How to Backtest a Trading Strategy

Step by step

1. Write the rules down

List what opens a trade, what closes it, how big it is, and what happens when two things clash. "Buy when it looks strong" cannot be tested. "Buy at the next open after the 14 bar RSI closes below 30; sell when it closes above 50 or after 10 bars" can. If you change a rule after seeing the result, you have started a new test, and you should count it as one.

2. Get clean data

You need the same bar size the bot will trade, with open, high, low, close and volume, covering more than one type of market. Check for gaps, bad ticks, and splits or contract rolls that make a price jump. Data can also differ from what you will trade live, since the quote you see from a broker is not the quote from another source.

3. Decide the fill rules before you run

Say at which price each order fills. A signal on a bar's close can only be traded at the next price, not the close itself. A limit order that the price only touched may not have filled in real life. A stop can fill worse than its level. Choose the pessimistic version when you are unsure.

4. Add every cost

Include the spread, commissions or fees, slippage and any overnight or funding charge. The table below shows what a small fixed cost does to a strategy with many small trades. The costs vary by broker and venue, so use the numbers from your own account.

5. Run it and keep the trade list

Save every trade with its entry, exit and result, not only the total. You will need the list to compute the rest and to see which trades carry the result.

6. Read the result table

Compute the number of trades, the win rate, the average win and loss, expectancy, profit factor and the max drawdown. A total with 8 trades behind it is a story, not a measurement.

7. Test it again on data you kept back

Split the data before you start. Build and tune on one part. Test once on the part you hid. If the hidden part is much worse, the tuning fit the noise.

A worked example on 12 made-up trades

A bot backtest produces the trade list below. Each result is a percent of the traded amount, and every figure is made up. The cost is 0.15 percent per trade, which stands in for the spread, the fee and slippage together. Everything was calculated by us on 29 Oct 2026 and is exact.

The result table for the same list reads like this:

• Trades: 12. This is too few to trust, and is used only to show the arithmetic.
• Win rate: 6 winners of 12 before costs (50.0 percent). After costs, 6 (50.0 percent), because the cost turns no winner into a loser here.
• Average win and loss, after costs: +0.93 and -0.95 percent.
• Expectancy per trade: +0.14 percent before costs, -0.01 after. Total +1.70 before and -0.10 after.
• Profit factor, after costs: 0.98. It is the sum of the winners divided by the sum of the losers.
• Worst fall from a peak: 2.50 points, between the highest running total and the lowest point after it.

Look at how much the cost takes. It is the same 0.15 on every trade, and it removes 1.80 of +1.70. A bot that trades twice as often with the same edge per trade loses twice as much to cost. This is why a backtest with no costs is a common way to get a result that disappears live.

The traps, one by one

Look-ahead

The test uses information that was not available at the time of the trade. A common form is trading at a bar's close using an indicator that needs that same close. Another is using a daily high or low to decide an entry earlier the same day. The check: for each signal, ask what was known at that moment. See look-ahead bias.

Survivorship

The data only holds the instruments that still exist, so the ones that failed or were delisted are missing. A stock test run on today's index members looks better than one run on the members of that day. See survivorship bias.

Costs left out

See the example above. Add spread, fees and slippage to every trade, and test with higher costs as well. If the result vanishes when costs rise a little, it was thin.

Overfitting

Trying many settings and keeping the best one finds noise. As a made-up example, suppose a test of 12 settings gives results from -0.60 to +0.90, an average of +0.03. Reporting only the best, +0.90, hides the other eleven. The more settings and rules you try, the more likely the top one is luck. The fix is the data you kept back, simple rules, and settings that still work when moved a little. See overfitting.

What a backtest does not tell you

It does not tell you the future, how the market will treat your orders, or how you will behave in a losing run. A backtest also cannot show events that were not in the data. For a martingale or grid bot, whose worst case is rare, a clean backtest can hide a loss that ends the account. Run the result through the Monte Carlo simulator to see how the order of trades changes the fall.