Personal
Paper Trader
A backtesting lab built to catch itself overfitting — evolved strategies, strict train/test splits, lookahead audits.
- Year
- 2026
- Role
- Solo
- Stack
- Python, pandas, NumPy, SciPy
The idea
Most backtests lie, usually by accident: they peek at future data or get tuned until they look good. This is a sandbox for comparing trading models on historical data with fake money, designed so it's hard to fool yourself.
How it works
Strategy families evolve on 2020–2023 data only; 2024 onward is held out for testing. Finalists have to pass statistical gates — probability of backtest overfitting and a deflated Sharpe ratio — before they're compared against simply holding VOO.
An audit script checks every model for lookahead bias and pricing errors, option pricing is calibrated against real quotes, and runs are pinned to be exactly reproducible. Each (starting cash, start date) pair is its own fresh simulation, since a small account can't take some trades at all.