Rolling Returns: Formula and Examples
Table of Contents
Rolling returns show how an investment’s compounded performance changes when
you move the starting date while keeping the holding period fixed. A strong return over the
latest three years may be one favorable observation in a much less consistent record.
Calculating every available three-year window makes that dependence on timing visible.
The method is straightforward: calculate one annualized return, shift the window forward
one month, and repeat. The interpretation requires more care. Nearby windows share most
of their returns, so a long list of results does not represent equally many independent
investment experiences. This guide uses a fully hypothetical ten-year record to show the
calculation, its distribution, and what you can reasonably conclude from it.
What Are Rolling Returns?
A rolling return measures performance over a fixed-length period that advances through
time. A monthly-updated three-year series calculates a return for each complete 36-month
window. A five-year series uses 60 months. Each observation answers: what compounded
annual rate would match the growth of an investment held over this window?
Every observation in a three-year rolling series measures three years of performance.
The start and end dates change together. The latest trailing three-year return is simply
the final observation in that same series.
“Rolling” describes how the measurement period moves. “Annualized” describes how a
multi-year result is expressed. Rolling returns can instead be reported cumulatively,
so check the label. This article reports compounded annualized total returns. For the
underlying annualization mechanics, see annualized return and CAGR.
Rolling Returns vs. Trailing and Calendar-Year Returns
These presentations can use the same source data while answering different questions.
A single trailing number is useful for the most recent experience; a rolling series adds
the other eligible starting dates.
| Presentation | What stays fixed? | Main question |
|---|---|---|
| Trailing three-year return | Three-year horizon and latest cutoff | What happened over the latest three years? |
| Rolling three-year returns | Three-year horizon; dates move | How did outcomes differ across eligible starting dates? |
| Calendar-year returns | January–December boundaries | What happened in each separate calendar year? |
| Since-inception CAGR | Original start date; horizon expands | What annual rate summarizes the entire record so far? |
A three-year CAGR and a five-year CAGR ending on the same date cover different
experiences. Annualizing both does not make their underlying windows identical. Likewise,
the average of many overlapping CAGRs is not the CAGR earned by one continuously held account.
The Rolling Return Formula
Let rt be the simple total return during month t,
expressed as a decimal. Construct a wealth index beginning at
W0 = 100, immediately before the first month’s return.
Each later level incorporates one more month:
For a window of L months ending at month t, compare the ending
level with the level exactly L months earlier:
Equivalently, compound the monthly growth factors from month t − L + 1
through month t, then raise that product to 12/L and subtract one.
The annualized rate summarizes the window; it does not assert that every year actually
earned that rate. Geometric linking and multi-year geometric annualization are consistent
with the performance-measurement principles explained in the GIPS handbook.[1]
Count Returns, Index Levels, and Windows Separately
With N complete monthly returns, you need N + 1 index levels,
including the initial level. If N is at least L, the first complete window ends at month L
and the last at month N. Moving one month at a time therefore gives:
For 120 monthly returns, there are 121 index levels, 85 three-year windows
(120 − 36 + 1), and 61 five-year windows (120 − 60 + 1).
If you instead have only 120 month-end index levels, their consecutive differences supply
119 monthly returns. Treating levels as returns creates an extra window that does not exist.
What Changes When the Window Moves?
Suppose the record begins with the January 2016 return, measured from December 31, 2015
to January 31, 2016. The first three-year window ends on December 31, 2018. Move both
boundaries forward one month for the next window, ending January 31, 2019. The table
separates the return that leaves, the returns that remain, and the return that enters.
| Window boundaries | January 2016 return | February 2016–December 2018 returns | January 2019 return |
|---|---|---|---|
| December 31, 2015–December 31, 2018 | Included: 1 month | Included: 35 months | Excluded |
| January 31, 2016–January 31, 2019 | Excluded | Included: 35 months | Included: 1 month |
Only one monthly growth factor leaves and one enters. The two annualized returns
are built from almost the same experience. Their 35 shared months explain why a rolling
line can look smooth even when the underlying monthly investment returns are volatile.
Worked Example: 120 Hypothetical Monthly Returns
Consider a synthetic investment with an initial total-return index of 100 on
December 31, 2015 and monthly observations through December 31, 2025. All income is
reinvested. Returns are nominal and USD-denominated, with no investor contributions,
withdrawals, investment fees, transaction costs, or taxes modeled.
The annual totals below are invented teaching inputs, not actual market returns
for those years. To specify all 120 monthly returns exactly, every month within
a given year earns the same rate: (1 + that year’s return)1/12 − 1.
This simplified pattern makes the example reproducible; actual monthly returns vary.
| Year | Hypothetical annual total return | Year-end wealth index |
|---|---|---|
| 2016 | 12% | 112.000 |
| 2017 | 8% | 120.960 |
| 2018 | −30% | 84.672 |
| 2019 | 25% | 105.840 |
| 2020 | 10% | 116.424 |
| 2021 | −20% | 93.139 |
| 2022 | 18% | 109.904 |
| 2023 | 6% | 116.499 |
| 2024 | −8% | 107.179 |
| 2025 | 14% | 122.184 |
Year-end index levels are rounded to three decimals; calculations retain full precision.
For example, the first three-year window compounds 2016–2018:
100 × 1.12 × 1.08 × 0.70 = 84.672. Its cumulative return is −15.328%,
and its annualized return is (84.672/100)1/3 − 1 = −5.40%.
Now calculate the same 36-month formula at every month-end from December 2018 through
December 2025. Each point in the next chart is one complete three-year holding window,
labeled by its ending month. The zero line separates windows with a
nominal gain from windows with a nominal loss.
| Window end (month) | Annualized return (%) |
|---|---|
| 2018-12 | -5.40 |
| 2019-01 | -5.11 |
| 2019-02 | -4.82 |
| 2019-03 | -4.53 |
| 2019-04 | -4.23 |
| 2019-05 | -3.94 |
| 2019-06 | -3.65 |
| 2019-07 | -3.35 |
| 2019-08 | -3.06 |
| 2019-09 | -2.76 |
| 2019-10 | -2.46 |
| 2019-11 | -2.17 |
| 2019-12 | -1.87 |
| 2020-01 | -1.82 |
| 2020-02 | -1.77 |
| 2020-03 | -1.72 |
| 2020-04 | -1.67 |
| 2020-05 | -1.62 |
| 2020-06 | -1.57 |
| 2020-07 | -1.52 |
| 2020-08 | -1.47 |
| 2020-09 | -1.42 |
| 2020-10 | -1.37 |
| 2020-11 | -1.32 |
| 2020-12 | -1.27 |
| 2021-01 | -0.90 |
| 2021-02 | -0.53 |
| 2021-03 | -0.16 |
| 2021-04 | 0.21 |
| 2021-05 | 0.58 |
| 2021-06 | 0.96 |
| 2021-07 | 1.33 |
| 2021-08 | 1.71 |
| 2021-09 | 2.09 |
| 2021-10 | 2.47 |
| 2021-11 | 2.85 |
| 2021-12 | 3.23 |
| 2022-01 | 3.06 |
| 2022-02 | 2.90 |
| 2022-03 | 2.73 |
| 2022-04 | 2.57 |
| 2022-05 | 2.41 |
| 2022-06 | 2.24 |
| 2022-07 | 2.08 |
| 2022-08 | 1.91 |
| 2022-09 | 1.75 |
| 2022-10 | 1.59 |
| 2022-11 | 1.43 |
| 2022-12 | 1.26 |
| 2023-01 | 1.16 |
| 2023-02 | 1.06 |
| 2023-03 | 0.95 |
| 2023-04 | 0.85 |
| 2023-05 | 0.74 |
| 2023-06 | 0.64 |
| 2023-07 | 0.54 |
| 2023-08 | 0.43 |
| 2023-09 | 0.33 |
| 2023-10 | 0.23 |
| 2023-11 | 0.12 |
| 2023-12 | 0.02 |
| 2024-01 | 0.41 |
| 2024-02 | 0.80 |
| 2024-03 | 1.19 |
| 2024-04 | 1.59 |
| 2024-05 | 1.98 |
| 2024-06 | 2.38 |
| 2024-07 | 2.78 |
| 2024-08 | 3.18 |
| 2024-09 | 3.58 |
| 2024-10 | 3.98 |
| 2024-11 | 4.39 |
| 2024-12 | 4.79 |
| 2025-01 | 4.69 |
| 2025-02 | 4.59 |
| 2025-03 | 4.49 |
| 2025-04 | 4.39 |
| 2025-05 | 4.29 |
| 2025-06 | 4.19 |
| 2025-07 | 4.09 |
| 2025-08 | 3.99 |
| 2025-09 | 3.89 |
| 2025-10 | 3.79 |
| 2025-11 | 3.69 |
| 2025-12 | 3.59 |
Hypothetical; all 85 complete 36-month windows from the specified 2016–2025 monthly series. Returns rounded to two decimals; no fees, taxes, or external cash flows.
The observed annualized outcomes span about 10.19 percentage points,
despite every point using the same three-year horizon. The final 3.59% trailing return
therefore tells only part of this synthetic record’s story. A higher line means a better
return over the preceding three years; it does not mean the investment earned that rate
during the month printed below it.
Read the Distribution Alongside the Latest Trailing Return
A time-series chart preserves the sequence of window outcomes. A frequency chart asks
how those outcomes are distributed. For this comparison, use all 85 three-year
windows and all 61 five-year windows. Group the unrounded annualized rates into
the same four bands; each band includes its lower boundary and excludes its upper boundary.
For example, 0% belongs in “0 to <5,” while a negative return belongs below zero.
Each series is expressed as a percentage of its own window count, allowing comparison
despite the different denominators. These are distributions of overlapping observations
in the hypothetical record, rather than estimates of future investment probabilities.
| Annualized return band (%) | 3-year | 5-year |
|---|---|---|
| −10 to <−5 | 2.35 | 0.00 |
| −5 to <0 | 30.59 | 34.43 |
| 0 to <5 | 67.06 | 57.38 |
| 5 to <10 | 0.00 | 8.20 |
Hypothetical; shares use 85 three-year and 61 five-year windows. Bands classify unrounded returns; displayed shares may not sum to exactly 100% after rounding.
Both latest returns—3.59% for three years and 0.97% for five years—sit in the 0% to
below-5% band. Yet 28/85 = 32.94% of three-year windows and
21/61 = 34.43% of five-year windows have negative annualized returns.
The positive latest observations conceal those loss windows if shown alone.
Compare the Range, Middle, and Final Observation
The following summary retains exact dates for the extreme windows and distinguishes
the sample median from the latest observation. A median is the middle ordered window
return; it is not a return earned by holding through the full ten-year record.
| Measure | Three-year windows | Five-year windows |
|---|---|---|
| Number of windows | 85 | 61 |
| First window end | December 31, 2018 | December 31, 2020 |
| Worst annualized return | −5.40%; December 31, 2015–December 31, 2018 | −3.62%; December 31, 2016–December 31, 2021 |
| Median annualized return | 0.96% | 0.67% |
| Best annualized return | 4.79%; December 31, 2021–December 31, 2024 | 6.59%; December 31, 2018–December 31, 2023 |
| Latest trailing annualized return | 3.59%; December 31, 2022–December 31, 2025 | 0.97%; December 31, 2020–December 31, 2025 |
The latest three-year cumulative growth factor is
1.06 × 0.92 × 1.14 = 1.111728, so
1.1117281/3 − 1 = 3.59% annualized.
For the latest five years, 0.80 × 1.18 × 1.06 × 0.92 × 1.14
= 1.049471232, giving 1.0494712321/5 − 1 = 0.97% annualized.
Both calculations end on the same date, but the five-year window also includes the
earlier loss and recovery.
Notice that the five-year series does not have a smaller loss-window share in this
example. Longer horizons do not mechanically ensure better outcomes. These full-sample
distributions also have different eligible end dates: the five-year record begins two
years later. To isolate a comparison across horizons more carefully, repeat the analysis
using common end dates and report that restriction.
Why Overlapping Windows Are Not Independent Trials
One difficult market episode can affect many consecutive rolling results. In a
36-month series, adjacent windows share 35 monthly returns; five-year windows share
59 of 60. Adding another plotted observation usually adds just one new month’s data.
It does not add another wholly separate three-year or five-year history.
This matters when moving from description to statistical inference. Treating all
windows as independent can overstate the precision of an estimated average or loss
frequency. Research on long-horizon predictive regressions explicitly develops methods
to address overlap in inference; those methods depend on the statistical problem and
assumptions.[2]
A descriptive rolling-return chart does not supply those assumptions or a forecast model.
“28 of 85 windows lost money” describes the specified sample. It does not establish
a 32.94% probability of losing money over the next three years. Overlap, the selected
history, and changing investment conditions all matter; this article’s invented
record supplies no empirical forecast evidence.
Non-overlapping windows offer a useful sensitivity check. Starting at the beginning
of a 120-month record gives only three disjoint 36-month windows, with 12 months left
over, or two disjoint 60-month windows. That illustrates how little separate long-horizon
history ten years contain. Disjoint periods still need not be statistically independent,
and shifting their calendar alignment can change the observed results.
Do not attach an independent-binomial confidence interval to these overlapping loss
counts or assume the standard error of the mean equals the sample standard deviation
divided by √85. Nor can you replace 85 with a guessed “effective sample size.” Formal
uncertainty estimates require a method that addresses the actual time dependence and
question being studied.
How to Prepare Data and Compare Investments Fairly
The formula cannot repair inconsistent source data. Establish the return definition
before ranking funds or comparing a portfolio with a benchmark.
- Match the return basis. Use documented total-return data with a
consistent reinvestment convention. Total return includes income as well as changes
in asset value.[1]
Do not compare a dividend-reinvested fund series with a price-only index. Confirm what
a vendor’s “adjusted” price actually adjusts. - Handle external cash flows first. Raw account balances rise when
you deposit money, even with no investment gain. For strategy performance, build an
appropriate cash-flow-adjusted return series before linking windows. For the experience
of your own invested dollars, consider the different question addressed by
time-weighted versus money-weighted returns. - Align currency, fees, and dates. Compare the same share class,
reporting currency, gross/net basis, and month-end convention. Check whether a fund
series already deducts expenses before subtracting them again. SEC investor guidance
highlights fees, methodology, and benchmark selection when evaluating performance
claims.[3] - Keep months consecutive. Investigate missing or duplicate
observations. Do not silently replace a missing return with zero or compress 36
available rows from more than 36 calendar months into a “three-year” window.
Obtain the missing observation or omit affected windows and disclose the gap. - Handle total losses and nonpositive levels explicitly. A window
from a positive index level to zero records a complete loss; one starting at zero
has no valid percentage-return denominator. Negative levels require a different
performance framework. Do not silently omit a failed investment to preserve a
positive-index sample. - Use common windows for rankings. A new fund and an older fund
should be compared on identical eligible start/end dates for the main comparison.
Present the older fund’s longer record separately. Within each matched window,
subtracting benchmark CAGR from fund CAGR gives a percentage-point gap; it is not
automatically a compounded relative-wealth return.
For one window, the annualized return calculator
can help check the endpoint arithmetic. A full rolling analysis still requires repeating
that calculation over every eligible pair of dates. No separate rolling-return calculator
or particular software capability is assumed here.
Rolling returns measure outcomes, rather than all dimensions of risk. Pair them with
maximum drawdown to examine interim losses and
standard deviation to examine return variability.
If the question is consistency of active performance relative to a benchmark,
the information ratio addresses a different task.
Common Mistakes with Rolling Returns
1. Changing the horizon while comparing starting dates. A three-year
window and a five-year window answer different questions. Keep the horizon fixed within
a series, and label any comparison across horizons explicitly.
2. Averaging monthly returns instead of compounding them. Additive
averages do not reproduce realized multi-month wealth growth. Use the product of monthly
growth factors or the ratio of a correctly constructed total-return wealth index.
3. Calling 85 windows 85 independent tests. Most observations repeat
the same underlying months. More frequent window updates create a denser description,
without manufacturing more independent long-horizon history.
4. Treating smooth rolling returns as low monthly risk. Long windows
can dilute one difficult month among many other observations. They also stay similar
because most inputs are shared. Inspect the investment’s actual path and drawdowns.
5. Reporting only the best window or latest favorable cutoff. Show
the full stated sample, the latest observation, and the relevant adverse periods.
Do not present a best starting date identified afterward as a rule you could have
followed in advance.
6. Treating every moving-window calculation as a rolling holding return.
A rolling regression in cross hedging estimates a relationship;
rolling factor analysis in hedge fund risk management
monitors exposures. A moving-average signal in managed futures
can help define a trading rule. Those moving samples have different outputs and purposes.
Limitations: What the Distribution Leaves Out
The sample bounds the conclusions. Showing every eligible window
within a selected decade does not eliminate selection of the decade itself. A record
beginning after a severe loss or ending before a difficult period can still flatter the
investment. The SEC cautions against selectively presented performance periods and
against projecting past results into the future.[3]
The worst observed window is not the worst possible outcome.
Future losses can exceed the sample minimum. A longer horizon also creates fewer
complete windows from a fixed history. Neither a narrower observed range nor an absence
of sampled losses proves that holding longer eliminates risk.
The underlying universe can be biased. A comparison restricted to
funds that still exist excludes funds that closed. A database admitting a strategy only
after a strong early record can also be selective. Rolling arithmetic cannot reconstruct
the missing investments or performance history; investigate inclusion rules and gaps.
Annualized endpoints omit the journey. A positive three-year result
can include a large temporary loss, a long recovery, and poor liquidity. It also does
not measure the outcome of a contribution or withdrawal plan. Positive nominal returns
need not preserve purchasing power after inflation, fees, and taxes.
A descriptive pattern is not a timing rule. You know a window’s
completed return only at its end. Using it to design a trading strategy requires a separate
decision rule, information available at the time, implementation assumptions, and testing.
The distribution alone establishes neither a successful entry signal nor manager skill.
A Practical Rolling-Return Checklist
A useful report states the horizon, update frequency, source coverage, total-return
convention, and fee/currency basis. Then show the sequence or distribution, identify the
latest trailing observation, and retain the dates of the best and worst complete windows.
For a decision, investigate why outcomes differed across windows, compare investments
on common dates, and check interim losses alongside terminal growth. Report negative-window
fractions as sample descriptions with their denominators. Keep any forecast or statistical
test separate and explain its additional assumptions.
Frequently Asked Questions
Each annualized rolling observation is a CAGR for its own window, using a suitable
total-return index or no-cash-flow investment values. Rolling analysis repeats that
calculation across start/end dates. A single since-inception CAGR summarizes one
entire record; a rolling series summarizes many equal-length portions of it.
Ten complete years supply 120 monthly returns. The first complete 60-month window
ends at month 60, and the last at month 120. Counting both endpoints gives
120 − 60 + 1 = 61. You need 121 wealth-index observations, including the level before
the first month’s return, to construct that record.
For total-return performance, include investment income and use a documented
reinvestment convention. A price-only return is a different measurement and should
be labeled accordingly. Match the fund and benchmark conventions rather than
assuming every adjusted-price field includes the same distributions or adjustments.
It describes the chosen record and sampling convention. It does not, by itself,
estimate the probability of a future gain. Overlapping windows reuse months, and
future conditions may differ from those in the sample. Even a record with no losing
windows cannot establish that a holding period is loss-free.
Use the horizon that answers the question, and show more than one when useful.
A five-year measure covers more months per observation but supplies fewer complete
observations from a fixed history. It can also include older conditions that differ
from the latest three years. Compare common end dates before interpreting a
difference between the two full-sample distributions.
An unadjusted balance ratio works for investment growth only when external
deposits or withdrawals do not distort it. Otherwise, obtain or calculate an
appropriate cash-flow-adjusted performance series first. Time-weighted return
evaluates the investment strategy separately from external cash-flow timing;
money-weighted return measures a different investor experience.
Sources and Methodology
- CFA Institute, GIPS Standards Handbook for Firms, November 2020; provisions 2.A.8, 2.A.24, and 4.B.2. Performance-measurement reference; no claim of GIPS compliance.
- Erik Hjalmarsson, New Methods for Inference in Long-Horizon Regressions, Journal of Financial and Quantitative Analysis 46(3), June 2011, pp. 815–839; published abstract and the accessible earlier working-paper introduction informed the overlap discussion.
- SEC Office of Investor Education and Advocacy, Investor Bulletin: Performance Claims, September 15, 2022; staff investor education.
Sources accessed November 11, 2026. All example returns, wealth levels, rolling
observations, and frequency counts are author calculations from the fully specified
hypothetical series above. Calculations use unrounded values before presentation rounding;
no security, fund, index, or provider’s actual performance is represented.
Disclaimer
This article is for educational and informational purposes only and does not
constitute investment advice. Hypothetical examples illustrate measurement mechanics;
they are not forecasts or actual investment performance. Historical outcomes do not
guarantee future results. Consider your circumstances and consult a qualified
financial professional before making investment decisions.
Keep reading
More in Portfolio Management
-
Investment Policy Statement & RRTTLLU Framework: Complete Guide
Learn what an investment policy statement (IPS) is, how to write one using the RRTTLLU framework, and why it is the cornerstone of the portfolio management…
-
Systematic vs Unsystematic Risk: What’s the Difference?
Learn the difference between systematic (market) risk and unsystematic (company-specific) risk, why diversification eliminates only one type…
-
Annualized Return (CAGR): Formula, Calculation, and Examples
Learn how to calculate annualized returns (CAGR) and understand the critical difference between geometric mean and arithmetic mean for measuring multi-year…
