A backtest is a question: “Can this rule be executed, and what did it do in old data?” It is not a certificate.
I still run them. I no longer treat a pretty curve as permission to go large.
What I test
One rule, one market, one session.
If the description needs three indicators, two timeframes, and a discretionary “feel,” I do not have a test. I have a story. Write the rule in one sentence or do not start.
How I keep it honest
– Same stop definition on every trade. No “this one needed room.”
– Costs: a spread or commission guess, even if rough.
– Out of sample: leave a later stretch of dates untouched until the rule is frozen.
– No deleting losers after you see them. That is just curve-fitting with extra steps.
If I cannot stand to record the losers, I already know the method is a mood.
What I record from the test
– Number of trades (if it is 12, you learned almost nothing)
– Average R
– Max drawdown in R
– Longest losing streak
– How often I would have broken my own rule if this had been live
Win rate without the rest is a marketing slide.
After the test
A green test still goes to a live sample at half size or less. Twenty trades, kill lines written first. That handoff is sizing the first 20 and kill criteria.
A red test is useful. It saved live money. Retire it. Do not add an extra average until it turns green. That extra average is how tests become fiction.
Tools
Spreadsheet plus the chart you already use is enough to start. Pine or a Python stack can wait until the rule is stable. Fancy engines will not fix a rule you keep editing.
This sits next to the plan. The plan says what you may trade. The test says whether that sentence had any history worth a live bid. Neither one is advice to place a trade today.