your private and public leaderboard scores on the new test set

", "rawMarkdown": "Thanks for sharing Mohamed👍. \n\nWhat were your private and public leaderboard scores on the new test set", "replies": [ { "id": 3386746, "postDate": "2026-01-05T19:04:23.587Z", "content": "

Hi @tomy4reel,

\n

For the new test set, my solution achieved:\nPublic Leaderboard (97%)\nPrivate Leaderboard (96%)

", "rawMarkdown": "Hi @tomy4reel,\n\nFor the new test set, my solution achieved:\nPublic Leaderboard (97%)\nPrivate Leaderboard (96%)" } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T07:13:04Z", "competition_id": 114135, "competition_slug": "detecting-reversal-points-in-us-equities", "team_id": 15036546, "solution_url": "" }