earlyann / stock_predictions Goto Github PK

In this project, I delved into the world of stock predictions by comparing and evaluating the effectiveness of three distinct machine learning models - Support Vector Machine (SVM), Balanced Random Forest (BRF), and Easy Ensemble (EE) - based on stock-picking concept weights. My primary aim was to refine traditional models and tailor weighting combinations to identify the best fit for different investment scenarios. The dataset is from the US stock market historical database and contains performances of weighted scoring stock portfolios are obtained with mixture design. Source: https://archive.ics.uci.edu/ml/datasets/stock+portfolio+performance

Support Vector Machine (SVM) model: This supervised machine learning algorithm specializes in classifying stocks based on their potential for positive or negative annual returns. The algorithm identifies the optimal hyperplane that separates data points into two categories, creating a decision boundary.

Balanced Random Forest Model: A variation of the conventional Random Forest model, BRF balances class weights, allowing it to handle imbalanced datasets more effectively and produce more reliable predictions.

Easy Ensemble model: This ensemble learning method combines multiple base classifiers to improve prediction performance. It focuses on resampling the dataset to balance class distribution, which enhances overall model accuracy.

Python Pandas: Data manipulation Scikit-learn: SVM model training and evaluation Imbalanced-learn: Balanced Random Forest and Easy Ensemble model training and evaluation Matplotlib: Data visualization

• Data collection and preprocessing: I obtained and preprocessed a dataset from the UCI Machine Learning Repository, ensuring data quality and usability.
• Feature engineering: I transformed the dataset into SVM model inputs by extracting stock-picking concept weights.
• Model training: Utilizing the scikit-learn and imbalanced-learn libraries, I trained the SVM, Balanced Random Forest, and Easy Ensemble models.
• Model evaluation: I gauged each model's performance using accuracy, precision, and recall metrics to determine their effectiveness.

I leveraged a dataset containing historical US stock market data and stock-picking concept weights derived from a weighted scoring stock selection model to predict stocks' annual returns.

To visualize the relationships between the various stock-picking concepts, I employed pair plots, which included histograms of each feature along the diagonal. A key observation was the relative scarcity of negative annual results, which prompted me to fine-tune each model to successfully predict instances of the minority class (negative annual results).

Each model input (X1-X6) represented weights of Large B/P, Large ROE, Large S/P, Large Return Rate in the last quarter, Large Market Value, and Small Systematic Risk concepts. The output was a binary label indicating positive or negative annual return based on a 0.0 threshold. Higher weights implied stronger emphasis on the corresponding concept. Each model was designed to learn the relationship between stock-picking concept weights (X1-X6) and the binary label in order to make predictions.

After training, each model could predict annual returns based on the concept weights. The training, hypertuning, and testing of the SVM model are depicted in the following images.

I present the accuracy, precision, recall, and f1-score for each model:

Accuracy: 1.0 Precision: 1.00 Recall: 1.00 F1-score: 1.00

Accuracy: 1.0 Precision: 1.00 Recall: 1.00 F1-score: 1.00

Accuracy: 1.0 Precision: 1.00 Recall: 1.00 F1-score: 1.00

Each model was able to successfully identify positive and negative annual returns 100% of the time.

I developed a function that takes one of my predefined models (SVM, BRF, or EE) and a set of X1-X6 data and predicts whether that set of data will yield a positive or negative annual result. The model parameter in the function accepts any model that has a predict method, which is true for most machine learning models in scikit-learn and imbalanced-learn. I was able to pass my previously defined models svm_best_model, brf_best_model, or ee_best_model as the first argument to the function, and it used the respective model for making predictions. This made it easy to test and implement different models within my framework and offered flexibility for users to explore other machine learning models that fit their specific needs.

My project successfully demonstrated the application of three machine learning models - Support Vector Machine, Balanced Random Forest, and Easy Ensemble - for predicting positive or negative annual returns based on stock-picking concept weights. Each model showed excellent performance in terms of accuracy, precision, recall, and f1-score, emphasizing the potential for machine learning to enhance investment strategies.

By comparing the strengths of these models, I realized that investors can choose the most suitable approach to guide their investment decisions, optimizing their portfolio based on their preferences and risk tolerance.

I also provided a flexible prediction function, compatible with various machine learning models, making it easy for users to test and implement different models within my framework.

In the future, I could explore predicting other investment performance indicators (Y2-Y6) and using an epsilon-SVR model for continuous value predictions. Additionally, incorporating more diverse datasets and exploring feature importance could further enhance the model's performance and provide deeper insights into the stock-picking process.