Algorithmic Trading Backtesting: How to Test a Strategy Without Fooling Yourself

What Algorithmic Trading Backtesting Actually Tests

Algorithmic trading backtesting takes a strategy's entry and exit rules and runs them against years of historical price data. It goes bar by bar to see what would have happened. I usually start with five years of daily data. For a strategy that trades often, I add a year or two of minute-level data on top. The output isn't a prediction. It's a record of one possible past.

A backtest doesn't prove a strategy works. It shows how a set of rules would have handled one stretch of market history, under the assumptions you coded into it. Change those assumptions and the same rules can tell a very different story. That's the first thing that trips up someone building their first system.

A basic workflow looks like this. Pull clean historical data. Code the entry and exit rules exactly as you'd trade them. Run the simulation bar by bar. Then compare the result against a benchmark like buy-and-hold. Haven't nailed down your entry and exit logic yet? Our strategy development coverage walks through that step. Skip the benchmark check and a mediocre strategy can look impressive just because the market went up.

Survivorship Bias and Look-Ahead Bias: The Traps Hiding in Your Data

One trap is survivorship bias. Backtest a stock-picking strategy using today's S&P 500 list against 15 years of prices, and you're only testing companies that survived long enough to still be in the index. Enron isn't on that list. Neither are the mortgage lenders that vanished in 2008. Those losers were quietly removed first. A strategy that would have picked them looks better than it should.

Look-ahead bias is subtler. It happens when a backtest uses information that wasn't actually available on the date being tested. Say, reacting to a quarterly earnings figure on the last day of that quarter, when the real number wasn't published for three more weeks. I've caught this exact bug in a moving-average strategy that used closing prices adjusted after a stock split was announced. That let the model effectively see the split coming. It's an easy mistake to make and a hard one to notice, because the backtest just looks unusually good.

Both biases share a pattern. They make a strategy look better on paper than it would have looked in real time. Checking a dataset for these problems isn't optional. It's part of the test itself.

Fees and Slippage: The Gap Between a Backtest and a Live Account

Most beginner backtests assume every order fills instantly at the exact price the strategy asked for. Live markets don't work that way. A market order on a fast-moving stock might fill a few cents away from where you expected. Every trade costs something too, even on a broker advertising zero commissions, because the spread itself is a cost.

One of the first systems I tested, a simple moving-average crossover on five-minute bars, showed a 41% annual return with zero slippage assumed. Adding a conservative slippage estimate, just 0.05% per trade plus a flat commission, dropped that same strategy to under 9% a year.

The rules never changed. Only the assumptions did.

Slippage matters more for strategies that trade often. A system placing four trades a year can absorb a few cents of slippage without much damage. A system placing forty trades a day cannot. That difference alone explains why some high-frequency-style strategies that look brilliant on paper lose money the moment they go live.

Overfitting: Why a Flawless Backtest Should Worry You

A common assumption is that a better-fitting backtest means a better strategy. In my experience it's usually the opposite. Test enough variations of moving-average lengths, stop-loss levels, and filters against the same historical data, and you will eventually find a combination that performed extremely well purely by chance.

Mathematicians David Bailey, Jonathan Borwein, Marcos López de Prado, and Qiji Jim Zhu studied this in a 2014 paper for the Notices of the American Mathematical Society. Their finding: testing many strategy variations against the same data makes it easy to produce an impressive backtest with little real predictive power. The more configurations tried, the more likely one simply got lucky.

The systems that hold up in live trading tend to look almost boring in a backtest: solid, not spectacular. A smooth equity curve with almost no drawdown is usually a reason to dig into the assumptions, not a reason to trade it with real money.

Walk-Forward Testing: How Algorithmic Trading Backtesting Should End

A single backtest, however careful, is one experiment on one stretch of history. Walk-forward testing turns that into several smaller experiments. Split years of data into chunks. Tune the strategy on the first chunk, then test it, untouched, on the next chunk it hasn't seen. Move the window forward and repeat.

A strategy that holds up across five or six of these windows is a far stronger candidate than one judged on a single ten-year run. It's also where risk management rules earn their keep, since a strategy's real edge often comes from how losses are limited, not just how wins are picked.

Python libraries like backtrader and vectorbt make walk-forward testing far less tedious than doing it by hand. TradingView's Pine Script has its own tools for the same idea. See our trading platforms and software coverage for a closer comparison. Whatever tool you use, don't skip the last step before going live: paper trading the system for a few weeks with no historical hindsight to lean on.

FAQ

What's the difference between backtesting and paper trading?

Backtesting runs a strategy against historical data all at once. It can simulate years of trades in a few seconds. Paper trading runs the same strategy against live, real-time data in a simulated account. It shows how the system handles current conditions and any bugs in your code before real money is involved. Most solid systems go through both steps.

How much historical data do I need for a good backtest?

It depends how often the strategy trades. A weekly strategy needs several years of data to produce enough trades to judge. A strategy trading dozens of times a day can get a meaningful sample from a few months of minute-level data. As a rough starting point, aim for at least 100 trades before drawing conclusions.

What's a walk-forward test?

A walk-forward test splits historical data into sequential chunks. It tunes the strategy on one chunk, then tests it unchanged on the next chunk it hasn't seen. Repeating this across several windows shows whether a strategy's edge holds up on new data, or only worked on the exact period it was built around.

Can a backtest guarantee future performance?

No backtest can guarantee how a strategy performs going forward, since markets change and no historical period repeats exactly. A well-built backtest still has value. One that accounts for fees, slippage, and out-of-sample testing can show whether a strategy's logic holds up under realistic conditions. That's the most any backtest can honestly promise.