If Your Backtest Looks Perfect, It's Wrong.
By CryptoTraders · Strategy · 2026-09-30
Anyone can produce a backtest with a smooth equity curve and an 80% win rate. It takes about an afternoon of tweaking parameters. The smoothness is not evidence the strategy works. It is usually evidence that you fit the rules so tightly to the past that they learned the noise. A perfect backtest is a red flag, not a green light.
The ways a backtest lies
Survivorship bias. If your historical universe only contains coins that still trade today, you quietly deleted every project that went to zero. Your strategy never had to survive the losers, so its results are inflated. Test on the universe as it existed at the time, dead coins included.
Look-ahead bias. This is using information the trade could not have had at the moment of the decision: a revised data point, a candle that had not closed yet, or simply tuning on the same data you then report results from. Keep training, validation, and forward data strictly separate.
Overfitting. The more parameters and conditions you bolt on, the easier it is to draw a perfect line through past noise. A few parameters that hold up across a range of values beat one ideal setting that breaks the moment the market regime changes.
Ignoring costs. Fees, spread, and slippage are real and they compound. On perps there is one more that people forget: funding. It accrues every 8 hours, and over a multi-day hold it can move your net result meaningfully. Depending on your side and the rate, funding is sometimes a cost and sometimes a credit, but it is never zero. A perp backtest that ignores it is fiction.
Too few trades. A handful of trades from one market regime proves nothing. You want enough trades across enough different conditions for the result to mean something. There is no magic minimum number to cite. The real test is whether it held up out of sample, not whether it crossed some round count.
What honest looks like
Out-of-sample testing and walk-forward analysis. Optimise on one window, then test on the next window the strategy has never seen. Roll the window forward and repeat. Model realistic costs, including funding. Use point-in-time data so you are not accidentally feeding it the future. The goal is not the best-looking curve. It is the curve that survives unseen data.
Where the Algo Alerts Dashboard fits
The strongest antidote to a flattering backtest is a live, forward record, because forward performance is the one dataset that cannot be curve-fit. That is what the Algo Alerts Dashboard provides for our algos: every signal tracked against real execution as it happens, outcomes scored in R with the stop-to-breakeven management applied mechanically, and rolling statistics that update as trades close, losers included. A backtest is a hypothesis. The dashboard is the experiment, running in public.