Case study · Software and quantitative research
Algorithmic trading systems
A safety-first research and execution system for studying trading ideas without confusing a backtest with a live result.
Trading research can look convincing while quietly mixing future information into features, treating incomplete data as final, or ignoring the mechanics of orders and fills. I wanted a system where those boundaries were explicit and testable.
View algos_public on GitHubWhat I built
I built the research and paper-execution workflow around deterministic replay, causal feature construction, walk-forward evaluation, portfolio planning, risk controls, and an auditable execution journal. The system is designed to fail closed when its assumptions or inputs are not satisfied.
The public repository is an earlier snapshot that predates my current generative-AI-assisted development workflow. I keep it public as an example of the research, system design, and coding approach behind the work; later development has continued privately beyond that snapshot.
Architecture and stack
The system is a Python research and execution codebase organized around typed domain contracts. Market data is normalized into canonical forms, stored and replayed deterministically, passed through causal research primitives, and evaluated through a next-bar event-driven backtester. Portfolio plans then pass through risk and execution state machines before reaching a deterministic paper exchange.
- Research: causal features, walk-forward splits, and economic evaluation
- Data: normalized market bars, quality checks, and checksummed local trade logs
- Execution: typed portfolio planning, risk controls, reconciliation, and paper fills
- Engineering: Python, strict contracts, property-focused tests, and reproducible local runs
Difficult engineering decisions
The most important decisions were about what the system would refuse to claim. Finalized bars cannot be fabricated, future data cannot be backfilled into a feature, and a paper result is not presented as proof of a profitable strategy. The execution path keeps planning, risk, order state, fills, and reconciliation separate so each boundary can be tested on its own.
I also treated durability and provenance as part of the engineering problem: records are checksummed, replay is deterministic, and the research identity includes the data, code, evaluation, and runtime assumptions behind a result.
What I learned
The hard part of quantitative software is not producing a signal. It is building enough discipline around data timing, evaluation, execution, and failure handling that the signal can be interpreted honestly. This project strengthened my interest in systems where correctness is a property of the entire workflow, not just one model or function.