Source: https://forexiapro.com/research/why-forex-strategies-fail-backtests/

# Why Forex Strategies Fail Backtests: Spread Does It, Our Tests Show

Dylan Forexia14 November 20267 min read

> **In brief:** Why forex strategies fail backtests: in our own tests the pattern was usually there before costs and gone after them. The setup we teach came to about -0.03R a trade with no costs and about -0.16R with spread and commission. A rule that trades often pays the spread each time, and that is larger than the pattern it chases.

Why do forex strategies fail backtests? In every study we have run, cost is the plainest answer. The patterns were mostly real before costs, a rule that trades often pays the spread on each trade, and the spread was larger than the pattern. A backtest that leaves costs out, or guesses them kindly, measures a pattern. A backtest that includes them measures whether a trader could collect it.

R is the amount risked on one trade, so -0.10R means losing a tenth of your risk, on average, each trade. Every figure below was already printed on one of our research pages, and each sits with its caveats there. Nothing on this page was rerun for it.

[![Why forex strategies fail backtests: bars of the loss per trade after costs for one fixed version of our setup in four periods, every period below zero](/seo/mechanical-setup-backtest-2015-2026/forex-strategy-backtest-2015-2026-loss-per-trade.webp)](/seo/mechanical-setup-backtest-2015-2026/forex-strategy-backtest-2015-2026-loss-per-trade.webp)

One fixed version of our setup as a rule, seven dollar majors and gold, 2012 to October 2026, after costs. Every period lost. This chart is from the mechanical setup backtest, where its figures and caveats are.

## Why do forex strategies fail backtests that include costs?

They fail because the move a rule catches is small and the cost of catching it is paid every time. A rule trading a few times a year can carry a wide spread. A rule trading daily cannot, unless the pattern is large.

Three costs matter, and a backtest has to say which it used.

- **Spread.** The gap between the price you buy at and the price you sell at. A fair test buys at the ask and sells at the bid, with the real spread of each bar, not a flat average.
- **Commission and slippage.** The fee a broker charges and the amount a fill differs from the price you saw. Slippage is worse around releases.
- **Financing.** What you pay to hold a position overnight. It does not matter to a trade closed within the day and matters a great deal to a hold of weeks.

A backtest that is silent on all three is usually showing a pattern, not a result.

## How much of a loss was cost, in our own test?

In our test of the setup we teach, most of it. We ran the setup as a fixed rule, with no judgement, on seven dollar majors and gold, and added the costs one at a time. Over the fitting years, February 2021 to December 2024, the same trades gave:

| Costs included | Result per trade |
| --- | --- |
| None | about -0.03R |
| The real spread on every bar | about -0.10R |
| Spread, commission and slippage | about -0.16R |

In the fixed version for 2021 to 2024, 8,931 trades lost 1,093R in all, and spread and commission accounted for about 821R of that, roughly three quarters. Take the costs away and what was left was close to zero, not a profit waiting to be freed. The full study, the split fixed before testing and the places it does not apply are on [our mechanical setup backtest](/research/mechanical-setup-backtest-2015-2026/).

On gold the same sum is harsher to read in dollars. The typical spread in the sample was $0.39, against a median hunt of $7 past the Asian range, which sounds small. It is not small once the rule fires thousands of times, and [our gold strategies page](/research/gold-strategies-tested/) shows the setup lost on buys and on sells in every period.

## Did the patterns in published research survive costs?

Mostly not. We took sixteen rules from published work, ran twenty tests on forex and gold with real bid and ask prices and financing, and wrote the pass mark before the first run. None passed. One was kept as a candidate and nineteen failed.

The detail is in [our test of published trading strategies](/research/published-trading-strategies-backtest/). Two findings from it belong here.

**The patterns were there before costs.** Gold made its whole rise outside the hours between the two London auctions, as the rule says, and EUR/USD fell in Europe's hours and rose in America's. The papers were not wrong about the pattern. Trading either one every day pays the spread every day, and the study judged that larger than the effect.

**Holding costs money too.** The monthly tests of carry, trend and momentum paid 2 basis points of turnover each way and 1% a year on gross exposure as a retail financing markup, and on the majors since 2006 none of them stayed above the mark. That markup is an assumption, and a desk paying less would face a different sum. We did not test that.

The study states most results in percent a year and not in R, because those tests have no stop, so no R exists. That is why some rows are described in words.

## Is cost the only reason strategies fail?

No. Cost is the cause we can measure most cleanly. The others we met are about how the test was run.

**Trying many settings.** Our test of Gann's methods ran 2,177 variations on 35 markets, with costs of 0.015% a round trip on the forex majors. A variation had to be chosen on one period, hold on a second and only then be read on a third. 43 passed all three, and the same test on shuffled candles passes about 15. That is what trying many things lets through by chance. Detail and caveats: [our Gann levels backtest](/research/gann-levels-75-markets/).

**Charts that were picked afterwards.** A marked-up chart is chosen once price has reversed, which is close to the definition of a good trade. The same program that looked clean on those days lost on the days that never became a markup. Our [liquidity sweep counts](/research/how-often-does-a-liquidity-sweep-reverse/) are of price behaviour, not of trades, for that reason.

**A rule that looks good on a slice.** A slice of markets, hours or weekdays that made money in one period nearly always lost in the next, about as much as the slices that had lost. Choosing by past profit chose noise.

## What should you check before you trust a backtest?

These questions would have caught most of what we found in our own work. You can ask them of anything you read, ours included.

1. **Were costs per bar, or a flat guess?** A flat spread understates the cost at the rollover and around news.
2. **Does it say how often the rule trades?** The more trades, the more spread paid.
3. **Was the split fixed before testing?** A rule tuned on all the data has not been tested on any of it.
4. **How many things were tried?** A winner among thousands needs a control, such as shuffled candles, to show it is more than luck.
5. **Does it hold in a period that shares no data?** Ours lost by about the same amount across periods that overlap nowhere.
6. **Were the best examples picked afterwards?** If so, they show what happened, not what the rule earns.
7. **What does it not prove?** An honest page has a section for it.

Our reading is that vendors publish results that pass, and that is why the failures are worth printing. The [archive of every test we ran](/research/forexia-research-archive/), losses included, lists them in one place. A reversal you spot on a finished chart is not a trade you could have taken at the time. The same filter belongs on any read of [a false breakout](/price-action/false-breakout-forex/) and of [the stop hunt](/glossary/the-stop-hunt/): did it pay after costs, on days that were not chosen afterwards?

## What does this page not prove?

It does not show that every strategy fails, or that costs are the only cause.

- **It is our rules and the published rules we chose.** Twenty tests are not all of published research, and most of that literature is on stocks.
- **It is one data source for each kind of data.** Another broker's spreads and swaps will differ, and the financing markup is an assumption.
- **Five-minute candles flatter the fills.** Real fills during a release are worse than they show.
- **It is past data.** Nothing here was traded live. A pattern that failed with costs can still appear in a market with smaller costs, and one that passed can fade.
- **A rule that fails after costs is not always useless.** It may still describe what price did. Describing a day and earning from it are different claims.

For the one slower idea that did better in our work, with its caveats beside it, read [the COT crowd study](/research/fading-the-crowd-cot-level-hunt/). It is not proven. The whole set is indexed on [the research page](/research/).

## How was this page made?

It was written on 13 November 2026 from the research posts named above, after the owner's decision on 11 November 2026 to republish them. Every figure is the one printed on its own page, after the rerun of 11 November 2026. No new test was run for it, and no figure was added.

Course 3 of ForexiaPRO, Reverse Price Action, teaches how to read what the hunt did, and its lesson on a hunt that keeps going shows why a rule needs more than a shape. See [ForexiaPRO](/pro).

## Frequently asked questions

### Why do most forex strategies fail a backtest?

In our tests the main reason was cost. The patterns were usually there before costs, and a rule that trades often pays the spread every time, which is larger than the pattern. Our own setup came to about -0.03R a trade with no costs and about -0.16R with spread and commission. Chosen settings that only look good by luck are the other common reason.

### Do you have to include the spread in a forex backtest?

Yes. A backtest on mid prices, or with a flat guessed spread, can show a profit that a real account would not have. We bought at the ask and sold at the bid, used the real spread on every bar and added commission and slippage. Spreads are wider at the rollover and around news, which is when many rules trade.

### Does a backtest with costs predict live results?

No. It tells you whether a rule could have paid for itself on past data, which is a necessary test and not a sufficient one. Real fills during a release are worse than five-minute candles show. A rule that fails with costs is unlikely to work live, and one that passes still has to be tried on days it has not seen.

### How many trades does a forex backtest need?

There is no single number. What mattered in our tests was the same result in periods that share no data, with the costs stated and a split fixed before testing. A large count of trades with no costs stated tells you very little.

### Why did the strategies in published papers fail in your test?

We tested whether a retail trader could collect the pattern, not whether the pattern exists. Many patterns were there before costs. After the spread, financing and a pass mark written before the run, none of the twenty passed, and one was kept as a candidate.

## More from the blog

### [Confluence Trading Forex: Why One Reason Is a Guess, and What a Stack Is](/signature-trade/confluence-trading-forex/)

Confluence trading forex means more than one reason on a chart. What counts as one reason, what counts twice, and one day that stacked six of eight.

### [Best Forex Trading Books: Seven Worth Reading, and What Each Gets Wrong](/learn-forex/best-forex-trading-books/)

Best forex trading books: seven worth reading, from psychology to candles, and what each gets wrong. A reading order for beginners. No returns promised.

### [Average Daily Range on Tuesday: What ADR Met Tuesday Means](/weekly-cycle/average-daily-range-tuesday/)

What it means when the average daily range is spent by Tuesday, how often it happens (28% of Tuesdays), and why it is one reading and not a signal.