Backtesting is the process of applying a trading strategy to historical data to see how it would have performed. Done right, it tells you whether your strategy has a statistical edge. Done wrong — which is most of the time — it tells you a convincing lie. Here's how to do it right.
Backtesting takes a set of explicit trading rules (when to enter, when to exit, how much to trade) and runs them against historical price data. The output is a simulated equity curve showing what your account would have done if you'd traded those rules in the past.
The purpose isn't to predict the future — it's to answer the question: "Does this strategy have characteristics consistent with a real edge, or is it noise?"
As the algorithmic trading literature emphasizes, a backtest is the first filter in a pipeline: hypothesis → backtest → paper trade → small live → full live. Most strategies die at the backtest stage, which is exactly the point — you want them to die cheaply on historical data rather than expensively with real capital.
Step 1: Define explicit rules. Every decision must be a rule — no discretion. Entry condition, exit condition, position size, stop-loss, and any filters. If you can't write it as pseudocode, you can't backtest it.
Step 2: Get clean data. Historical OHLCV (open-high-low-close-volume) data for your target market. For crypto, this means exchange-specific data (Binance BTC/USDT perpetual is different from Bybit BTC/USDT perpetual). Data quality matters — missing candles, incorrect timestamps, or adjusted prices will corrupt your results.
Step 3: Run the simulation. Apply your rules to the data chronologically. For each bar, check entry conditions, manage open positions, check exit conditions. Track every trade: entry price, exit price, position size, P&L.
Step 4: Measure results. Calculate performance metrics (below). Compare to benchmarks — both "buy and hold" and a random baseline.
Step 5: Validate. Apply the strategy to out-of-sample data it wasn't designed on. If performance degrades dramatically, you've overfit.
| Metric | What It Tells You | Good Target |
|---|---|---|
| Total return | How much the strategy made | Context-dependent |
| Sharpe ratio | Return per unit of risk | >1.5 |
| Max drawdown | Worst peak-to-trough decline | <25% |
| Win rate | % of trades profitable | 40-60% for trend, 55-70% for MR |
| Profit factor | Gross profit / gross loss | >1.5 |
| Number of trades | Statistical significance | >100 minimum |
| Avg win / avg loss | Reward-to-risk ratio | >1.5 for trend following |
A critical rule: if you have fewer than 100 trades, your results are statistically meaningless. You need enough trades to distinguish indicator from noise.
Overfitting is the #1 cause of backtest-to-live failure. It happens when your strategy is tuned to historical quirks rather than genuine market patterns.
Signs of overfitting:
López de Prado's financial ML framework provides rigorous tools for this: purged k-fold cross-validation ensures that your test data doesn't leak into your training data through overlapping time windows. The purging removes training observations whose label intervals overlap with test labels, and an embargo window removes observations immediately after the test set.
Survivorship bias. If you backtest on "top 50 altcoins by market cap," you're only testing on coins that survived. The ones that went to zero aren't in your dataset. This inflates returns.
Regime shifts. Crypto market structure has changed dramatically — 2017 was a different market than 2021, which was different from 2024. A strategy that worked in 2020's DeFi summer may be irrelevant in 2025. Test across multiple regimes.
Funding and liquidation costs. Most crypto backtesting tools don't account for funding rate costs or liquidation risk. A strategy that shows 50% annual return but holds leveraged positions through dozens of funding intervals may actually be unprofitable after costs.
Exchange-specific execution. Slippage, fees, and fill quality vary between exchanges. A backtest that assumes instant fills at the candle close price overstates performance. Add realistic slippage assumptions: 0.05–0.10% per trade for liquid BTC pairs, 0.2–0.5% for altcoins.
It's your first line of defense against bad ideas. Most trading ideas don't work. Backtesting lets you kill them cheaply. If a strategy can't even work on historical data (which it has the advantage of being designed for), it certainly won't work live.
It quantifies your edge. Instead of "I think this works," you can say "over 500 trades across 3 years, this strategy has a 1.8 Sharpe ratio, 55% win rate, and 18% max drawdown." That's a basis for capital allocation. A feeling is not.
It reveals hidden risks. A strategy may show great annual returns but with a 45% max drawdown. Would you hold through a 45% drawdown? If not, you won't capture the returns. Backtesting surfaces the pain points you'll face in live trading.
Look-ahead bias. Using information that wasn't available at the time of the trade. Example: using the day's close to make a decision at the day's open. In crypto, this is especially common when using "daily close" indicators — whose daily close? UTC? Exchange-specific? The timezone matters.
Ignoring transaction costs. A strategy that trades 20 times per day with 0.02% edge per trade sounds great — until you add 0.04% in fees per trade (maker + taker), which turns the edge negative. Always include realistic fees, spread, and slippage.
Optimizing on the full dataset. If you use all available data to tune parameters, you have no out-of-sample period to validate. Split your data: 60% for development (in-sample), 20% for validation, 20% for final test. Touch the final 20% only once.
Python with pandas, numpy, and a backtesting library (backtrader, vectorbt, or custom) is the professional standard. TradingView's Pine Script is accessible for simpler strategies but limited for complex logic. For serious work, build your own framework — commercial tools hide assumptions that can distort results.
Enough to generate 100+ trades across at least 2 distinct market regimes (bull + bear minimum). For daily strategies, 3-5 years is typical. For hourly strategies, 1-2 years may suffice. More data is better, but only if the market structure during the early data is still relevant.
Almost certainly not at face value. Returns that high usually indicate one or more of: overfitting, missing transaction costs, survivorship bias, or look-ahead bias. Stress-test aggressively: add 2× your expected slippage, test on out-of-sample data, vary parameters by ±20%, and check if returns survive. If they drop to 30-50% after realistic adjustments, you may have something worth paper trading.
---
*This article is part of The Codex — PARAGON's structured learning library.*