Manual backtesting without fooling yourself

Why test

On social media every method seems to work, because only the successful examples get shown. The only way to know whether a setup has an edge is to measure it over many cases, recorded with rules fixed in advance.

Procedure

• Write the entry, stop and target rules before looking at the data
• Scroll the historical chart candle by candle hiding the future (with the replay function or by covering the screen)
• When the conditions occur, record the trade: entry, stop, target, result in R
• Don't skip the uncomfortable cases and don't change the rules while testing
• Collect at least 100 trades over different periods (trend, range, high and low volatility)

Biases to avoid

• Look-ahead: seeing what follows and 'recognising' the setup in hindsight
• Selection: recording only the good-looking setups
• Overfitting: tweaking the rules until the past fits, then discovering it doesn't work on the future
• Ignored costs: spread, commissions and slippage reduce every trade, especially on low timeframes
• Small sample: 20 trades tell you nothing

Splitting the data

Use one part of the data to develop the rules and another, never looked at, to verify them. If the result collapses on the second sample, the rules were fitted to noise.

Exercise

• Test the module 3 model on 100 historical trades of one pair
• Subtract 1 pip of cost from every trade and recalculate the result
• Repeat on another period and another pair, without changing the rules