Featured research

A Python research and execution workflow that keeps backtests, paper orders, and risk checks traceable.

The problem

Keep a strategy’s signals, orders, and risk checks traceable.

What I built

Cost-aware backtesting, paper-order execution, and services for data quality, exposure, VaR, and drawdown checks.

Scope

Backtests are historical simulations. Orders run through Alpaca’s paper endpoint by default.

Implementation

I built EMA–ADX trend-following and Z-score mean-reversion strategies, a vectorized backtester that accounts for trading costs, and a paper-trading executor for limit orders. Docker services and a shared database connect order and fill updates, data freshness checks, and a separate worker for value at risk (VaR), exposure, and drawdown controls.

Reviewing runs

Saved settings and results let me compare parameter choices and trace signals through orders, fills, and risk checks.

Research output

The strategy finishes below buy-and-hold in the saved simulation. This run gives me a concrete benchmark gap to investigate. Saved settings and results make that investigation repeatable across backtests and paper orders.

The featured figure is an unchanged saved backtest from the public repository. View the original figure ↗ or read the project report ↗.

What I built

I used Python and SQL to standardize Bloomberg equity and FX data, identifiers, trading calendars, and corporate actions. Input checks, saved configurations and results, and structured logs made the research runs repeatable.

I also used hidden Markov models (HMMs) and clustering to study market regimes, their stability, and drawdowns within each regime.

Comparing allocations

I checked the matrices and portfolio constraints, then compared mean–variance, Black–Litterman, and hierarchical risk parity allocations against saved benchmarks. I evaluated Sharpe ratios, drawdowns, turnover, and risk contributions.

Columbia · December 2024

What I built

I used a pseudoinverse to estimate residual returns and winsorized the targets to limit the influence of extreme values. I compared regularized weighted least squares with a neural-network baseline, then used the Woodbury identity to calculate portfolio weights from factor and stock-specific risk estimates.

Portfolio risk

The notebook reports the market values of long and short positions, profit and loss from residual returns, and total and stock-specific risk alongside the portfolio weights.

Data availability

Reproducing the notebook requires input data that are not included in the public repository.

CQF · June 2025

What I built

I created features from price, volatility, momentum, and calendar data. I compared mutual-information filtering, recursive feature elimination with random forests, and XGBoost feature selection. For tuning, I used a chronological holdout and five-fold time-series cross-validation.

Evaluation

The notebook includes feature-selection experiments, baseline and tuned model results, confusion matrices, and classification reports.

Limitations

The labels use same-day returns, and feature selection runs before the training and test data are split. To evaluate next-day forecasts, the target would need to be shifted forward and feature selection fitted only on training data.