Nicolas Liebaert Français

The 2s5s10s butterfly: a small edge, mostly earned before 2010

A butterfly measures whether the middle of the yield curve is out of line with its two ends. I tested a simple rule on it over 36 years to find out if it has an edge after costs.

0.33 Sharpe ratio after costs no stop, 95% interval 0.03 to 0.60
0.17 With a fill one day later 95% interval -0.11 to 0.43
82% Winning trades 88 closed trades in 36 years
-39 bp Worst closed trade worst open loss -89 bp

Data to 3 November 2026.

What I did
The rule buys the 5-year Treasury and sells the 2-year and 10-year when the butterfly is far above its recent average. It does the opposite when the butterfly is far below. I ran it on 36 years of daily yields. Each day it uses only data known that day, and every trade pays a cost.
Why it matters to a trader
Rates desks and macro funds trade the shape of the yield curve as well as its level. A butterfly is sized so that a parallel move of the whole curve has almost no effect. It pays only if the middle moves against the ends.
What I found
The edge is small. The Sharpe ratio (average daily profit divided by its standard deviation, scaled to a year) is 0.33 after costs without a stop, and 82% of trades win. But the worst trade lost 8 times the average trade. With a fill one day later, about half of the butterfly’s Sharpe ratio and a third of the crack’s disappear. The rule worked less well after 2010. It also earned close to nothing in quiet markets and when the curve was inverted (short-term yields above long-term yields).
What I could not show
That anyone could have traded this in practice. I use the yields the US Federal Reserve publishes for each maturity, not the prices of bonds or futures. I did not model financing, carry (the income or cost of holding the position), exchange margin (the collateral a broker requires) or the cost of rolling from one contract to the next.

Try the rule on the butterfly

This demo reads the latest data: after 3 November 2026 its figures may differ from the text.

Stop
Fill

At these settings, a small edge: Sharpe ratio 0.26, and 86% of the profit came before 2010.

Cumulative profit of the rule, in basis points of the butterfly, after costs. The thin line is the other stop setting.
Sharpe ratio
0.26
after costs
Winning trades
71%
of 107 trades
Average trade
2.78 bp
in basis points, after costs
Worst closed trade
-33.0 bp
11.9 times the average trade
Worst open loss
-42.0 bp
before costs
Total profit
297 bp
in basis points, all years
Maximum drawdown
-88 bp
fall from a peak
Time in a trade
38%
of all days

Profit by year

Profit in each calendar year. Bars above zero made money, bars below it lost money.

Profit in each year, as a table
YearProfit

Results by period

PeriodDaysSharpeProfit
1991 to 1999 2,253 0.37 90 bp
2000 to 2009 2,501 0.44 165 bp
2010 to 2019 2,501 0.02 6 bp
2020 to 2026 1,711 0.18 37 bp
Data to 3 November 2026. The rule is a JavaScript copy of the code in the research repository, so it responds at once on a phone. It is checked against the Python by a test script. The exit level moves with the entry level: it is always one third of it.

How the trade works

A butterfly compares three points on the US Treasury yield curve: the 2-year, the 5-year and the 10-year. It equals 2 × the 5-year yield minus the 2-year yield minus the 10-year yield, in basis points (one basis point is 0.01 percentage point). It shows how far the middle of the curve sits above its two ends.

The logic: a high butterfly means the 5-year yield is high against the two ends, so the 5-year bond looks cheap. The trade buys it and sells the other two. The legs are sized so that a parallel move of the whole curve has almost no effect. The trade makes or loses money only if the middle moves against the ends.

Bond prices fall when yields rise. So when the rule sells the butterfly, measured in yield, it buys the belly: long the 5-year (the belly), short the 2-year and the 10-year (the wings). Each short leg carries half of the 5-year’s DV01, which is the change in dollars for a one basis point move in yield. One basis point of the butterfly is worth half the 5-year DV01: at $10,000, the average trade with the stop, 2.8 bp, is about $14,000 and the worst, -33 bp, about -$165,000.

The rule is the same as for the crack spread. The z-score counts how many standard deviations today’s spread is from its average over the last 250 trading days. The rule enters above 1.5 or below -1.5, and exits between -0.5 and 0.5. A stop at 2 standard deviations against the trade is optional. The relative value page explains the rule in full.

What the results say

A Sharpe ratio above 1 is good. Around 0.3 to 0.5 is small. The 95% confidence interval comes from a block bootstrap, explained on the relative value page. If it includes zero, the data cannot rule out a true Sharpe ratio of zero.

Data to 3 November 2026.

No stopWith stop
Sharpe ratio, after costs0.330.26
95% confidence interval0.03 to 0.60-0.01 to 0.55
Sharpe ratio, fill one day later0.170.14
95% confidence interval, fill one day later-0.11 to 0.43-0.13 to 0.39
Closed trades88107
Winning trades82%71%
Average trade5.0 bp2.8 bp
Worst closed trade-39 bp-33 bp
Worst open loss (before costs)-89 bp-42 bp
Traded years with a profit24 of 3621 of 36

Without a stop, a short opened on 18 May 2026 was still open on 3 November 2026, at -8.5 bp. It is left out of the trade figures.

With a stop, the 95% confidence interval includes zero, so the true edge may be zero. With a fill one day later it includes zero in both versions. A Sharpe ratio of 0.33 over 36 years is a t-statistic of about 2, below the hurdle of about 3 often asked of a rule found after many trials.

The stop lowers the profit and only trims the worst closed trade a little. What it cuts is the worst open loss: -89 bp without the stop, in 2010, and -42 bp with it, in 2014. A win rate means little without this line: the rule without a stop wins often because it sits through open losses two to three times the size of its worst closed trade. Many small gains and a few large losses are the usual shape of a mean-reversion trade (a bet that a gap will close). This is why position size matters more than the signal.

When could the rule work?

Each point is marked Tested if I measured it in this data. It is marked Idea if it is how the market works in theory and I did not test it. Each subset of the data is small, so treat the numbers with caution. The Sharpe ratios and the losing years below are for the rule with the stop.

The calm and busy groups are thirds of the whole history, so they use hindsight: nobody could have drawn those lines at the time. With lines drawn only from past days the butterfly’s Sharpe ratio is 0.08, -0.04 and 0.58 (calm, middle, busy). Nearly all of these slices have a 95% confidence interval that includes zero.

When it could work

  • Idea When the central bank is on hold and rates move in a range. With no large new driver, a 5-year yield that is out of line tends to come back.
  • Idea After a one-off push on a single maturity, such as a heavy 5-year auction or index buying at month end. The gap opens for a few days, then closes.
  • Tested When the curve has a normal shape and rates are moving. The Sharpe ratio was 0.48 when the 10-year was 0 to 100 basis points above the 2-year. It was 0.50 in the third of days when rates moved most.

When it can lose money

  • Tested When the curve is inverted (the 2-year yields more than the 10-year). The Sharpe ratio was -0.03. The same is true in quiet markets: -0.06. Costs are not the reason: before costs the Sharpe ratio is still only 0.05 on inverted curves and 0.04 in quiet markets. I do not know why.
  • Tested In many years. The rule lost money in 15 of 36 traded years (1991, 1995, 1998, 2000, 2001, 2004, 2010, 2013, 2014, 2016, 2019, 2021, 2022, 2024, 2026). I did not test the cause. My guess is a change in central bank policy, which reprices the whole curve.
  • Tested After 2010. The Sharpe ratio was 0.44 from 2000 to 2009, 0.02 from 2010 to 2019 and 0.18 since 2020. I do not know why. It may be that central bank bond buying kept the curve in place.
  • Tested One year does a lot of the work. 2008 is 31% of the profit with the stop. Without it the Sharpe ratio of the rule without a stop is 0.25, and its 95% confidence interval (-0.03 to 0.53) includes zero.

The latest position. The worked example applies the rule to the latest data and shows these situations next to it.

How much should you trust it?

Other settings. This table shows the Sharpe ratio of the rule without a stop for other windows and entry levels. Across the 12 settings it goes from 0.16 to 0.45.

Data to 3 November 2026.

Window \ entry level1.01.52.0
60 days0.450.400.42
120 days0.300.190.29
250 days0.450.330.16
500 days0.340.320.29

The outlined cell is the setting used on this page. The stronger the tint, the higher the Sharpe ratio.

Costs. With a cost of 0, 0.25, 0.5 and 1 basis point each time the rule trades, the Sharpe ratio (with stop) is 0.36, 0.31, 0.26, 0.17. It falls as the cost rises but stays above zero at these four levels. You can try other costs in the demo above.

The trade is built on yields, not on prices. The butterfly 2 × 5-year minus 2-year minus 10-year is the standard 50/50 DV01 butterfly. It is neutral to a parallel move only, and the curve rarely moves in parallel. The 5-year moves more than the average of the wings: when yields rise 1 bp on average, the butterfly rises about 0.23 bp. Level and slope together explain 23% of its daily variance (27% since 2010). A PCA-weighted butterfly would remove most of that. I did not do it, so each trade carries a small bet on rates. Over the whole sample this washes out: the rule’s profit has no measurable link to the direction of rates.

What I did not model. I did not test on bond or futures prices. The yields are constant-maturity yields. They are computed from the whole curve and are not prices you can trade. A real trade uses on-the-run bonds (the most recently issued) or futures. It has a bid-offer spread, financing, exchange margin and a contract roll every quarter. I did not model any of that. Carry and roll-down (the gain from a bond moving down the curve as it ages) are missing too, because a change in yield is not a total return. A rough estimate (T-bill financing, straight lines between the three yields) puts them between about -44 and +44 bp of the butterfly a year (10th and 90th percentiles, for a position held all year), against about 8 to 12 bp a year of net profit for the rule. Over the sample the missing amount looks positive, about +1 bp a trade, or about a quarter of the rule’s net profit. A position held all year could see several times that profit, so a real test needs it.

The futures sizes on the worked example are estimates. They use the duration of a par bond of the same maturity, not that of the cheapest-to-deliver bond (the bond a seller of the future would deliver).

An earlier version. An independent review showed that the conclusions of my first version of this study were wrong. I rebuilt it, and this page is the new version. The correction is on the relative value page.

Research only

This is research. The rule has never been traded, and nothing here is investment advice. The rule earned a small amount in the past, and it may not do so in the future.

Source and tests at github.com/Nicolas8330/relative-value-strategies. Every figure outside the demo comes from one run of examples/make_results.py, or from the rule itself, on yields to 3 November 2026. The one-off checks use data to 3 November 2026: the intervals with a later fill, the level and slope fit, carry, the groups drawn from past days and the test without 2008. See also the crack spread page.