Six Strategies, Four Currency Pairs

00The verdict in three lines

• A second, unrelated market is a robustness test: an edge that holds on several markets is more convincing, although an edge can also exist in one market only, so failing to transfer does not rule that out. The bar used here is this study's bar, not a universal definition of an edge. We took the six strategies already tested on ES, kept every rule and default setting, and ran them on four currency pairs: 24 variants per pair, 96 in total, with the rules, the gates, the costs and a state-matched random control fixed in a dated specification before any result.
• None passed: 0 of 96 cleared every gate, and no variant came near the bar for support, a pass on at least 3 of the 4 pairs. Every one of the 96 failed the development t-statistic gate (the highest was 1.61, 2.5 needed). 65 were positive before costs and 30 after: across 347,256 trades the average gross result was about 0.36 pip, against a cost of about 1 pip. The closest call was EURUSD A, which failed only gate 3.
• Cause of death: No proven edge. Costs turned 35 of the 65 results that were positive before costs into losses, and the other 30 that stayed above zero after costs did not meet all the criteria. Bollinger, whose long side was our best hint on ES, finished above zero in none of its 16 results here.

01What we tested

The six strategies from our ES posts, each with its four variants (long and short or long only, with a simple exit or with a fixed stop and target), at the same default settings and nothing tuned for FX. The rules are those in each ES post and are listed in the specification.

Differences from the ES versions, all forced by the market: the cost model and units; stops and targets in price units; the Supertrend 5- and 15-minute “trading hours only” versions are not run; and the RSI(2) entry window (10:00 to 16:00 New York) is kept because it is part of that strategy's rules, even though FX trades around the clock. Long-only versions have no natural drift rationale on FX; they are kept so the variants stay identical to the ES tests.

The cross-market rule, fixed in advance: a variant counts as supported across markets only if it passes every gate on at least 3 of the 4 pairs. A pass on one or two pairs is a single-pair result, not support.

02Data and costs

Data: spot FX, 1-minute bid history from an IC Markets MT4 demo server, converted to New York time, 2004 to late September 2026. This is a broker demo download, and early years may contain interpolated bars. It is not futures data. Periods: development 2004 to 2012, validation 2013 to 2019, holdout 2020 to 2026. The holdout is not clean: the 2020 to 2026 FX data was looked at in an earlier test of a different idea.

Costs: Our IC Markets account is a Standard account (no commission, wider spread). The specification priced costs at the Raw Spread commission plus an assumed spread, about 1.0 pip per round trip in total (1.2 on USDJPY), which is a middle figure for either account type. The half-cost and double-cost results below show how much that choice matters. The MT4 history records no spreads, so the cost is an assumption and the same in every year; spreads in 2004 to 2012 were wider than today's, which flatters the development period.

Engine check: the engine is the ES engine with the costs made a parameter. Before any FX run it reproduced all 24 published ES results trade for trade, and the published ES control percentiles (fx_selftest_result.txt).

03Results: every strategy on every pair

Each cell is the net result in pips after costs, and how many of the six gates that result failed. The full numbers for all 96 results are in the tables below.

RSI(2) Dip Buyer

Supertrend

Donchian Channel Breakout

Bollinger Band Reversal

RSI Divergence

MACD Crossover

By strategy

The strategies that earned something before costs were Supertrend (16 of 16) and MACD (15 of 16). Bollinger was positive before costs in 4 of 16 and, after costs, in 0: the long-side hint from ES did not travel.

04Did anything work on more than one pair?

This is the robustness test the study was built for: a market-specific edge could still exist, and the question here is whether anything holds across these four pairs. Variant by variant, the number of pairs (out of four) where the result finished above zero before costs, above zero after costs, above zero in the holdout, and where it cleared all six gates.

* Holdout already used before this test. “Supported across markets” needs a pass on at least 3 of the 4 pairs.

No variant was supported. No variant passed on even one pair. The variants that finished above zero after costs on three pairs were ST60-L (1 of 4 in the holdout) and MC-L-F (2 of 4 in the holdout), and none of them cleared the development t-statistic or the validation profit factor gate on any of the four pairs.

05The gates

Gate 3 is the wall: no result reached a development t-statistic of 2.5, and none reached 2.0 either. The matched control was cleared by 11 of 96 results (EURUSD A, EURUSD ST60-LS-T, EURUSD DV-L-S, EURUSD DV-L-F, GBPUSD ST60-LS-T, GBPUSD ST60-L-T, USDJPY MC-LS-S, USDJPY MC-L-S, AUDUSD DV-LS-S, AUDUSD MC-LS-S, AUDUSD MC-L-S); at the 95th percentile chance alone would clear about 5. Many of these share a signal across correlated pairs, so they are not 96 independent chances, but a few signals do seem to carry some timing information before costs. They fail on the other gates. Gates failed per result: 1 result failed 1 gate, 3 results failed 2 gates, 6 results failed 3 gates, 20 results failed 4 gates, 66 results failed 5 gates. The only result that failed a single gate was EURUSD A, on the development t-statistic (1.36, with a validation profit factor of 1.21 and a holdout of +44 pips).

06Costs

The 96 results made +125,671 pips before costs and -238,575 pips after, on 347,256 trades. Costs turned 35 of the 65 results that were positive before costs into losses. Because the cost per trade is a constant, the result at another cost follows exactly: net = gross minus trades times cost.

At half the cost 44 of 96 results would be positive, at double the cost 4. We have not re-run the other gates at other costs, so this shows only the sign of the result. Even at half the cost the total across all 96 results would be -56,452 pips.

07Every result

All 96 results, one table per pair. “vs matched random” is the percentile of the average net per trade among 500 random runs of the control matched on each signal's entry state (95 needed). The trade logs for every pair are zipped next to this page.

EURUSD

GBPUSD

USDJPY

AUDUSD

* Holdout already used before this test. PF = profit factor, in pips. Results are in pips per unit, no compounding. Net after the 10 best trades are removed is in pips.

08Cause of death

What this test covers, exactly: these six rule sets at their default settings on hourly bars of four major pairs, 2004 to 2026, with a flat cost of about 1 pip per round trip. It says nothing about the same ideas on other timeframes, other pairs, or with carry. What it does show is that the tested configurations did not demonstrate the robustness the protocol requires across these four pairs.

What we did not test: daily bars, 5- and 15-minute bars, cross rates and minor pairs, carry (swap) income, session or news filters, volatility-based position sizing, and a cost that varies by year. We used the ES defaults on purpose: tuning would raise the odds of finding a lucky setting, not a real edge.

What would change our mind: a variant, declared in advance, that clears every gate on at least three of the four pairs, ideally on data we have not looked at, or a forward sample of new trades with a clearly positive result after real costs.

TWhat this shows, and what to take into your own tests

What the results show

• None of the 96 variants passed every gate, and none came near the bar for support (a pass on at least 3 of the 4 pairs).
• All 96 failed the development t-statistic gate: the highest was 1.61 against the 2.5 required.
• 65 of 96 results were positive before costs and 30 after. Across 347,256 trades the average gross result was about +0.36 pip a trade, against a cost of about 1 pip.

What remains unknown

• Other timeframes (daily bars, or 5 and 15 minutes), other pairs and cross rates, none of which we tested.
• Carry (the swap a position held for days earns or pays), which we did not model.
• Real spreads over 2004 to 2012: our flat cost may be too low or too high for individual years. At half the cost, 44 of the 96 results would have been net positive (we have not re-run the other gates at other costs).
• Whether the small gross edge Supertrend and MACD show on most pairs could be kept with far fewer trades.

Why it failed

These strategies trade a lot, about 3,617 trades per variant over 22 years, and the gross result per trade (0.36 pip on average) is smaller than the cost per trade (about 1 pip). Costs turned 35 of the 65 results that were positive before costs into losses. The other 30 stayed above zero after costs, and none of them met all the criteria: all 30 failed the development t-statistic, 27 the validation profit factor, 20 the matched control, 22 the ten-best-trades test and 15 the holdout. Supertrend (positive before costs in 16 of 16 results) and MACD (15 of 16) came closest, with 9 and 6 results above zero after costs; none of these 15 cleared the development t-statistic or reached a validation profit factor of 1.2. The picture also faded over time: of the 96 results, 50 finished above zero in development, 26 in validation and 23 in the holdout.

What you can apply to your own tests

• A second, unrelated market is a good robustness check before you trust a first result. An edge can still be specific to one market, but then it needs other evidence. The Bollinger long side was our best hint on ES; here it finished above zero in none of its 16 results.
• Count your costs in the unit of the edge. With about 1 pip a trade, a strategy has to earn more than a pip gross on average before it earns anything.
• Write the bar for “works on more than one market” before you see the results. Ours was a pass on at least 3 of 4 pairs, and it removed any temptation to promote a lucky pair.
• Split results by period. 50 of 96 results were positive in development against 23 in the holdout, which is what results that fade look like.

A note on the numbers: drawdown figures are measured on closed trades unless a table says it includes open losses. Open losses inside a trade are not counted in a closed-trade drawdown, so real account risk is higher, especially for systems with no stop. Backtests are not live trading.

FAQCommon questions

Do technical indicators work on forex?

In our backtest of six popular indicator strategies on four major pairs (96 variants), 0 passed every gate. 30 finished above zero after costs, but each of those failed other tests, such as the development t-statistic.

Does a strategy that works on one currency pair work on the others?

We set the bar in advance: a variant counts as supported only if it passes every gate on at least 3 of the 4 pairs. No variant came close, and none passed on even one pair. That bar is this study's, not a universal definition of an edge.

How much do trading costs matter in forex strategies?

We charged about 1 pip per round trip (1.2 on USDJPY). Across 347,256 trades the 96 variants made +125,671 pips before costs and -238,575 pips after. Costs turned 35 of the 65 results that were positive before costs into losses.

How was the forex study tested?

On 22 years of spot FX, with rules written and dated before the data was run, costs included on every trade, results split into development, validation and holdout periods, a random-entry control as the benchmark, and every variant reported. The specification was published before any result. Read the protocol.

09Check our arithmetic

Code provenance: our own code, written from the standard definitions of the indicators. No third-party source code is used or published.

10Clock audit (added 9 Oct 2026)

We checked the time stamps of this broker's MetaTrader 4 history against an outside event: the price jump at the US data releases that come out at 08:30 New York (the payrolls, the weekly jobless claims). Week by week, the jump should sit in the minute 08:30 to 08:31. For the currency pairs it sat one hour early in many weeks: on EUR/USD, for example, in 541 of 1,205 weeks, mostly 2004 to 2014, plus a few weeks a year after that when the US and European clocks change on different dates. In 297 of those weeks there was no clear release, so the week took the call of the nearest week that had one; that part is an estimate. We moved the time stamps of those weeks by one hour (prices untouched) and ran the same code again on the same data (clock_audit_compare.txt). The same pattern holds on GBP/USD, USD/JPY and AUD/USD. Two rules of this study depend on the time of day: the RSI(2) entries are allowed only from 10:00 to 16:00 New York, and every position is closed at the Friday bar closing at or after 16:00. So we re-ran all 96 variants with the corrected time stamps.

What changes. No verdict: 0 of 96 pass with either clock, and none comes near a pass on three of the four pairs. The figures quoted in this post come from the run with the clock as published, and several of them move: 4 results change sign after costs (USDJPY B, USDJPY DC-LS-S, AUDUSD DV-L-S, AUDUSD MC-L-S), and the number of gates passed changes for 6 (EURUSD A (5 → 4), EURUSD B (2 → 4), EURUSD DV-L-S (2 → 1), GBPUSD ST60-L-T (3 → 4), USDJPY ST60-L (3 → 2), AUDUSD DV-LS-S (4 → 3)). The closest call, EURUSD A, passes 4 gates instead of 5 with the corrected clock, because its holdout turns negative. Every variant's figures under both clocks are in clock_audit_compare.txt.

11Corrections

8 Oct 2026, after outside review: (1) the introduction no longer says a real edge must show up on a second market. Transfer to other markets is a robustness check; an edge can be specific to one market, and the 3-of-4-pairs bar is this study's bar, not a universal definition of an edge. (2) The “Why it failed” text no longer says the gross edges of Supertrend and MACD were too small to pay the costs: 9 Supertrend and 6 MACD results finished above zero after costs. It now says costs turned 35 of the 65 results that were positive before costs into losses, and that the 30 that stayed above zero did not meet all the criteria. No result changed. 9 Oct 2026, clock audit: a check against the US data releases at 08:30 New York showed the time stamps of the four currency files one hour early in most weeks of 2004 to 2014 and in some clock-change weeks after that. Two rules of this study depend on the time of day (the RSI(2) entry window and the Friday close), so all 96 variants were re-run with corrected time stamps; the comparison is in the new “Clock audit” section. No verdict changed; several figures quoted in this post moved, and the closest call, EURUSD A, passes 4 gates instead of 5 with the corrected clock. The figures above stay as the record of the published run. 8 Oct 2026, search and navigation: added an answer-first line, a FAQ and links to related tests and guides. No result changed. 22 Oct 2026, navigation: added a link to the new guide comparing trading costs across all our tests. No result changed. 22 Oct 2026, navigation: added a link to the new cross-market scoreboard, which includes this test. No result changed. 22 Oct 2026, navigation: added a link to the new guide comparing trend following and mean reversion across all our tests. No result changed. 22 Oct 2026, navigation: added a link to the new guide comparing versions with and without a stop loss across our tests. No result changed.

+More tests and guides

• RSI 2 strategy backtest
• Supertrend indicator backtest
• Bollinger bands strategy backtest
• Gold opening range breakout backtest
• Did the stop loss help? This test next to the other with-and-without pairs
• Trend following vs mean reversion: this test next to the others of its style
• 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