Does Backtesting Only Coins Still Trading Today Overstate Returns?

Trading Strategies

26 October 2026

Yes, and the overstatement can be large. Backtesting only coins that still trade today means you are systematically deleting from your sample the coins that went to zero, got delisted, or lost liquidity entirely. Those "vanished coins" are exactly the part of a backtest where a strategy should be forced to experience losses. Remove them, and the equity curve looks smoother, drawdowns look shallower, and the Sharpe ratio looks higher. But that result reflects how "survivors" performed, not the full reality of what was actually tradable at the time.

OKX Exchange
A leading global cryptocurrency platform,suitable for both beginners and experienced traders.
New user benefit: 20% off trading fees upon registration!！

Survivorship bias is especially severe in crypto

Traditional stock markets also have delistings, but the failure rate in the crypto market is on a completely different scale. As of September 2026, CoinPaprika tracked 61,463 coins, of which only 13,154 were still active. That means 78.6% of coins are no longer trading. If you use "trading pairs that still exist on Binance/OKX today" to backtest a strategy for 2021 or 2022, your sample pool has already automatically filtered out nearly 80% of the historical participants. And most of those filtered-out coins did not disappear because they performed too well.

A reproducible example comes from the Chinese quant community: a momentum strategy showed an annualized return of 35% when backtested on "current mainstream coins." After adding back coins that were once listed but later delisted or collapsed, the annualized return dropped to 22%. That is a reduction of more than one third. This is not a minor adjustment; it is a change in magnitude.

What gets deleted, and why it matters

You are not deleting a random sample. You are deleting failed samples. LUNA, FTT, and many tokens that were active in 2021 but later went to zero or were delisted cannot be found in today's trading pair list. If your strategy had held LUNA in 2021, the real outcome might have been close to a total loss. But if the backtest sample only includes coins that are still alive today, that loss never happened. The strategy's "maximum drawdown" is therefore systematically underestimated, and its "win rate" is systematically overestimated.

Academic research points in the same direction. One study covering 378 million Monte Carlo simulations found that after correcting for survivorship bias, the median excess return of holding 378 non-stablecoin crypto assets over a 2-3 year window was -28.4%. Only the top quartile earned extremely high returns. In other words, the true return for most assets is negative, but if you only look at coins still trading today, your sample is naturally tilted toward the few winners.

How to handle this in backtesting

The most direct approach is to use the list of trading pairs that existed at each historical point in time, not today's list. Data providers such as CoinAPI offer interfaces for querying historical symbols, including trading pairs that have been delisted or become inactive. You can use these to rebuild the asset pool that was "actually tradable" on each backtest date. This is called a point-in-time universe: on date t, include only assets that actually existed and had data at that time, including those that later disappeared.

For individual backtesters without a paid data source, a practical compromise is to at least manually add back the well-known coins you know were listed during the backtest period and later went to zero or were delisted. It does not need to be complete, but it will show you how the strategy behaves on those failed assets. If returns drop sharply after adding them back, the previous backtest result was largely supported by survivorship bias.

Another point to keep in mind: survivorship bias cannot be fixed by walk-forward testing. Walk-forward can reduce overfitting, but every window still uses the same group of assets that "survived to today," so the bias exists in every window.

How to tell if your backtest is contaminated

When looking at a result, first ask: where did the sample pool come from? If the answer is "filtered from Binance's current trading pair list" or "selected from CoinGecko's current top 100," survivorship bias is almost certainly present. If the backtest period covers the collapse of 2021 or 2022, and the sample contains none of the coins that disappeared during those crashes, the return numbers have low credibility.

A rough check: does your backtest sample contain any coin that went to zero, was delisted, or stopped trading after the backtest period? If none at all, your sample is too "clean" — clean to the point of being unrealistic. In the real crypto market, failure is the norm, not the exception.