# Stacked LightGBM Ensemble for Detecting Market Reversal Points

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