How Accurately Can We Measure Cycle Periods?

When a piece of cycles research reports a period like "26.65 million years" or "3.41 years," there's an obvious follow-up question that too often goes unanswered: how precisely is that actually known? A number without a stated precision invites two opposite mistakes — dismissing a real match because it "isn't exact," or accepting a false match because it "looks close." Both are avoidable with a simple rule.

The rule

Accuracy = cycle period ÷ (10 × number of full cycles present in the data)

The "number of full cycles present" is just the total length of the data divided by the candidate period. So a longer data span, or a shorter candidate period, both mean more repeats of the cycle are on record — and the more repeats you have, the tighter the accuracy the rule assigns.

A necessary condition: the method has to allow a fractional number of cycles

The rule only holds for spectral methods that let a trial period fit a non-integer number of cycles into the data window — techniques in the least-squares spectral analysis family, which search continuously across frequency rather than only at fixed grid points. 

CATS software does do fractional numbers of cycles in the data, down to 0.01 variations which is more than adequate for getting within 0.1 cycles.

An ordinary DFT (or the FFT that computes it efficiently) doesn't have this freedom: it's restricted to a fixed harmonic grid, which effectively forces a whole number of cycles to fit exactly into the data span. That constraint alone costs a factor of about ten in the achievable precision. So the rule above assumes a fitting method that isn't hamstrung this way — with a plain FFT, the same data yields roughly ten times the uncertainty.

Why it works

The intuition is straightforward: each time a cycle completes another full repeat within the data, that's another independent confirmation of exactly where its peaks and troughs fall. Fitting the centre of a known, clean waveform shape against many repeats pins down its period far more tightly than a single repeat could, because the errors in locating each individual peak partly cancel out across the whole set rather than adding up.

The factor of 10 in the rule isn't a guess — it was established empirically, by testing the method on monthly commodity price data containing cycles of known, exactly-verifiable length: 12, 6, 4, 3, and 2.4 months, all arising directly from the calendar seasons. Because the true periods in that test case were already certain, the actual precision the method achieved could be measured directly against them, and a consistent ×10 improvement over the naive resolution limit held up.

Applied to the recent factor-analysis results from 281 years of spliced European commodity price data, the rule reproduces the exact stated precision for each cycle found:

Notice the pattern: shorter cycles, with many more repeats fitting into the same data span, get pinned down far more tightly than longer cycles with only a few repeats — exactly what the rule predicts.

Why the distinction matters in practice

Applying the rule honestly changes which apparent "matches" between a found cycle and a known reference cycle actually hold up.

A found period of 10.96 years, with a rule-derived accuracy of ±0.04, is a genuine match to the roughly 11-year sunspot cycle — the gap between them is well inside the stated uncertainty.

But a found period of 3.411 years, with a much tighter accuracy of ±0.004 owing to its many repeats in the data, turns out not to be a genuine match to the well-known 3.389-year (40-month) stock market cycle — the gap between them, though small in absolute terms, is more than five times the stated precision.

This is the central, sometimes uncomfortable lesson: a number that "looks close" is not the same as a number that matches within its properly calculated precision.

Appendix: how this relates to the uncertainty principle

The rule above answers a different question from the one usually raised when "uncertainty" comes up in cycles work, and it's worth being precise about the difference.

There are really two separate questions:

1. How close together can two real periods be before we can no longer tell them apart at all? This is a hard limit set purely by how much data you have.
2. Given that a single, clean cycle really is there, how precisely can we pin down its exact value? This is a much softer limit, set by the quality of the fit — and it's what the rule above answers.

The first question is where a version of Heisenberg's uncertainty principle comes in. It can be rearranged from its usual energy-time form (ΔE·Δt ≥ a fixed constant) into a statement about cycles: (time uncertainty) × (period uncertainty) is bounded, where the time uncertainty is simply the length of the data. This is the classical time-bandwidth product (the Gabor limit) that falls out of the same Fourier mathematics behind the quantum relation — not because it's literally quantum mechanics, but because both are consequences of the same fact about Fourier-conjugate pairs. Converted to a period uncertainty, it becomes the standard Rayleigh frequency-resolution criterion from spectroscopy and signal processing: period² ÷ data-span — the naive baseline that the rule above improves on by roughly a factor of 10.

The reason the two questions have different answers is the same reason a telescope's diffraction limit sets how close two stars can appear before blurring into one, while the position of a single isolated star can still be measured to a small fraction of that same diffraction width, given a clean enough signal. The Fourier/Rayleigh limit sets the width you can't beat without more data; the ×10 rule is the empirically-measured centroiding gain available within that width when a single clean cycle is being fitted rather than two cycles being separated. Both trace back to the same underlying relation — one from the physics-textbook side, one rediscovered independently from the data-analysis side.

A side note on terminology: the general family of fractional-cycle-fitting methods described above is usually called Least-Squares Spectral Analysis (LSSA), or the Vaníček method after its 1969 originator. The discussion here assumes evenly-spaced (regular) data throughout. For unevenly-spaced or gapped data, a related variant called the Date-Compensated Discrete Fourier Transform (DCDFT) — along with the closely related Lomb-Scargle periodogram — extends the same fractional-cycle fitting to irregular sampling, which is common in astronomy.