Backtesting: How to Test a Trading Strategy and the Pitfalls That Fake the Results

Backtesting is a simulation: you take a trading strategy with fixed rules, apply it to past prices bar by bar, and record every trade it would have made. The result is what the strategy would have earned, lost and risked over that period, if you had followed it exactly. A backtest does not predict the future; it tells you whether a set of rules had any edge in the past, and how painful it was to follow.

How to backtest a strategy, step by step

1. Write the rules down before you look at results. Entry signal, exit signal, stop, position size, and which market and timeframe. If a rule needs judgment, it cannot be backtested; turn it into a test you can compute.
2. Get clean data. Prices adjusted for splits and, for a total return, dividends; enough history to cover rising, falling and flat markets.
3. Split the history. Keep a recent part aside, the out-of-sample period, and do not look at it while you build the rules.
4. Replay bar by bar, with only the information available at each bar. A signal computed on a close can only be traded after that close, at the next open at the earliest.
5. Include costs. Commissions, the bid and ask spread, and slippage, on every entry and every exit.
6. Measure more than the return. Number of trades, share of winners, largest drawdown, time in the market, and the result of simply buying and holding over the same period.
7. Run the frozen rules once on the out-of-sample period. That result, not the in-sample one, is your estimate.

You can do this by hand on a chart, in a spreadsheet, or in software; the tools are covered on our strategy tester page, and the other tools of chart reading in our technical analysis section.

A worked example: moving average crossovers on the S&P 500 ETF

The rule: hold the ETF when its fast simple moving average of closes is above its slow one, stay out otherwise, as in a moving average crossover. A signal on a close is traded at the next open; costs are 0.05% per trade, each way. Data: the daily S&P 500 ETF, dividend-adjusted prices. When the fast average is already above the slow one on the first day of a period, as it was for the 5 and 20 day pair in both periods, the test buys at the second open of the period; a position still open at the end is valued at the last close.

In the in-sample period we tried 17 pairs: fast averages of 5, 10, 20 and 50 days with slow averages of 20, 50, 100, 150 and 200 days, the fast always shorter. The best was 5 and 20 days. We then ran the same pair, unchanged, on the out-of-sample period.

The pair that looked best in 2022 and 2023 was among the worst afterwards, and the pair that ranked first afterwards, 5 and 150 days, had come only 10th of 17 in the in-sample period. The ranking built on 2022 and 2023 did not carry over to the next period: that is what overfitting looks like, and only a period kept aside can show it. The example is an illustration of the method, not a strategy we recommend.

Five pitfalls that make a backtest look better than it is

1. Overfitting

Trying many variants and keeping the best one finds the variant that fitted the noise of that period. Every parameter you tune adds to the problem. The defenses: fewer parameters, an out-of-sample period, and suspicion of any result that only works for one exact setting.

2. Lookahead bias

Using information that was not available at the time of the trade. The classic case is acting on a signal at a price that came before the signal was known. A signal computed on a bar's close can be traded at the next bar's open at the earliest; even a fill at the signal bar's own close assumes you acted on a price that was only final once the bar was over.

In our example, buying and selling at the open of the bar whose close gives the signal, which no trader could do, turns the out-of-sample result of the 5 and 20 pair from +28.5% to +37.9% with the same costs. Swing points, which are only confirmed several bars after they form, are another trap: a rule that uses a swing low on the day it formed is looking into the future, as our page on the Aroon indicator and ZigZag shows with dated examples.

3. Ignoring costs

A strategy that trades often pays the spread and commissions every time. Without costs, the 5 and 20 pair made +31.3% out-of-sample instead of +28.5%; with fewer trades the gap is smaller, with daily trading it can erase the whole edge.

4. Survivorship bias

Testing a stock strategy on today's index members leaves out the companies that were removed, went bankrupt or were bought, and the losers they would have produced. Use data that includes the delisted names, or test on instruments that existed through the whole period, such as an index ETF.

5. Too few trades

With 14 trades over two years, as in our in-sample period, a handful of trades decides the result: too few to tell skill from luck. More markets, more history or a faster timeframe give a larger sample, and the conclusions should wait for it.

Backtesting, forward testing and paper trading

A backtest runs the rules on the past. Forward testing runs the same frozen rules on new data as it arrives, without changing them, either on paper or with a small position. It is slower, but it cannot be overfitted to data you have not seen yet, and it tests something the backtest cannot: whether you actually follow the rules. A trailing stop or a time exit, written in advance, is easier to follow in both. Simulated accounts for this step are compared on our page of paper trading apps.

Paper trading keeps one source of the backtest's optimism: a simulated order often fills in full at the quoted price, while a live order that is large for the market uses up the best price and fills partly at worse ones.

How we test chart patterns

The figure pages of this site, candlestick patterns included, are backtested with the same principles: a fixed detection rule for each pattern, only the data available at each bar's close, with every measurement starting on the next bar, assumed costs per market, and three comparisons: random bars of the same asset, the excess return over the rest of the market, and the same trade on random dates of the same asset in the same conditions. The whole pipeline is also run on random prices, where no pattern can have an edge, to check that it does not invent one. The method is described in how we test patterns.

Backtesting software

The best backtesting software depends on your market and on whether you write code; the detailed comparison is on our strategy tester page. The platforms most often used for it:

Every tool runs a rule with lookahead bias or an overfitted setting as readily as a sound one: the checks above are yours to make.

Learning to backtest

New Trader University sells an online course on the subject, Backtesting 101, which walks through backtests of moving average strategies. We have not taken it and do not rate it.