Synthetic view of the ML pipeline and the compared strategy curves.
Back to projects
Quantitative Finance · Data Science · Machine Learning
Machine Learning Allocation Project
This project is linked to the portfolio backtesting and wealth-tracking stack: it is the machine-learning layer used to test signals and compare their impact on portfolio performance. The repository structures a full pipeline, from market-data preparation to test-metric production, with a reproducible logic built around dedicated scripts and run_all orchestration. The approach combines classic baselines such as Equal Weight and Markowitz Minimum Variance with supervised models like Random Forest and Logistic Regression in order to predict the direction of daily returns and convert those probabilities into portfolio weights.
PythonPandasNumPyscikit-learnMatplotlibSeaborn
What this project demonstrates
Ability to turn a quantitative topic into a usable ML pipeline: feature engineering, variable selection, multi-model training, comparison against financial baselines, and clear reporting of results.
My role
I contributed to project structuring, to the integration of strategy-comparison modules, and to the design of an end-to-end flow covering data preparation, training, backtesting, and reporting.
Context
The objective was to go beyond an exploratory notebook and produce a rigorous, collaborative working base with modular code, dedicated scripts, exported metrics, and directly usable comparison charts.
Objective
Assess the contribution of supervised models to portfolio allocation in a structured way while preserving a robust financial reference through Equal Weight and Markowitz baselines.
Deep dive
Technical reading of the project
The project follows a full quant and machine-learning workflow: price ingestion, technical-feature construction, directional model training by asset, portfolio backtest, and systematic comparison against financial baselines.
- • Data preparation and exploratory analysis through normalized prices, return distributions, correlations, and rolling volatility.
- • Financial baselines: Equal Weight Buy and Hold and Markowitz Minimum Variance.
- • Supervised modeling by asset through Random Forest and Logistic Regression.
- • ML strategy backtests on the test period with portfolio weights derived from probabilities.
- • Export of metrics and figures for comparative reading.
Gallery
Key screens and visualizations
Visual comparison of the Random Forest model against the financial baselines.
Architecture
Technical organization
Data layer
Modules such as src/data/load_data.py and src/data/preprocess.py load prices, compute returns, and create a proper temporal train or test split.
Feature engineering
Technical indicators are built in src/features/technical_indicators.py and ANOVA selection is applied through src/features/feature_selection.py to keep the most relevant variables.
Models and strategies
Financial baselines live in src/baselines while ML models such as Random Forest and Logistic Regression are implemented with the backtest logic in the model layer.
Orchestration and reporting
Scripts such as scripts/run_*.py and run_all.py execute the full pipeline, produce metric tables, and generate equity-curve comparisons.
Pipeline
Data flow
- 1.Load price series and compute daily returns.
- 2.Create technical features and save the consolidated dataset.
- 3.Split train and test chronologically and train the models asset by asset.
- 4.Project signals or probabilities into portfolio weights over the test period.
- 5.Compute equity curves and cross-strategy comparison metrics.
- 6.Export results to reports/tables and reports/figures for decision-oriented analysis.
Technical choices
Structuring decisions
Systematic comparison with baselines
Every ML strategy is evaluated against Equal Weight and Markowitz to keep a stable financial benchmark rather than relying only on a classification score.
Explicit temporal split
The train or test split follows market chronology in order to limit information leakage in time-series data.
ANOVA selection plus GridSearchCV
Random Forest combines feature reduction and hyperparameter tuning to avoid a purely arbitrary modeling approach.
Asset-level modeling, then portfolio aggregation
Predictions are generated at the ticker level and then aggregated through a weighting logic, which makes the model closer to a real allocation use case.
Reliability
Quality and controls
- • Dedicated scripts for each stage improve reproducibility and make debugging more targeted.
- • Metrics and test visualizations are exported in versionable files.
- • The modular structure across data, features, models, baselines, and scripts is compatible with collaborative work.
Limitations
Current attention points
- • Performance remains dependent on the historical window and the market regime observed.
- • Real market frictions are simplified, especially transaction costs, liquidity, and advanced slippage.
- • Validation can still be strengthened through more robust walk-forward protocols and stress testing.
Roadmap
Next steps
- • Add walk-forward validation and stricter out-of-sample robustness protocols.
- • Extend the model library with approaches such as gradient boosting under the same comparison framework.
- • Connect ML pipeline outputs automatically to the wealth-tracking modules.
Challenges
Main project challenges
Build a clean temporal train or test pipeline to limit information leakage in a time-series setting.
Compare fundamentally different approaches, from static allocations to supervised models, under homogeneous metrics.
Maintain a readable architecture despite the multiplication of components such as features, models, baselines, and orchestration scripts.
Outcomes and learnings
What I take away
Reproducible execution chain through run_prepare.py, run_baselines.py, run_random_forest.py, run_logistic_regression.py, and run_all.py.
Systematic production of test reports such as metrics_test_* and comparative figures such as equity_*_vs_baselines_test.png.
Integration of Random Forest with ANOVA feature selection and hyperparameter search through GridSearchCV.
Addition of a Logistic Regression supervised benchmark compared with classical financial strategies.
Other projects
Continue exploring
Personal Finance · Python · PyQt6
Wealth Tracking Application
PyQt6 + SQLite desktop application to centralize multi-asset accounts, rebuild weekly history, and analyze portfolio performance.
View this projectQuantitative Finance · Analysis · Python
Portfolio Backtesting and Optimization
Python environment to backtest strategies, compare risk and return metrics, and analyze portfolio behavior.
View this projectDiscuss
I can detail technical choices and outcomes during an interview.
If this project is relevant for you, I can detail the initial need, data structure, assumptions, challenges encountered, and analysis limitations.