Paper trading and backtesting without fooling yourself

Paper trading and backtesting are two different activities that people treat as one. Backtesting is running a rule over past data to see what it would have done. Paper trading is making decisions forward in time without money. Both are useful, and they fail in opposite directions.

The distinction that matters: a backtest is retrospective and therefore exposed to knowing the answer; paper trading is forward and therefore exposed to being too short to mean anything.

Quick answer

Backtesting checks whether a rule is worth trying. Paper trading checks whether you can execute it. A rule that backtests well and that you cannot follow is worth nothing, and a test of your own discipline over a few weeks with no losses has tested nothing.

The failures that make a backtest meaningless

These are not exotic. Most backtests that look impressive contain at least one.

• Look-ahead bias. Using information that would not have been available at the time. The simplest version is testing on data you have already seen and shaped your rule around.
• Survivorship bias. Testing on instruments that still exist. The ones that failed are absent, and they are exactly the ones a rule might have bought.
• Overfitting. Adjusting the rule until the past fits. Every adjustment makes the past fit better and the future less likely to, and there is no way to detect this from the result alone.
• Ignoring costs. Fees, spread and slippage turn a marginal rule into a losing one. A backtest without costs is an arithmetic exercise, not a test.
• Testing on too little. A rule tested over a short window that happened to suit it has been tested on one example, not on a distribution.

What a backtest can honestly tell you

It can tell you whether a rule is worth paper trading. That is a modest claim and it is the correct one. It cannot tell you what the rule will return, because the conditions will differ, and presenting a historical result as an expectation is the single most common misrepresentation in this area.

The useful outputs are structural rather than numerical: how often the rule trades, how large the worst losing run was, whether the losses cluster, and how sensitive the result is to the parameters. A rule that only works at one parameter value is a coincidence that has been fitted.

What paper trading can honestly tell you

It can tell you whether you follow the rule when nothing forces you to. That is about behaviour, and behaviour is the part that survives into real trading.

It cannot tell you how you behave when the money is real. That gap is not a flaw in the method; it is the boundary of what the method can reach, and the honest conclusion is that the first live period is still a test rather than an application.

Doing both in the right order

1. Write the rule down precisely enough that someone else could follow it.
2. Backtest on a portion of the data and hold the rest back.
3. Run once on the held-back portion. Accept the result rather than adjusting the rule.
4. Paper trade forward for long enough to include losses and recoveries.
5. Move to a small real amount, and judge your behaviour rather than the result.

The next question after practice is how much a mistake costs, which is in risk and position size. What the practice is for is in why practise before you trade.

Parameters are where a backtest lies to you

Any rule has settings: how far a stop sits, how long a position is held, when a condition counts as met. Adjusting them until the past fits is the most common form of self-deception in this area, and it is invisible in the output because the output improves.

The check is sensitivity. If the rule only works at one setting and fails at the settings either side, it has been fitted to noise. A rule that degrades gracefully as the parameters move is a rule that is describing something real.

Costs change the answer more often than people expect

A rule that trades frequently spends a substantial part of any advantage on spread, fees and slippage. Backtests that omit costs routinely show rules that are negative once costs are included, and the frequency of trading is what decides it.

The practical move is to model costs pessimistically. A rule that survives pessimistic costs is worth paper trading; a rule that only works with optimistic ones has been tested against a market that does not exist.