Stock Minimum Price Classification

A research-oriented comparison of SVM and XGBoost using statistical feature screening and exhaustive feature-subset evaluation.

The research question

Can lagged financial ratios classify whether a stock's closing price is above a fixed minimum-price threshold?

The experiment treats this as a binary classification problem, screens candidate indicators with Cramér's V, evaluates every non-empty subset of the ten highest-ranked features, and compares Support Vector Machine and XGBoost models.

Evidence at a glance

Experimental signal	Recorded result
Candidate features after screening	10
Feature subsets evaluated per model	1,023
Best subset size	5 features
Best SVM test accuracy	70.7%
Recorded XGBoost test accuracy	69.5%
Research recognition	Best Paper · ICoDSA 2025

The best recorded SVM subset contains DERn-1, QRn-2, QRn-1, PTBVn-1, and PBVn-1.

Experimental pipeline

flowchart LR
    A["Financial-ratio dataset"] --> B["Clean and encode"]
    B --> C["Binary target: close > 50"]
    C --> D["Cramér's V screening"]
    D --> E["Top 10 candidate features"]
    E --> F["1,023 feature subsets"]
    F --> G["SVM experiments"]
    F --> H["XGBoost experiments"]
    G --> I["Kernel and parameter comparison"]
    H --> J["Model comparison"]
    I --> K["Evaluation and interpretation"]
    J --> K

Loading

Why exhaustive subsets?

Selecting the top ten features individually does not guarantee that they work well together. Enumerating all 2¹⁰ − 1 non-empty combinations tests compact and larger subsets under the same experimental setup.

This makes the feature-selection decision inspectable:

• Which ratios repeatedly appear in high-performing subsets?
• Does adding more variables improve generalization?
• Do nonlinear SVM and tree-based models prefer different combinations?
• Is a smaller model competitive with a larger one?

Model comparison

Model	Test accuracy	Main observation
Tuned SVM with RBF kernel	70.7%	Best recorded holdout result
XGBoost	69.5%	Competitive, but showed stronger overfitting in the recorded run

The selected features connect three interpretable financial themes:

• Leverage: Debt-to-Equity Ratio.
• Liquidity: lagged Quick Ratios.
• Valuation: Price-to-Book-related indicators.

Reproduce the notebook

Requires Python 3.8 or newer.

git clone https://github.com/akbaralqahri/Stocks_Classification.git
cd Stocks_Classification

python -m venv .venv

Activate and install:

# Windows
.venv\Scripts\activate

# macOS / Linux
source .venv/bin/activate

pip install -r requirements.txt jupyterlab
jupyter lab TA5.ipynb

Repository map

TA5.ipynb              complete experiment and recorded outputs
cramers_v_results.csv  exported statistical feature ranking
requirements.txt       reproducible Python dependencies
main.py                reserved application entry point

Methodological limitations

• The target is a fixed nominal-price threshold, not return, direction, or risk-adjusted performance.
• Exhaustive subset selection against a holdout score can produce an optimistic model-selection estimate.
• The current experiment uses a stratified split rather than a walk-forward market evaluation.
• External market regimes, macroeconomic variables, transaction costs, and liquidity are not modeled.
• The results are a research benchmark and should not be interpreted as a trading signal.

Recommended next steps are nested cross-validation, a time-ordered holdout, probability calibration, regime testing, and a target tied to future return rather than absolute nominal price.

---

Built by Muhammad Ali Akbar Al-Qahri as a transparent machine-learning research case study.