Testing
A backtest is an argument. A forward test is evidence.
Both answer "does this work", and only one of them does it without knowing the ending. That difference is not a technicality — it is most of the reason strategies that looked excellent in testing behave differently the week real money is attached.
The difference in one line
A backtest runs a rule over data that already happened. A forward test runs it on data as it arrives.
Everything else follows from that. A backtest is fast, repeatable and cheap, and it was built by someone who already knew what the market did next. A forward test is slow, unrepeatable and costs you real time, and nobody involved knows the ending — including you.
Neither is optional. A backtest that takes an afternoon can reject an idea that a forward test would have taken three months to reject, and that is genuinely valuable. It just cannot do the opposite: no backtest can confirm an idea, because confirmation requires not knowing.
The four ways a backtest flatters itself
All four are easy to introduce by accident, and three of them are invisible in the result.
- Overfitting
- Enough parameter tuning will fit the noise rather than the signal. The tell is fragility: if changing a moving average from 20 to 21 destroys the result, you have fitted the sample rather than found an effect.
- Look-ahead
- A calculation quietly uses information from a bar it could not have had yet. Entering on the open of the bar whose close triggered the signal is the classic version, and it is worth a great deal of imaginary money.
- Survivorship
- The instruments that stopped existing are missing from the data, so the sample only contains things that lasted. Less severe in futures than in equities, but it is not zero.
- Omitted costs
- Commission and slippage left out, or applied as a flat number that does not reflect the size or the hour. A strategy averaging $30 a trade loses an eighth of its edge to $4 of costs, and edges that thin are exactly the ones a backtest produces.
The number that actually matters
Not the return. The gap between the backtest and the forward record.
Every strategy degrades from backtest to live — that is normal and expected, because the backtest was the best case by construction. What tells you something is how much. A strategy that returned 40% in testing and 32% forward has degraded 20%, which is ordinary. One that returned 40% and is running at 4% has not degraded, it was fitted.
Which is why the two records should never be shown as one curve. Published strategies here keep them separate for exactly this reason: the backtest each was published with, and separately the live record since publication. The second is the one that can embarrass us, which is the only reason it is worth showing.
If a strategy is sold to you with a single blended equity curve, ask where the publication date sits on it.
How to forward test without waiting a year
The honest objection to forward testing is that it is slow. Three ways to compress it, in ascending order of what they prove.
- 01Paper accounts. The strategy fires its real signals and fills against live prices with commission and slippage you set per side. Real conditions, real timing, no capital. It answers how often it fires and what a bad session looks like — the two things a backtest answers worst.
- 02Simulated broker account. The same setup pointed at a broker's sim environment, which adds the broker's own behaviour: rejections, session boundaries, contract rolls. Nothing about the configuration changes when you later point it at a funded account.
- 03Live at minimum size. One micro contract. It costs almost nothing and it adds the two things no simulation contains: real fills against real liquidity, and your own behaviour when the money is real. That second one changes more results than most people expect.
- 04Watch frequency before P&L. In the first fortnight, how often it fires and when tells you more than the running total. A strategy that fires twice in two weeks has not been tested regardless of whether those two trades won.
- 05Judge the worst stretch, not the average. The number that decides whether you can actually run something is its worst run at your size, because that is the point at which people abandon a working strategy.
Why costs have to be in from the start
Adding costs at the end is the most common way a forward test ends up as optimistic as the backtest it was meant to check.
Commission is per contract per side, so a round trip on ten micros pays twenty commissions. Slippage is not a constant — it is worse on the open, worse in thin hours, worse on stops than on limits, and worse the larger you are relative to what is resting at the price.
A test that applies a flat half-tick to everything is still a useful test, but it is a friendly one. If the strategy only works at zero cost, that is not a strategy, and the sooner the arithmetic says so the less it costs to find out.
Paper accounts here take a per-side commission and a slippage figure precisely so the number you are looking at is after costs rather than before them. The expectancy calculator makes the same point with your own averages: an edge of $30 a trade with $4 of costs is a $26 edge, and the difference compounds across every trade you will ever take.
Common questions
What is the difference between forward testing and backtesting?
A backtest runs a strategy over historical data that has already happened; a forward test runs it on data as it arrives. The backtest was built by someone who knew what came next, which is why it can reject ideas cheaply but cannot confirm them.
Is forward testing better than backtesting?
They answer different questions. A backtest rejects bad ideas in an afternoon rather than a quarter, which is genuinely valuable. A forward test is the only one that produces evidence, because nobody involved knows the outcome while it runs. Use both, in that order.
How long should a forward test run?
Long enough to contain a bad stretch, which is a function of how often the strategy trades rather than of calendar time. A strategy firing twice a week needs months; one firing several times a day needs weeks. Judge frequency first and P&L second.
What is overfitting in trading?
Tuning parameters until the rule fits the noise in your sample rather than a real effect. The test is fragility: if a small change to a parameter destroys the result, you have described the sample rather than the market.
Can paper trading replace forward testing?
Paper trading is a form of forward testing — real conditions, simulated fills. What it cannot contain is real fills against real liquidity and your own behaviour when the money is real. Both change results, so a period at minimum live size is worth doing after paper.
Why do backtested strategies fail live?
Usually one of four things: the parameters were fitted to the sample, a calculation used information it could not have had, costs were omitted or understated, or the strategy was fine and the trader abandoned it during a drawdown the backtest had shown but nobody had felt.
Start on a simulated account.
Connect a sim or evaluation account and run the whole product against it before you point anything at live money. Nothing about the setup changes when you do.