Backtesting Guides
These guides exist because most backtests lie — not through bad math, but through look-ahead bias, overfitting, ignored costs and reckless sizing. Each article tackles one way a promising strategy falls apart, using general principles you can verify yourself on this site's backtester.
New to backtesting? Start with look-ahead bias and overfitting — they explain why results look too good. Then read the indicator guides (RSI, Bollinger, Supertrend) to build entries, and finish with position sizing and max drawdown to decide how much to risk.
Look-Ahead Bias: What It Is Worth, Measured by Injecting It
We rewrote the candle data so the engine filled at prices it could not have known. All three test strategies flipped from losing to winning, and the drawdown fell from 92.3% to 26.3%.
Overfitting in Backtests: How to Tell a Plateau From a Lucky Peak
A single peak in a parameter sweep is usually luck. Measured on an RSI grid: how neighbouring settings reveal a plateau, and how the sample quietly disappears as you tune.
Do Fees and Slippage Really Matter? We Ran 8 Strategies Twice
Eight strategies run at zero cost and again at 0.06% fee plus 0.02% slippage per side. Four were profitable without costs, and two of those four flipped to a loss with costs on.
Does an RSI Strategy Actually Work? A Measured Backtest
Mean reversion against momentum on seven years of Binance data. The momentum version has the higher win rate and loses more, and the exit rule changed the result more than the entry did.
Does the Supertrend Strategy Work? Tested Across 6 Coins
The ATR multiplier decides almost everything and the period barely matters. Taking only longs returned +1,361.6% against +462.6% both ways, and only two of six coins were profitable.
Win Rate vs Risk/Reward: 36 Tests, and Win Rate Predicted Nothing
One entry signal, 36 take-profit and stop combinations. Thirteen cells won more than half their trades and not one of them made money; an 8% stop won 73.2% and still lost 46.2%.
Compounding and Volatility Drag: When a Positive Average Still Loses
Two of eight strategies have a positive average trade and a losing account. Shuffling the trade order 200 times never changed the final return, only the drawdown, which ranged 37% to 90%.
Does the Bollinger Band Strategy Work? 16 Settings, None Profitable
Fifteen of sixteen parameter cells traded and all fifteen lost. Buying the lower band and buying the upper band both won exactly 40.1%, so at that setting the band supplied no direction.
Position Sizing and Leverage: 3x Turned +221% Into -58%
Leverage measured on a strategy that actually works: 2x halved the return and 3x turned +221.3% into -58.5%. Not one position was liquidated along the way, which was the surprise.
Does the MACD Strategy Work? One Parameter Flips the Sign
Moving only the signal period from 5 to 16 takes the same rule from -94.6% to +54.5%, and changing only the timeframe moves it from -100.0% to +932.8%.
How Many Trades Does a Backtest Need? We Resampled 2,126
A strategy that loses 98.1% showed a profit in 35.9% of random ten-trade samples. Its best thirty-trade stretch returned +43.1% with a 70% win rate, and every number in it is real.
Market Regimes: Trend Following Lost to Buy and Hold Every Up Year
Split by calendar year against BTC's own return. Trend following was profitable in all five up years and lost to simply holding in every one of them, which is not how it is usually sold.
Max Drawdown: Every Strategy Spent 93%+ of Its Life Below a Peak
Every preset spent 93.4% to 99.8% of its bars below a previous peak. A one-year test reported a 43.6% drawdown where the full history reported 99.1%, so the figure is a floor, not a ceiling.