A Machine Learning-Based Approach to Detecting Financial and Market Patterns Leading to Significant Stock Price Changes in Energy Companies

The study proposes machine learning models which help to detect financial and market patterns that result in significant stock price changes in European energy companies. The period taken for the analysis is from 2020 to the first quarter of 2025, because it is characterized by geopolitical tensions, policy shifts and transition towards renewable energy. The research focuses on ten major energy firms across oil and gas and electric utilities sectors. The dataset consists of weekly stock prices, technical indicators, macroeconomic data, and financial ratios. Given the complexity of data, the primary objective is not only predictive performance of the models but also model interpretability, achieved through SHAP analysis. The target variable sorts observations into three categories based on how the stock is expected to move over the next 10 days: it could either see a significant drop, remain neutral, or experience significant growth. The study employs six different models – Logistic Regression, Logistic Regression with L2 regularization (Ridge), Logistic Regression with L1 regularization (Lasso), Logistic Regression with Elastic Net regularization, Random Forest, and XGBoost. Logistic Regression with Lasso and Logistic Regression with Elastic Net demonstrate the most generalizable performance across training, validation, and test sets. Initial analysis reveals problems related to strong class imbalance and the aggregation of different energy sectors, which could lead to model bias. SHAP analysis shows that the most influential predictors vary significantly depending on the scope of analysis: when both sectors are combined, financial stability indicators dominate, but when focusing only on oil and gas, technical indicators and macroeconomic variables gain importance. Moreover, even within the same sector (i.e. utilities), companies show notable variation in stock price behavior. This suggests that machine learning techniques could be incorporated at the sector level, but they should be adjusted to account for company-specific characteristics.