Think in samples
A strategy can lose even when its long-run expectation is positive, and it can win temporarily without a real edge. A single outcome says very little.
- Number of observations
- Win and loss size distribution
- Maximum drawdown
- Costs and slippage
- Stability across regimes
Backtests are experiments
Historical results depend on data quality, parameter choices and assumptions about execution. The more ideas are tried, the easier it is to discover a pattern that exists only by chance.
- Keep validation data separate
- Avoid choosing parameters solely for the best curve
- Stress assumptions
- Confirm with forward and demo testing
Risk is part of the model
Position sizing and loss limits are not an add-on. They define whether the strategy can survive the normal variability of its own outcomes.