Why Does a Trading Strategy Work in Backtests but Fail in Live Trading?

A backtest is a story about the past. It can be a good story and still have no edge in it. Here is why so many beautiful equity curves die in the first month of live trading.

What people on X are asking

In October 2026 one account asked why a profitable backtest can lose money in live trading and listed three things to check: overfitting, trading costs and changing markets. In January 2026 an algo trading account claimed that 80% of algo traders fail in their first year because their perfect equity curves fall apart live; we found no study behind that 80% figure. And in March 2025 a systematic trader argued that strategy design is not about optimising every parameter but about simple rules and few parameters that capture a real effect.

What people in the thread said

• @arcadia_eth replied that robustness tests are not optional, which matches the research below.
• @MarginCall4 asked why anyone would trade a simple system whose backtest looks ugly. The author answered that a simpler system without curve fitting beats a complex fitted one. That trade-off is real: a slightly worse but robust backtest is often more trustworthy than a perfect one.

These are community comments, summarised in our words; the data and sources below come first.

The fact-checked answer

1. Overfitting, the biggest cause. A 2014 article in the Notices of the American Mathematical Society, "Pseudo-mathematics and financial charlatanism", modelled what happens when researchers try many strategy variations on the same history. Its examples:

• With five years of data, no more than about 45 independent model configurations should be tried; beyond that you are almost guaranteed to find a strategy with an in-sample annual Sharpe ratio of 1 whose expected out-of-sample Sharpe is zero.
• With a two-year backtest, just seven independent configurations are enough for the best to show an expected in-sample Sharpe of 1 with zero true skill.
• The authors' implication: a backtest that does not report how many trials were run cannot be judged.

Every time you tweak a moving average length, add a filter or change a stop and re-run, you add a trial.

Backtest-to-live gaps and how to test for them

Our checklist. For scalping and other high-frequency strategies, costs are often the decisive item; see our page on scalping after costs. Tap a column header to sort.

Cause	What it looks like live	Check before trading
Overfitting	Edge disappears from the first weeks	Count trials; test on data never used in design
Missing costs	Small winners become losers	Add commission, spread and slippage per trade
Look-ahead bias	Signals that could not have been known	Use only data available at the bar's close
Survivorship bias	Universe of today's winners	Include delisted symbols
Regime change	Works in trends, fails in ranges	Split results by volatility regime
Execution	Fills worse than the backtest assumed	Paper trade or trade tiny size first

2. Costs and fills. A strategy averaging a few ticks per trade can be wiped out by commissions and the spread. Backtests often assume a fill at the exact signal price; live orders slip, especially on stops and around news.

3. Data errors. Look-ahead bias uses information that was not available at the time, such as a daily close to trigger a trade during that day. Survivorship bias tests only on assets that still exist.

4. Markets change. A trend-following rule tuned in a trending year will struggle in a choppy one. Splitting results by regime shows whether the edge depends on conditions that may not return.

Why regulators insist on warnings. CFTC Rule 4.41 requires commodity pool operators and trading advisors who show simulated results to state that "these results are based on simulated or hypothetical performance results that have certain inherent limitations" and that they do not represent actual trading. NFA Compliance Rule 2-29 sets standards for promotional material that uses hypothetical results, and the SEC's marketing rule for investment advisers restricts hypothetical performance in advertisements. A backtest is, legally and practically, hypothetical.

UXS decision diagram

Is your backtest ready to trade with real money?

Text version of the diagram

1. Did you record how many variations you tried? If no: Assume overfitting; reduce parameters and retest
2. Does it still work on data you never used in design? If no: The edge is probably in-sample luck
3. Is it profitable after commissions, spread and slippage? If no: It is not a strategy yet
4. Does it hold up across volatility regimes? If no: Trade it only in the regime it suits, with care
5. Did paper or tiny-size live trading match the backtest? If no: Investigate execution before scaling
6. Otherwise: Trade small, track live results against the backtest

Diagram drawn by UXS. A simplified decision aid for education, not a recommendation.

Before trusting a backtest, check the sample size; see how to tell if a strategy actually has an edge.

Decision helper

Related reading: what a good Sharpe ratio is, win rate vs expectancy and whether scalping can be profitable after costs.

Common mistakes

• Optimising every parameter until the curve looks perfect.
• Leaving out costs and slippage.
• Testing only on the data used to design the rules.
• Scaling up before live results match the test.

FAQ

Why do backtests fail in live trading?

Mostly overfitting, missing costs, data errors and changing markets.

What is overfitting in trading?

Tuning rules so closely to past data that they capture noise rather than an edge.

How many backtests are too many?

A 2014 AMS paper found that seven independent tries on two years of data can produce a fake Sharpe of 1.

What is out-of-sample testing?

Testing a strategy on data not used to design it.

Should I include slippage in backtests?

Yes. Ignoring it is one of the most common reasons strategies fail live.

Are backtest results legally "hypothetical"?

Yes. US futures rules require a hypothetical-performance warning.