What Is Backtesting? What It Proves and What It Doesn't
Backtesting runs a set of trading rules against historical market data to see how they would have behaved. You write the rules, replay the past, and count what the rules would have done. The output is a table: number of trades, win rate, return, largest decline.
Most guides stop there and tell you how to run one. This one assumes something different about you.
Most people never run a backtest. They are shown one. A bot vendor posts an equity curve. A signal channel publishes a performance table. A platform puts a number on its homepage. The skill that matters is not running the test. It is reading the one in front of you.
So this covers both: what a backtest genuinely proves, and the seven questions that tell you whether the one you are looking at means anything.
How backtesting works
Three parts.
Rules. When to enter, when to exit, how large a position, when to stop. If the rules are vague there is nothing to test. "Buy when the chart looks strong" is not a rule.
Data. Price history for a specific asset over a specific period. Usually candles: open, high, low, close.
Simulation. Software replays the data and marks every moment the rules would have fired. The result is a trade list, which gets summarised into statistics.
That summary is the most misread document in retail trading.
What backtesting actually proves
Your rules are unambiguous. If a strategy can be backtested at all, you have written it precisely enough for a machine to follow. That alone has value. An instinct cannot be tested.
Whether the logic had an edge in the past. If the rules lose money systematically across historical data, the logic is probably broken. A backtest establishes that cheaply, before any capital is at risk.
The shape of the risk. How deep the worst decline was, how long it lasted, how many losses came in a row. These numbers describe what holding the strategy would have felt like far better than the return does. We cover this in more depth in our guide to maximum drawdown.
Notice that all three are statements about the past. None of them is a statement about the future.
The five ways a backtest lies
This is the section most guides compress into one sentence. It is where the actual information lives.
Overfitting
The most common failure. You tune a strategy's settings so precisely to historical data that the strategy stops learning the market and starts memorising that period's noise.
It happens like this: a 14-period setting performs poorly, so you try 17, then 21, and at 19 the equity curve looks excellent. That 19 is not a discovery. It is a coincidence, and it does not survive contact with live data.
The tell is fragility. Nudge the parameter slightly and if the result collapses, the logic is not robust. Sound logic performs reasonably across neighbouring settings, not at exactly one.
Look-ahead bias
The simulation uses information that would not have been available at that moment. The classic version: deciding based on a candle's closing price, then entering a trade inside that same candle. In reality you only know the close once the candle has closed.
The error looks tiny and inflates results enormously.
Survivorship bias
Testing on assets that exist today. Dozens of tokens listed in 2021 and went to zero by 2023, and they are not in the dataset. The strategy has been tested exclusively on survivors, which is to say on winners.
In crypto this effect is far larger than in equities, because the delisting and abandonment rate is higher.
Ignoring costs
Fees, slippage, and liquidity. A backtest that assumes every order fills at the price on screen is not describing anything that can happen.
The impact compounds with frequency. A strategy placing several trades a day is affected many times more than one placing several trades a month. A high-frequency backtest can invert completely once realistic costs are applied.
Testing one market regime
A test covering 2020 to 2021 makes almost any long-biased strategy look excellent, because that period trended up. Run the same rules through 2022 and the picture changes.
A meaningful test spans an uptrend, a downtrend, and a range.
AlgoTitan strategies run for 14 days on live market data in paper mode. No exchange connection, no payment, no card. Paper trading uses demo funds. This is not investment advice. See the strategies
How to read a backtest someone shows you
Seven questions. If most of them go unanswered, you are holding a marketing asset rather than a measurement.
1. What period does the test cover? If it is only an uptrend, it tells you very little. Look for at least one full cycle including a sustained decline.
2. Are fees and slippage included? If not, the result is a theoretical ceiling rather than an outcome. This is usually not stated. If it is not stated, assume they are excluded.
3. How many trades? Twenty trades is not a sample. It is anecdote with a chart attached. Hundreds of trades produce something you can actually read.
4. Is maximum drawdown shown? If only return is displayed, half the picture is missing. Presenting return without risk is a deliberate choice, and it is worth noticing who made it.
5. When were the parameters chosen? If the settings were picked after looking at results from the test period, that is not a backtest. It is a fit. The honest method selects parameters on one period and tests on another it has never seen.
6. Was the same logic tried on another asset or another period? Logic that works on exactly one asset over exactly one window has probably been shaped to fit that window.
7. Is there any forward test? Historical testing and real-time testing answer different questions. More on that below.
Several of these overlap with the broader warning signs covered in how to spot a trading bot scam.
Backtesting versus forward testing
These are not two grades of the same exam. They answer different questions.
Backtesting uses historical data. It is fast: a decade of data runs in seconds. But it is an exam whose answers are already known. There is always the possibility, conscious or not, of adjusting the rules after seeing how they scored.
Forward testing, also called paper trading or demo trading, runs in real time. The strategy trades current market conditions with simulated funds. It is slow: fourteen days of evidence takes fourteen days. But nobody knows the outcome in advance, which makes fitting impossible.
That makes forward testing the harder exam, not the weaker one.
In practice the two run in sequence. Backtesting eliminates logic that never worked, cheaply. Forward testing shows how the surviving logic behaves under real conditions. Our paper trading guide covers that second stage in detail.
There is a third difference that marketing material never mentions: forward testing also tests you. Whether you stay with a strategy through three days without a single trade, whether you switch it off during a drawdown, whether a quiet week reads as broken to you. A backtest cannot measure any of that, and it is frequently the thing that decides the outcome. We wrote about the quiet-period problem separately in why your trading bot is not trading.
Frequently asked questions
How reliable is backtesting? If the rules are precise and realistic costs are included, it reasonably indicates whether logic had an edge historically. It offers no guarantee about future results. Past performance does not guarantee future performance.
Is backtesting the same as paper trading? No. Backtesting uses historical data; paper trading uses live real-time data. Paper trading is slower but cannot be fitted after the fact.
How many trades should a good backtest include? There is no fixed threshold, but a few dozen trades can be explained by chance. Hundreds produce a more readable picture.
Why do backtest results fail to hold up live? Most often three reasons: overfitting, omitted costs, and testing across a single market regime.
Is backtesting enough for crypto strategies? Not on its own. Liquidity and volatility shift quickly in crypto, which makes real-time forward testing a separate and necessary step.
What is look-ahead bias? Using information in a simulation that would not have been available at that point in time, such as acting on a candle's close before the candle has closed.
The short version
A backtest is an elimination tool, not evidence. It cheaply demonstrates that logic does not work. It cannot demonstrate that it will.
When a performance table is put in front of you, the question is not "what was the return." It is "how was this number produced." Test period, costs, trade count, drawdown, and when the parameters were chosen. If those five are not disclosed, you are looking at a presentation rather than a measurement.
AlgoTitan strategies run for 14 days on live market data in paper mode, before any exchange connection or payment. Paper results are simulated and may differ from live results due to fees, slippage, and liquidity. Trading involves risk of loss, including loss of principal. This is not investment advice.
