I built a workflow that connects strategy research with paper execution and operational review. It starts with EMA–ADX trend following and Z-score mean reversion, then follows the resulting orders, positions, and risk measures through separate services. The aim was to make a research run inspectable beyond its final equity curve.

The Python backtester writes results and configuration metadata to CSV and JSON. A separate executor submits paper orders through Alpaca, while an order listener observes broker updates. An independent risk worker reads positions and market data, records snapshots, and can enforce configured controls. PostgreSQL stores order and risk telemetry, and a Streamlit dashboard brings that information together with saved research runs. A standalone data guard checks freshness, gaps, and spreads.

The report follows the strategies from signal definitions and trading assumptions to historical comparisons with buy-and-hold. Saved figures show the equity paths; run metadata records the configuration behind them. The architecture below distinguishes the independent services and their shared data, so the implementation can be examined alongside the analysis.

This study connects return modeling to portfolio construction. I compared linear and neural-network models of factor-adjusted equity returns, then combined their outputs with a factor risk model to calculate portfolio weights. The analysis follows the modeling choices through to historical portfolio profit, exposures, and risk.

The notebook forms a market-cap-filtered universe, winsorizes raw returns, and removes exposures to 59 industry and six style factors. Dates divide the data into an initial 80% training period and a final 20% test period. Regularized weighted least squares provides the linear baseline, with a neural network as the comparison. Both modeling branches feed portfolio-weight calculations using a diagonal factor-covariance approximation, stock-specific variances, and the Woodbury identity.

The report develops the weight calculation and presents regularization studies alongside portfolio diagnostics. Its saved plots separate total risk from the stock-specific component and inspect long and short market values. The method and risk pages below let you follow the mathematics into those outputs, rather than judging the models only by a fitted prediction error.

I built an XGBoost classification study around SPY market data, exploring how features are selected, how model settings are tuned, and how the results are evaluated. The executed notebook labels whether the current day’s return exceeds 0.25%; it therefore studies same-day classification rather than a next-day forecasting target.

The workflow downloads daily price and volume history and constructs rolling statistics, technical indicators, lagged returns, and calendar features. Feature exploration combines mutual information, random-forest recursive elimination, and XGBoost importance. A separate seven-feature model set is used for the baseline and tuned classifiers, with a chronological 80/20 split and TimeSeriesSplit for tuning. The feature-selection stage occurs before the final holdout, an important distinction when interpreting the evaluation.

The report walks through preprocessing, selection, and model diagnostics. The saved ROC and historical strategy comparison use baseline probabilities, rather than the tuned model’s predictions. The comparison shifts the trading signal by one day and omits trading costs. Those details are kept visible so the original figures can be read as methodology evidence, with their assumptions intact.