A rule set that returned 60 percent on historical data and loses money in its first month live is not necessarily a bad strategy. More often it is a strategy that was tested against a version of the market that never existed.
Almost every gap between backtest and live performance traces back to one of six things, and five of them are data problems rather than trading problems. Here is how to find which one is affecting you.
1. Fees and slippage were left out or guessed
This is the first thing to rule out because it is the easiest to fix. A strategy trading twenty times a month at 0.1 percent per side gives up 4 percent a year before anything else happens. Add maker and taker differences, and a strategy that assumed maker fills while actually taking liquidity can lose its entire edge.
Slippage is the harder half. Backtests fill at the price on the candle. Live orders fill against a book that moves, and the difference grows with position size and shrinks with liquidity. Test the same rules at three position sizes. If the returns fall away sharply as size increases, the edge lives in a part of the book you cannot reach with real money.
2. The candle told you something it could not have known
Look ahead bias is the classic silent killer. It appears whenever a rule uses information that was not available at the moment the trade would have been placed. Using the closing price of a candle to enter within that same candle is the most common version, and it flatters results dramatically on volatile pairs.
The test is simple. Shift every entry one candle later and rerun. A robust strategy loses some performance. A strategy built on look ahead bias collapses entirely, and better to learn that in a backtest than in a drawdown.
3. The historical data has gaps you never looked for
Exchange history is not clean. Pairs get delisted, symbols get renamed, exchanges go down during exactly the volatile periods your strategy claims to profit from, and some datasets silently interpolate the missing candles rather than leaving a hole.
Interpolated data is the worst case because it makes the market look smoother than it was. A strategy tested on it learns that price moves in orderly steps, which is precisely wrong about the moments that matter. Before trusting any dataset, plot the count of candles per day for a year and look for the dips. They will be there, and they usually sit on the days you would most want to be right about.
4. Survivorship bias in the pair selection
Testing a strategy across the current top fifty coins by market cap tells you how it would have performed on assets that survived to become the top fifty. Every token that collapsed on the way is missing from the sample.
In crypto this effect is much larger than in equities, because the failure rate is higher and the listing history is shorter. If your backtest universe was chosen using today’s rankings, treat the results as an upper bound rather than an expectation, and rerun on the universe as it stood at the start of the test period.
5. Your live data feed is not the feed you tested on
Backtests usually run on clean historical exports. Live rules run on whatever your connection delivers in real time, and those are different animals. Rate limits are the usual point of failure. Major exchanges limit by API key and separately by IP address, so a strategy polling several pairs across several exchanges can hit a limit that has nothing to do with its trading volume.
When that happens you do not get an error you notice. You get a delayed price, and the rule fires late. The fix is partly architectural, moving from repeated REST polling to websocket streams wherever the exchange offers them, and partly about not sending all requests from one place.
Traders who collect wider datasets, funding rates across venues, order book snapshots, news and social sentiment, run into the same wall faster, because public endpoints and web sources are stricter than authenticated trading endpoints. Distributing that collection is routine work, and an affordable residential proxy plan from a provider such as ProxyWing costs less per month than a single bad fill. What matters is keeping the collection layer separate from the execution layer, so a data job that hits a limit never delays an order.
One boundary is worth stating plainly. This applies to gathering public market data. Using an address in another country to reach an exchange that does not serve your jurisdiction breaches the terms you accepted and in several countries breaks the law outright. It also tends to end with a frozen withdrawal at the worst possible moment.
6. The market regime changed and the rules did not
A strategy tuned on 2021 conditions is tuned on a market with different volatility, different correlation to equities and different participants. Nothing about the code breaks, it simply stops describing the market it is trading.
Splitting the test period is the standard defence. Optimise on the first two thirds, validate on the final third, and never touch the validation period again once you have looked at it. If performance in the untouched period is materially worse, the parameters were fitted to noise.
What good validation looks like in practice
| Check | What it catches | Effort |
| Shift entries one candle later | Look ahead bias | Minutes |
| Rerun at three position sizes | Slippage sensitivity | Minutes |
| Plot candles per day across the test period | Data gaps and interpolation | Under an hour |
| Rebuild the universe as of the start date | Survivorship bias | A few hours |
| Walk forward on an untouched final third | Overfitting to parameters | A few hours |
| Paper trade for a full month | Feed, latency and rate limit problems | One month |
The last row is the one people skip and the one that catches the most. A month of paper trading with the real feed, the real polling schedule and the real rules costs nothing but patience, and it surfaces every infrastructure problem before money is involved.
A reasonable order of operations
Fix fees and slippage assumptions first, because they are cheap to correct and they eliminate a good share of false positives. Then test for look ahead bias, which either invalidates the strategy immediately or clears it. Then examine the data itself. Only after all three should you spend time tuning parameters, since tuning on flawed inputs just produces a more confident version of the same mistake.
Strategies that survive this sequence tend to look duller than the ones that do not. That is the point. A rule set returning a modest number consistently across regimes is worth considerably more than one returning a spectacular number in a backtest nobody stress tested.
FAQ
How much historical data do I need for a reliable backtest?
Enough to cover at least two different market regimes, which in crypto usually means a full cycle. Six months of data during a single trend tells you how the strategy behaves in that trend and nothing more.
Is paper trading actually useful, or just a delay?
It is the only way to test your infrastructure end to end. It will not validate the edge, because fills are simulated, but it reliably exposes feed delays, rate limits and rule logic that behaves differently on live data.
Why do the same rules perform differently on two exchanges?
Fee structures, liquidity and the exact timestamping of candles all differ. Order books on smaller venues also move more against larger orders, which shows up as slippage rather than as a strategy problem.
Should I optimise parameters continuously?
Reoptimising frequently is usually a slow way to overfit. Set a fixed review interval, use a held out period each time, and require a meaningful improvement before changing anything.
What is the single most common cause of live underperformance?
Transaction costs that were understated in testing. It is unglamorous and it accounts for more failed strategies than every other item on this list combined.
Disclaimer
This article is for informational and educational purposes only and does not constitute financial, investment, legal or tax advice. Crypto trading involves substantial risk, and past, simulated or backtested performance does not guarantee future results.
Coinrule provides trading automation software and does not guarantee the performance of any strategy, exchange, data source or third-party service. Users are responsible for their own trading decisions and for complying with applicable laws and platform terms. Coinrule’s services are not intended for, directed at or marketed to residents of the EU or EEA, and Coinrule does not hold authorisation as a Crypto-Asset Service Provider under MiCA.
This article contains sponsored content and third-party links. Coinrule does not endorse or guarantee third-party products or services.