Do Technical Trading Rules Work in FX? Fifty Years of Evidence, and a Test on Fed Data

Simple technical trading rules, such as buying a currency when its price crosses above a moving average, did earn statistically significant profits in the major dollar exchange rates in the 1970s and 1980s, and those profits survived realistic transaction costs. Since the early 1990s the same rules have earned roughly nothing in the major currencies, which is why a chart-based forex signal today needs far stronger evidence than “this pattern has worked before.”

In short: the academic record on technical rules in FX is unusually clear. Richard Levich and Lee Thomas (1993) found that filter and moving-average rules on currency futures from 1976 to 1990 produced profits that ranked in the top 1% of thousands of randomly reshuffled price series. Christopher Neely, Paul Weller and Joshua Ulrich (2009) then tested those exact rules on the following 15 years and found the profits had disappeared. Our own test on Federal Reserve data, net of costs, reproduces both halves: a 5-day/20-day moving-average rule earned 8.0% a year across five major currencies from 1976 to 1990, then 0.35% a year from 1991 to 2005 and minus 1.15% a year from 2006 to November 2026. More complex rules and emerging-market currencies have done better, but the simple patterns sold in most signal channels belong to a market that no longer exists.

The distinction between chart-based and news-based signals is covered in technical vs fundamental signals. This piece is about evidence: what the papers tested, what they found, and what happens when you rerun the classic rules yourself.

The misconception: “if it worked in the past, the pattern is real”

Most technical signal marketing rests on a backtest or a hand-picked chart where a crossover or breakout “called” a big move. The academic literature takes the question seriously, because the stakes are high: if mechanical rules applied to public price data earn risk-adjusted profits, the market is not efficient in even its weakest form.

Two separate questions get confused:

Was the profit real in the sample? Or was it the product of trying many rules and reporting the winners (data mining)?
Does the profit persist after it is discovered? A real inefficiency can still be competed away once enough traders exploit it.

The FX evidence answers yes to the first question for the 1970s and 1980s, and mostly no to the second. Both answers matter for anyone paying for a technical signal today. (The general problem of searching many rules and keeping the best is the subject of overfitting and data snooping.)

The rules being tested

The literature focuses on a handful of rules simple enough to have been in common use before anyone wrote about them:

Moving-average rule, MA(S, L): compute a short moving average of the last S daily prices and a long one of the last L prices. Be long the foreign currency when the short average is above the long one, short otherwise. MA(1, 200) compares today’s price with its 200-day average; MA(5, 20) compares a one-week average with a one-month average.
Filter rule (x%): buy when the currency rises x% above its most recent trough, sell and go short when it falls x% below its most recent peak.
Channel (breakout) rule: go long when the price exceeds its highest level of the previous L days, short when it falls below the lowest.

All of these are trend-following rules. They make money when exchange rates move in persistent runs and lose money in choppy, mean-reverting markets, where each false signal costs a spread.

A correct test measures the excess return: the change in the exchange rate plus the interest-rate differential earned or paid while holding the position, minus transaction costs every time the rule switches sides. Ignoring the interest differential overstates or understates results, because holding a high-yielding currency earns interest and holding a low-yielding one costs it.

The early evidence: Levich and Thomas (1993)

Richard Levich of NYU and Lee Thomas, “The significance of technical trading-rule profits in the foreign exchange market: a bootstrap approach,” Journal of International Money and Finance 12 (1993), circulated as NBER Working Paper 3818 (1991). Their design is the reason the paper is still cited.

Data: daily settlement prices of Chicago Mercantile Exchange futures for five currencies (British pound, Canadian dollar, German mark, Japanese yen and Swiss franc), January 1, 1976 to December 31, 1990, about 3,800 observations each. Using futures automatically builds the interest differential into the return.
Rules: filter rules from 0.5% to 5% and moving-average rules MA(1, 5), MA(5, 20) and MA(1, 200).
Test: they reshuffled each currency’s daily price changes 10,000 times, which destroys any trend structure while keeping the same distribution of daily moves, then ran every rule on every shuffled series. If a rule’s profit on the real data ranks above 99% of the shuffled series, the trends it exploited are very unlikely to be luck.

The results were striking. Across 45 combinations of currency and rule, the paper reports that “in 31 cases the original exchange rate series produced profits in the top 1% of all times series,” eight more were in the top 5%, and the remaining six were positive but not significant. Average moving-average profits were 7.0% to 9.0% a year for four of the currencies and 2.7% for the Canadian dollar.

The paper also did two things that marketing backtests rarely do. It estimated transaction costs for a large institution in the futures market at about 2.5 basis points per transaction (4 basis points as a conservative figure) and showed that for the fastest rules, trading about 65 times a year, costs would cut profits by 1.62% to 2.60% a year, enough to “nearly decimate” the 3.3% return of MA(1, 5) in the Canadian dollar. And it split the sample into three five-year periods, finding that profitability “declined in the 1986-1990 period” although it remained positive and often significant.

The out-of-sample verdict: Neely, Weller and Ulrich (2009)

Christopher Neely (Federal Reserve Bank of St. Louis), Paul Weller (University of Iowa) and Joshua Ulrich, “The Adaptive Markets Hypothesis: Evidence from the Foreign Exchange Market,” Journal of Financial and Quantitative Analysis 44, no. 2 (2009), pp. 467-488. The figures below are from the St. Louis Fed working-paper version (2006-046, revised March 2007), which uses data through June 2005.

Their method is the cleanest possible out-of-sample test: take rules that earlier papers had already published as profitable, and run exactly those rules on the data that came after the original sample ended. Nothing is chosen with hindsight.

They also took transaction costs seriously. Citing the earlier literature, they note that costs for a switch from long to short fell from about 10 basis points in the 1970s to about 2 basis points by the mid-2000s, as electronic trading spread, with spot spreads of “2 basis points or less, for transactions in the $5 to $50 million range.” They model this as a linear decline from 10 basis points per switch on January 1, 1973 to 1.88 basis points on June 30, 2005.

For the Levich and Thomas rules, the reversal is stark:

Levich-Thomas rules, replicated by Neely, Weller and Ulrich	Original sample, 1976-1990	Ex post sample, 1991 to June 2005
1% filter, net return, average of DEM, JPY, CHF, GBP	8.47% a year	-1.14% a year
MA(5, 20), the best MA rule in sample, net return	9.09% a year	MA rules fell 7 to 10 points, mostly negative
Average Sharpe ratio, all rules and currencies	0.63	-0.01
Uniform portfolio rule, mean Sharpe ratio across currencies	0.90	-0.02

In the paper’s words, “A reliably profitable strategy over the original time period would have lost money over the 1991-2005:6 period.” The MA(1, 5) rule did not earn a positive return for any currency in the later period, and no moving-average rule was positive for more than two of the four currencies.

The authors then asked why. If the original profits were pure data mining, they should have vanished abruptly at the end of each original sample. Instead, they found “little support for the existence of a mean break at the end of the original sample period, other than for the Levich and Thomas (1993) study,” and concluded that “the excess returns of the 1970s and 1980s were genuine.” The profits faded gradually as markets learned. Returns to less-studied rules (channel rules, ARIMA rules, genetic-programming rules and a Markov-switching model) also declined but “have probably not completely disappeared.”

That pattern is what Andrew Lo’s adaptive markets hypothesis predicts: profit opportunities exist, competition erodes them, and more complex strategies survive longer than simple ones. It does not fit a strict efficient-markets view, under which the early profits should never have existed.

Dennis Olson and the decline over time

Dennis Olson, “Have trading rule profits in the currency markets declined over time?” Journal of Banking and Finance 28 (2004), pp. 85-105, attacked the same question from a different angle, using 18 exchange-rate series. According to the paper’s abstract, he optimized moving-average rules over successive five-year in-sample windows starting in 1971 and tested each choice on the following five years, through 2000, and found that risk-adjusted profits fell from an average of over 3% a year in the late 1970s and early 1980s to about zero in the 1990s. Neely, Weller and Ulrich summarize the result the same way: returns “declined from the 1970s to about zero in the 1990s.” In their account, he suggested the earlier profits reflected a temporary inefficiency and concluded that “To ‘beat’ the currency market in the future may require more complicated trading rules, which in turn may represent temporary inefficiencies that will be eliminated once they are identified.” (We quote that sentence from Neely, Weller and Ulrich’s paper rather than from Olson’s original.)

Where trend returns went: momentum and emerging markets

Two later papers keep the story honest, because they show the effect did not vanish everywhere.

Lukas Menkhoff, Lucio Sarno, Maik Schmeling and Andreas Schrimpf, “Currency momentum strategies,” Journal of Financial Economics 106 (2012), pp. 660-684. Using up to 48 currencies from January 1976 to January 2010, they found a cross-sectional spread of “up to 10% per annum” between past winner and past loser currencies. Three findings matter for signal buyers:

Costs bite hard. Applying the full quoted bid-ask spread, the best momentum strategy’s return fell “from nearly 10% to about 4% p.a.,” and costs wiped out most of the profit of many other variants, partly because momentum portfolios tilt toward currencies with high transaction costs.
The profits sit in risky corners. Returns were about 8% a year in currencies with high idiosyncratic volatility versus about 4% in low-volatility ones, and were concentrated in countries with high country-risk ratings. The authors conclude that “very effective limits to arbitrage” stop momentum from being easily exploited.
It is not the classic moving-average rule. Benchmark MA(1, 20), MA(5, 20) and MA(1, 200) rules across their whole currency sample earned mean excess returns around 5% a year with Sharpe ratios of 0.77 to 0.88, but regressions of momentum returns on those rules explained at most 26% of the variation. Cross-sectional momentum ranks currencies against each other and buys relative winners, while a moving-average rule looks at each currency’s own past in isolation. They are different strategies.

Neely and Weller, “Lessons from the Evolution of Foreign Exchange Trading Strategies,” Journal of Banking and Finance (2013). Simulating a trader who re-selects among technical rules and the carry trade, they find returns “dip significantly in the 1990s but recover by the end of the decade,” and that “trading rule returns still exist in forex markets… though they have migrated to emerging markets to a considerable degree.” Citing Burnside and coauthors (2007), they also note that emerging-market bid-ask spreads over 1997-2006 were two to four times those of developed currencies, so an emerging-market edge is expensive to harvest.

Worked example: the classic rules on Federal Reserve data

We reran the Levich and Thomas moving-average rules ourselves, in Python, on primary data:

Exchange rates: the Federal Reserve’s H.10 daily noon buying rates in New York (via FRED), for the German mark spliced into the euro in January 1999 at the official 1.95583 conversion rate, the yen, the pound, the Swiss franc and the Canadian dollar, January 1976 to November 6, 2026.
Interest differentials: OECD monthly short-term rates (via FRED), using the previous month’s average so the test never uses information it could not have had. For the yen before July 1985 the only long series available is the Bank of Japan discount rate, used as a proxy.
Costs: the Neely, Weller and Ulrich schedule, 10 basis points per switch in 1973 falling linearly to 1.88 basis points in mid-2005, held there afterward.
Rule mechanics: the signal is computed from the day’s fixing and the position is held from that fixing to the next. Results are an equally weighted portfolio of the five currencies, unlevered.

Rule (switches per year per currency)	1976-1990 net	1991-2005 net	2006 to Nov 2026 net	Full period net
MA(1, 5) (about 65)	4.53% (t = 2.6)	-2.33%	-1.58%	0.00%
MA(5, 20) (about 15)	8.00% (t = 4.4)	0.35%	-1.15%	1.99%
MA(1, 200) (about 8)	7.04% (t = 3.9)	0.89%	0.12%	2.39%

Annual excess returns, simple averages of daily returns. The Sharpe ratios of MA(5, 20) and MA(1, 200) were 1.14 and 1.02 in 1976-1990, then 0.06 and 0.14 in 1991-2005, and -0.20 and 0.02 since 2006.

The numbers line up closely with the published work: our 1976-1990 MA(5, 20) figure of 8.0% compares with 9.09% in the Neely, Weller and Ulrich replication, and the collapse after 1990 is the same. By decade, MA(5, 20) earned 7.7% a year in the 1980s, 1.1% in the 1990s, -0.4% in the 2000s, -2.5% in the 2010s and 0.9% from 2020 to November 2026. MA(1, 200) went 6.9%, 1.9%, 1.7%, -0.2% and -1.3% over the same decades.

Three further checks tell you how fragile the fast rules are:

A one-day delay. If MA(1, 5) trades one day after its signal instead of at the signal’s own fixing, its 1976-1990 return falls from 4.53% to 1.84% a year and is no longer statistically significant. The slower rules barely change (MA(5, 20) goes from 8.00% to 7.35%). Any fast rule whose edge depends on executing at exactly the price that generated the signal is fragile, and a signal relayed through a Telegram channel always arrives late.
Costs relative to gross profit. Over the full sample, costs took 2.57 points a year from MA(1, 5), its entire gross return, against 0.57 points for MA(5, 20) and 0.27 for MA(1, 200).
Retail-style costs. Charging 3 basis points per switch plus a 1% a year financing markup on the position (an illustrative retail assumption, since overnight financing at retail brokers is usually priced at a markup to interbank rates) turns the 2006-2026 results into -3.3% for MA(1, 5), -2.3% for MA(5, 20) and -1.0% for MA(1, 200). Cost mechanics are worked through in transaction costs in retail FX.

What this test does not show: it uses one fixing a day, so it cannot speak to intraday rules; it ignores risk management overlays; and five currencies are a narrow universe. It does show, on public primary data that anyone can download, that the textbook rules taught in most trading courses have had no edge in the major dollar pairs for more than 30 years.

Why the profits faded

The literature offers several explanations, and they are not mutually exclusive.

Learning and competition. This is the adaptive-markets reading. Once rules were published and computing became cheap, more capital chased the same trends, and the trends got shorter.
Lower costs, paradoxically. Cheaper trading should make rules more profitable net of costs, yet profits fell anyway. That is strong evidence that the gross edge itself shrank.
Not risk, and not intervention. Neely and Weller (2013) summarize the earlier research as finding that explanations based on risk adjustment or central bank intervention “do not plausibly justify the observed excess returns produced by simple technical trading rules.” Whatever drove the early profits, it was not obvious compensation for risk.
Data mining in part. Neely, Weller and Ulrich found little evidence of it for most rules, but the Levich and Thomas rules did show a break at the end of their sample, which is consistent with some selection of the rules that happened to work best.
Migration. Where trends persist, as in some emerging-market currencies, Neely and Weller (2013) and Menkhoff and coauthors show the edge sits alongside higher costs, wider spreads and crash-prone currencies.

What this means for a technical forex signal

A signal service sells you, in effect, a trading rule. The evidence suggests four questions before you trust one.

Is the rule simple and well known? Moving-average crossovers, RSI thresholds and breakout rules on EUR/USD or GBP/USD are close cousins of the rules tested above. The evidence says their edge in major pairs disappeared around 1990.
Is the evidence in sample? A backtest that chose its parameters on the same data it reports is in-sample. The relevant comparison is the drop from the Levich and Thomas sample to the Neely, Weller and Ulrich ex post sample: from about 8% a year to below zero. See backtests vs live results.
Does it trade fast? Our MA(1, 5) test shows that a rule trading 65 times a year lost its whole gross return to institutional-level costs. A retail scalping signal trading hundreds of times a year faces costs several times higher.
Is the track record long enough? Our five-currency rule portfolios ran at about 6% annual volatility. At that volatility, a true 2% a year edge needs roughly 36 years of data to reach a t-statistic of 2, so a few months of results cannot separate skill from noise. The arithmetic is in statistical significance of a track record.

What can go wrong when you test this yourself

Leaving out interest. Spot-only backtests ignore the interest differential, which can be several percent a year in a high-yield pair and is correlated with the rule’s position.
Look-ahead in the data. Using a month’s interest rate before it was published, or the closing price that generated the signal as the execution price for a rule that cannot actually trade there, quietly inflates results.
Ignoring the switch cost. Each change from long to short crosses the spread on twice the position size. Fast rules switch dozens of times a year.
Picking the winner. Running 50 parameter pairs and reporting the best one is data snooping. The honest number is the out-of-sample performance of a rule chosen in advance.
Treating FX as one market. Results in the yen, the pound and emerging currencies differ widely. A rule that worked in one pair proves little about another.
Assuming the past regime returns. The 1976-1990 Sharpe ratios above 1.0 came from a market with wider spreads, fewer algorithmic participants and frequent central bank intervention.

The short version

Simple trend rules earned real, statistically significant profits in major currencies in the 1970s and 1980s: Levich and Thomas found profits in the top 1% of randomized series for 31 of 45 rule and currency combinations from 1976 to 1990. Neely, Weller and Ulrich then showed that the same rules lost money from 1991 to 2005, with an average Sharpe ratio of -0.01, and Olson found trading-rule returns declining to about zero by the 1990s. Our own net-of-cost test on Federal Reserve H.10 data reproduces the collapse and shows it continuing to November 2026. Trend returns survive in less-studied rules, cross-sectional momentum and emerging markets, but with higher costs and harder limits to arbitrage. This explains the evidence; it is not a recommendation to use or avoid any strategy.