What If You Do the Exact Opposite of a Losing Trading Strategy? 435 Backtests Reversed

00The short answer

• Reversing a losing strategy rarely works. We picked losers on early data only, then measured them on later data. Of the 110 versions that lost before costs in their early years, 27 (25%) ended positive after costs once reversed.
• Leaving them alone did just as well. The same 110 versions, kept as they were: also 27 positive. The other 56 lost in both directions.
• The reason is arithmetic. Reversing flips the result before costs. It does not flip the costs. A reversed strategy ends positive only if the original lost more before costs than it paid in costs. Of the 257 versions that lost after costs, 149 lost only because of costs, and reversing cannot rescue those.
• Nothing passed our tests. 0 of 435 reversed versions passed all six of our gates. The strongest early result had a t-statistic of 1.60; the gate asks for 2.5.
• One thing did carry information. A version that lost before costs early was a better candidate than one that did not: 25% against 10% positive once reversed, more than chance explains (shuffle p 0.002). A better candidate, not a good bet: three in four still did not finish positive.

01Why reversing does not reverse the costs

A trade has two parts: the move, which changes sign when you take the other side, and the cost, which does not. So the reversed result is minus the result before costs, minus the costs. Three versions of the same sum, with a cost of $30 a trade:

The test fits on one line: reversing pays only when the loss before costs is larger than the costs. This condition is not new; guides on reversing forex strategies state it for the spread. What was missing is a count on real logs, with the choice made before the result.

A real pair: the same trades, both sides

We have already tested one exact mirror. On our VWAP test on ES futures, one version trades with the VWAP and the other trades against it, at the same times:

One side made $106,088 before costs and the other lost exactly that. Each then paid the same costs. Our reversal formula applied to VW-TREND gives VW-REV to the cent, trade by trade, which is one of the checks this study had to pass.

02What we did

• The data: every version with a per-trade log on this site, 435 versions of 68 tests, each in its own unit (dollars, R, pips or percent). No price data was read and no strategy was run again. Tests in different units are compared only through shares and counts.
• The reversal: the same entry and exit on the other side, paying the same cost. Reversed result before costs = minus the original; after costs = minus the original before costs, minus the cost. Trade by trade.
• Early and later data: the same three periods each test already used, by entry date. “Early” is the development period; “later” is validation plus holdout. All of the holdout periods had already been used by the original tests.
• Written down first: the rules, the checks and a prediction were frozen before any reversed number was computed (the specification). We chose this study after seeing every published result, which is why the choice of what to reverse comes from early data only.

03With hindsight: 37 of 257

Start with hindsight: take the whole history of every version, look at who lost, and reverse them. This is not evidence, because you can only pick the losers after the fact.

Even knowing the answer, only 37 of 257 losers (14%) turn positive, about one in seven; among the 108 that lost before costs too, about one in three. The zeros are arithmetic, not findings: a version with a profit before costs has a loss before costs when reversed, and then pays the costs as well.

04The real test: choose early, measure later

A trader cannot reverse a strategy on results that have not happened yet. So each rule below sees only the development period, and the outcome is measured only on the validation and holdout periods that follow. 423 versions in 67 tests have at least 30 trades in both parts; 12 do not. In each cell, the first number counts versions; the number in brackets gives each test the same weight, because versions of one test share their data.

* This column is not in the specification; we added it because it is the comparison a trader needs. The shuffle compares chosen against not chosen: the rule’s label moves among the versions inside each test, 10,000 times.

Rule A is the obvious rule, and it did worst. Half of what it picks (108 of 218) had a profit before costs in development, so reversing them starts from the wrong side.

Rule B is about a coin flip before costs and one in four after. Before costs, the reversed side was positive later 46% of the time (51% with each test weighing the same). The costs take it from there. By count it ties with leaving the versions alone, 27 each; with each test weighing the same it is 28.5% against 22.1%. Neither is a strategy you would want to fund.

The difference from the versions not chosen looks real, but read it carefully. It is 17.6 points, of which the make-up of the tests alone gives about 4.0; only 0.2% of shuffles reached it. This says the label carries information, not that reversing pays. Part of it is the other group: versions with a profit before costs in development kept one later 70% of the time, so reversing them was especially bad.

Rule C has too few versions to say anything. Only four versions were that clearly bad: Fibonacci 61.8% (EUR/USD 1H) FB-LS; Fibonacci scalping, Google AI (ES) FO-1M-VWAP; WaveTrend (gold, 1-hour) WT-LS-X; Order block + fair value gap, SMC (EURUSD, 5-min) OB-FVG-C. Reversed, three were positive later before costs and one after costs (WT-LS-X). As they were, none was positive after costs.

Our prediction, frozen before the run: rules A and B would leave fewer than half positive after costs, which held. Rule C would do best, which did not: at 25% it is level with rule B by count and below it with each test weighing the same, on four versions.

Without the 61 versions whose logs have no column for the result before costs (we used their average cost per trade, section 09), rule B gives 28.4% (34.5%), with a shuffle p of 0.001. The conclusion does not change.

05What cheaper trading would change

Costs are charged per trade, so the same shares follow exactly at any other cost. Reversed versions positive later, by count and with each test weighing the same:

Free trading would bring rule B only to about a coin flip. There was little to rescue here even before costs. For how much costs take from strategies as they stand, see how much trading costs eat across our backtests.

06Would any reversed version pass our tests?

We put every reversed version through the same six gates as a new strategy. Testing 435 reversed versions is 435 more tries, so anything that passed would be a candidate for a new test, not a result. Nothing passed.

For comparison, 3 of the 435 originals passed all six in their own runs. One of them, the 1-hour golden cross on gold, was later changed to Inconclusive after a data check; the other two are the New York noise area versions on NQ, now being forward-tested. The control gate reads as 100 minus the version’s published percentile: moved trades reverse as well, so a version at the 3rd percentile becomes the 97th. The real pair confirms it: VW-TREND sits at 55.4 and VW-REV at 44.6. This holds exactly only if a trade’s cost does not change when it moves. 28 versions have no readable percentile (three golden cross tests) and count as not assessed.

The closest two

Both come from our Bitcoin CME gap test, where fading the gap did worse than random timing. Reversed, each passes five gates and fails the early t-statistic, on only 128 later trades. CG-S1 is also the only one of the 27 rule B winners with a later t-statistic above 2 (2.34). That is a reason to write a proper rule and test it, not a result. Another 8 versions pass four gates, 5 of them on USD/JPY in our forex study.

07Where reversing came closest

A description of the 110 rule B versions, not a verdict: the groups are small and overlap. Style labels come from our trend following vs mean reversion guide.

Reversed breakout rules, which means fading breakouts, came closest to half (9 of 19), but the largest later t-statistic among those 9 was 1.15. Reversed trend rules did worst of the larger groups. Long-only rules reversed became short-only in markets that mostly rose. No SPY version was chosen: none with enough trades lost before costs in its early years.

The 10 most reversible in development, and what happened next

3 of the 10 stayed positive. The one that looked best early, the 61.8% Fibonacci pullback on EUR/USD, then went to −161.3 R.

08Every test, reversed

One row per published test, by market. “Reversed positive” is the whole history, with hindsight. The next three columns are the real test: versions chosen by rule B in development, how many ended positive later once reversed, and how many ended positive later as they were.

The golden cross tests with no readable control percentile can reach at most five gates here.

09What this study does not say

• It is not a new run of a mirrored rule. We took the other side of each trade at the same prices. In a real mirrored rule a stop becomes a target and a target becomes a stop, and fills differ. A reversed version that looks good here still has to be written as a rule and tested.
• It is generous to reversing. There is no short-borrow fee, and a dividend or swap already inside a result before costs simply changes sign.
• It is not a sample of all strategies. These are 68 tests of mostly popular rules we chose to test. A strategy built to lose by a wide margin before costs would behave differently, if you could find one in advance.
• The versions are not independent. Versions of one test share data and many tests share markets, so 435 versions are far fewer than 435 separate experiments. The shuffle keeps labels inside each test, which handles only the first of these.
• It is exploratory. Every period had been seen before these rules were written, even though the rules were not fitted to them.
• 61 versions in 15 tests use an average cost. Their logs have no column for the result before costs, so each trade is charged the version’s average cost from its published summary.

10What to do with a losing strategy

1. Split the result into before and after costs. If it made money before costs, reversing can only make it worse. The question is whether costs can come down or the strategy can trade less: our trading cost calculator does the sum.
2. Compare the loss before costs with the cost per trade. Reversing pays only when the loss is the larger of the two. On most of our versions it was not.
3. Choose on one period and measure on another. Reversing over the whole history shows you 37 winners that were only visible afterwards. In our test a version’s early loss before costs said little about the sign of its later result.
4. If it still looks interesting, write the mirrored rule as a new strategy, with its own stop and target, count it as another try, and test it against random entries. See how to backtest a trading strategy without fooling yourself.
5. Otherwise bury it. A strategy with no signal has no signal in either direction.

FAQCommon questions

Does reversing a losing trading strategy work?

Rarely in our tests. Choosing on early data only, 27 of 110 versions that lost before costs (25%) ended positive after costs on later data once reversed. The same 110 left as they were also gave 27, and 56 lost in both directions.

If a strategy loses money, why not just trade the opposite?

Because taking the other side flips the result before costs, not the costs. The reversed result is minus the result before costs, minus the costs, so it is positive only if the original lost more before costs than it paid in costs. Of the 257 versions that lost after costs in our tests, 149 lost only because of costs, and reversing cannot rescue those.

Can both sides of a strategy lose money?

Yes. On our VWAP test on ES futures, the same 11,707 trades on one side made $106,088 before costs and on the other lost exactly that. After $360,107 of costs each, one side ended at -$254,020 and the other at -$466,195.

Can I reverse a losing EA or trading bot?

You can, but the loss before costs per trade has to be larger than your spread and commission, and you should measure it on data you did not choose it from. A real reversed EA also swaps its stop and target, which our numbers do not: we took the other side at the same prices, which is generous to reversing.

Is a consistently losing strategy a good one to reverse?

A better candidate than most, not a good bet. Versions that lost before costs early ended positive once reversed 25% of the time, against 10% for the others, more than chance explains. But only 4 of our versions were clearly bad before costs in their early years (t-statistic of -2 or lower), and 1 of those ended positive once reversed.

How was this tested?

On the published trade logs of every version on this site, with no new backtest. The rules were frozen before any reversed number was computed, what to reverse was chosen on the development period only, and the outcome was measured on the validation and holdout periods. A second, independent script matched every version. Read the protocol.

11Check our arithmetic

12Corrections

None since publication.

+More tests and guides

• Trading costs
• Trend following vs mean reversion
• Vwap strategy backtest
• Bitcoin cme gap
• Free prop firm challenge calculator
• What all the backtests found so far
• The backtesting protocol
• How to backtest a trading strategy without fooling yourself
• Backtesting glossary