The question.
A promising historical strategy can disappear once survivorship bias, transaction costs, slippage, and repeated parameter searches are accounted for.
This is a proposed research environment, rather than a live trading service. Its emphasis is reproducibility and risk-adjusted evaluation; a profitable backtest is not a promise of future returns.
The proposed approach.
- Version data snapshots and corporate-action handling
- Compare candidate signals with simple market and factor baselines
- Separate training, selection, and final out-of-sample periods
- Model fees, liquidity, slippage, exposure limits, and drawdowns before paper execution
Target architecture.
Market and event dataFeature researchWalk-forward testsExecution simulatorPaper portfolioRisk report
The difficult parts.
- Look-ahead leakage and stale market data
- Multiple testing and regime changes
- Translating theoretical signals into feasible fills
How to evaluate it.
- Out-of-sample risk-adjusted performance after modeled costs
- Drawdown, turnover, liquidity sensitivity, and concentration
- Parameter stability across periods and market regimes
Planned deliverables.
- A reproducible signal-research notebook collection
- A bias-aware backtesting and paper-trading workflow
- A strategy assumptions and risk report
Design principles.
Treat every backtest as an experiment.
Keep research conclusions separate from execution authority.