Why a fitted dominant cycle is not a forecast

A finished sample is not a forecast

A dominant cycle is an estimate of the primary periodic length and phase inside a finished lookback of ordered price, volume, or breadth observations. An editorial reply states that applying a Fast Fourier Transform to past prices measures historical cycle content and helps design filters, and that this use does not by itself forecast later prices.

The same reply states that recovering cycle length and phase from a finished sample is a fit to that history, not evidence that those cycles will continue. Editorial vocabulary for that reconstruction is a historical cycle fit: past observations are written as a sum of cycles whose lengths and starting positions are chosen after the sample is already complete.

Editorial vocabulary also names a phase unwrap as the extra step of extending a fitted cycle's starting position beyond the last observation. A description of the past does not automatically authorize this step.

What the letters disputed

A letter writer reports unsatisfactory cycle-analysis results on daily, weekly, and monthly broad-index series and challenges any claim of satisfactory cycle parameters to publish those parameters for comparison.

A correspondent argues that Fourier analysis is only a first-pass sorter for approximate cycle lengths when building a current cyclic model, not the foundation of that model.

A correspondent argues that a straight-line detrend before a Fourier transform distorts the spectrum because a market trend is a curve formed by a residual trend plus waves longer than the transform window. Editorial note: a linear detrend subtracts a straight line so the series is centered for spectral estimation, and the choice of detrend changes the cycles that come back.

The editorial reply counters that a non-linear detrend would itself skew recovered cycles, that Fourier phase has not been shown to extend into the future with statistical reliability, and that long-cycle bins in a Fast Fourier Transform are too widely spaced, which is why a maximum-entropy method was pursued.

Filters, other methods, and the extra choice

An editorial comment states that, except case by case, no simple formula gives an optimum pairing of moving averages or exponential averages, and that published practice used trial-and-error charting or fast Fourier analysis of historical cyclical content to tune those filters.

An editorial reply already separates that historical measurement from later-price forecasting. Editorial vocabulary therefore treats Fourier analysis here as a decomposition of a detrended historical series into cycle length and phase, used as a filter-design and description tool rather than as a standalone forecast engine.

The editorial reply names maximum-entropy spectrum analysis as an alternative it found more useful for short-horizon prediction than treating a Fourier decomposition as a forecast. The same reply lists autoregressive integrated moving-average methods and adaptive filters as two further approaches it associates with statistically framed forecasts.

Editorial reading: a maximum-entropy spectrum is a short-record spectral estimator offered when long-cycle Fourier bins are too coarse and a short-horizon prediction, not just a historical fit, is the intended output. Switching from Fourier analysis to that estimator is an extra modeling choice. It still needs an out-of-sample rule.