Why Most Backtests Are Lies
2026-02-13
The Fantasy vs Reality Gap
Every crypto bot on Twitter shows a backtest with insane returns. +500%. +2,000%. “Proven profitable.”
Then you run it live and lose 30% in a week.
This isn’t bad luck. The backtest was a lie from the start.
The Three Lies of Crypto Backtests
1. Look-Ahead Bias
The most common and most dangerous.
Backtest thinks: "The candle closed bearish, volume spiked → SHORT"
Reality at 10:01: "The candle is still forming. I can't see the close."If your backtest uses the current candle’s data to make decisions, it’s using information that doesn’t exist yet in live trading. This alone can turn a losing strategy into a “winning” one.
PRUVIQ’s rule: Only use the previous completed candle for signals. The current candle is only used for entry price.
2. Overfitting
You test 50 parameter combinations and pick the best one. Of course it looks good — you optimized it to fit historical data perfectly.
But markets change. What worked in 2024 Q1 might fail in 2024 Q3.
How to detect it:
- Does it work across multiple time periods? (2023, 2024, 2025)
- Does it work on out-of-sample data?
- Is the win rate suspiciously high? (>70% in crypto = suspicious)
3. Ignoring Real Trading Costs
Backtests assume perfect fills at exact prices. Live trading has:
- Slippage — you don’t get the price you want
- Funding rates — holding futures costs money
- Partial fills — your order might not fully execute
- Latency — signals arrive faster than execution
A strategy that backtests at +50% might be +5% or negative after costs.
Our $14,000 Lesson
We learned this the hard way. A momentum strategy backtested at +400%. We ran it live. It lost $14,115.
The problem? A one-candle indexing error. The backtest was using current instead of previous candle data. One line of code. Fourteen thousand dollars.
How to Verify a Backtest
Ask these questions:
| Question | Red Flag |
|---|---|
| Which candle data is used for signals? | Current candle = look-ahead bias |
| How many parameter combinations were tested? | Many = likely overfit |
| Is out-of-sample testing shown? | No = probably only works on training data |
| Are trading costs included? | No = unrealistic returns |
| Does the backtest code match live code? | Different = apples to oranges |
The Rule
If you can’t prove a strategy works on data it has never seen, don’t trade it with real money.
This is the foundation of everything PRUVIQ builds. Every strategy goes through out-of-sample validation before it touches a live exchange.
This is educational content. Not financial advice.