The Bartels Test: Statistical Significance for Cycle Analysis

Julius Bartels was a German geophysicist who spent much of his career at the Carnegie Institution in Washington. In the 1930s he developed a statistical test for evaluating whether a claimed periodicity in a time series was real or a chance pattern in noisy data.

The Bartels test, sometimes called the Bartels cycle significance test, became standard in geophysics and solar-terrestrial physics. It's now the reference method for anyone doing serious cycle analysis in climate data, solar activity records, or financial time series.

The problem it solves

Cycle analysis without statistical testing is nearly useless. Any sufficiently noisy time series will show apparent periodicities at many frequencies if you look hard enough. The question is always: is this cycle real, or did I find it by looking at enough possibilities?

This is the multiple testing problem. If you test 100 frequencies and accept a 5% false positive rate, you expect 5 spurious "discoveries" even if the data is pure noise.

Bartels developed his test specifically because geophysicists of his era were drowning in claimed cycles, 11-year, 27-day, 35-year, various planetary periods, with no rigorous way to separate the real ones from the artefacts of wishful analysis.

How the Bartels test works

The test uses a specific structure. You take a time series and fold it over the candidate period. That is: you stack the data in rows of length P (the candidate period in days, months, or years), with one row per cycle.

You then compute the variance of the row means against the total variance. If the period P is real, if the data genuinely oscillates at that frequency, the row means will vary more than chance predicts. If the period is noise, the row means will look like random samples from the overall distribution.

The test statistic has a known distribution under the null hypothesis (no real cycle). You compute the probability that the observed row-mean variance would occur by chance. That probability is the Bartels significance level for the candidate period.

A significance level below 0.05 is the conventional threshold for "probably real." Below 0.01 is strong evidence. The test is two-tailed: it will also detect anti-cycles (data consistently below average at phase 0.5 and above at phase 0.0) that some analysts miss with simpler methods.

Application to financial cycle analysis

The Bartels test was designed for geophysical data but it applies directly to financial time series.

If you want to test whether the stock market has a genuine 40-week cycle, fold the price series into 40-week segments, compute the row-mean variance, and compare to the Bartels null distribution. If the cycle is real, you should see meaningful variation across the 40-week phase. If it's noise, the row means will be indistinguishable from random.

This is more rigorous than looking at a spectral plot and noticing a peak. Spectral peaks can appear from data-snooping even in random series. The Bartels test explicitly controls for this by computing the probability under the null hypothesis.

The cycle lengths most commonly tested in financial analysis: - 40-week (Kitchin inventory cycle) - 4-year (presidential cycle) - 9.2-year - 18.6-year (lunar nodal cycle) - Seasonal patterns within years

Crohamhurst.app includes Bartels testing for user-uploaded time series as part of its cycle analysis toolkit.

Application to rainfall and climate cycles

Bartels himself applied his test extensively to rainfall data, solar activity records, and geomagnetic indices. The most significant findings from that work:

• The 11-year sunspot cycle passes the test in many solar and geomagnetic records
• The 27-day rotation cycle (solar rotation period as seen from Earth) passes in geomagnetic data
• Claimed 35-year rainfall cycles pass in some regional datasets, fail in others, the evidence is genuinely ambiguous

This last point is important for understanding the Inigo Jones legacy. Jones built long-range rainfall forecasts on the 35-year Bruckner cycle. Whether that cycle passes rigorous significance testing in Australian data depends heavily on which rainfall stations you use and which period you analyse. The cycle appears genuine in some records. In others it's within noise.

See our post on Inigo Jones and the Crohamhurst Observatory for more on the historical context.

Bartels test vs FFT spectrum analysis

The Bartels test and spectral analysis (FFT) are complementary, not competing.

FFT gives you the full power spectrum, the amplitude at every frequency, in one calculation. It's fast and comprehensive. The limitation is that spectral peaks don't come with automatic significance levels. A peak in a noisy spectrum might be real or might be an artefact.

The Bartels test addresses a single candidate period with explicit statistical testing. It's slower, you test one period at a time, but it gives you a proper significance measure.

In practice, the workflow is: use FFT to identify candidate periods, then apply Bartels testing to the significant peaks to determine which are worth taking seriously. This combination is covered in detail in our FFT, Hurst, and Bartels cycle analysis guide.

What to do with a significant Bartels result

A period that passes the Bartels test at p < 0.01 tells you the cycle is almost certainly real in the historical data. It does not tell you:

• That the cycle will continue in the future
• That the cycle is stable (amplitude and phase can drift)
• That the cycle is tradeable on its own

Cycles are real but they're not mechanical. A genuine 40-week stock market cycle will occasionally extend to 45 weeks or compress to 35 weeks. The Bartels significance tells you there's structure; the backtesting primer tells you whether that structure is exploitable.

Practical access

The original Bartels test was published in 1935 and is in the public domain. Python implementations are straightforward. For a pre-built environment with Bartels testing alongside other cycle analysis tools, see Crohamhurst.app.

Related: FFT, Hurst, and Bartels: Cycle Analysis for Financial Data | Backtesting Cycle Methods: A Primer | Inigo Jones and the Crohamhurst Observatory