In one sentence — Prony's method tries to describe a series as the sum of a few waves, estimating each wave's frequency, amplitude, and phase. The result shows how well those waves reconstruct the sample used; it does not prove that they will continue into the future.
What is Prony's method according to Hurst?
In Appendix VI of The Profit Magic of Stock Transaction Timing, Hurst starts from the output of an Ormsby band-pass filter. The filter has already reduced the series to a frequency interval; Prony's method tries to express what remains with a limited number of sinusoids.
The idea resembles identifying instruments inside a chord. The combined sound is observable, while the calculation tries to separate the individual notes and measure their intensity and position within the cycle.
The fitted model
For observations at times (t_k), Hurst postulates a sum of (m) components:
S(t_k) ≈ Σ [Aᵢ cos(ωᵢt_k) + Bᵢ sin(ωᵢt_k)]
Each (ωᵢ) is a frequency. Coefficients (Aᵢ) and (Bᵢ) describe the same component on the cosine and sine axes. There must be more observations than coefficients: Appendix VI sets the condition N > 2m.
This inequality permits an overdetermined fit, but it does not choose the correct value of (m). The number of components, the data window, and the preceding filter remain research assumptions.
The four-step procedure
1 — Fix the scope
Choose the filtered series, the (N) observations, and model order (m). Changing these elements after seeing the result changes the problem rather than merely repeating the calculation.
2 — Recover the frequencies
When the frequencies are unknown, Hurst builds N − 2m equations to estimate (m) intermediate coefficients (α) by least squares. These enter a polynomial written with Chebyshev polynomials, where (x = cos ω):
T₀(x)=1 · T₁(x)=x · Tᵣ₊₁(x)=2xTᵣ(x)−Tᵣ₋₁(x)
The admissible roots yield candidate frequencies (ωᵢ). This is the core Prony step.
3 — Estimate cosine and sine coefficients
The frequencies are substituted back into the original model. A second least-squares problem estimates (Aᵢ) and (Bᵢ) by minimizing the sum of squared differences between the observed series and the trigonometric sum.
4 — Obtain amplitude and phase
Each pair is recombined as Cᵢ sin(ωᵢt + φᵢ). Under Hurst's convention:
Cᵢ = √(Aᵢ²+Bᵢ²) and φᵢ = atan2(Aᵢ,Bᵢ).
The output is therefore a set of estimates (ωᵢ, Cᵢ, φᵢ), together with the residuals left by the model.
“No additional lag” does not mean real time
Hurst notes that the fit can smooth the high-frequency residue without applying another filter and therefore without adding a new time displacement to the series already available.
The boundary matters: the input may come from a centred filter requiring data on both sides of the estimated point. Prony's method does not recover missing future observations and does not remove the edge described in the Ormsby filter entry.
Where the method can fail
Model order. An (m) that is too low can merge distinct components; one that is too high can assign frequencies to noise.
Noise and close frequencies. Polynomial roots can become unstable. Modern literature classifies Prony as a high-resolution parametric method that is sensitive to noise; preliminary filtering reduces the issue without eliminating it.
The same sample. A small residual measures fit on the data used. It is not a forecast error and is not independent validation.
Changing parameters. Frequency, amplitude, and phase may vary across windows. Calling them “objective” means that they result from a fixed numerical procedure, not that they are permanent market properties.
How to evaluate it today
A reproducible evaluation requires researchers to:
- freeze the filter, window, sampling interval, and order (m) before inspecting the test segment;
- estimate parameters only on the calibration sample;
- carry them unchanged into later out-of-sample data;
- report residuals, frequency stability, and sensitivity to the initial choices.
Hurst's technical method should not be confused with curve fitting in trading slang. Fitting a model to data is a normal statistical operation; overfitting occurs when selection exploits the sample so heavily that the result fails on new data.
Sources
- J. M. Hurst, The Profit Magic of Stock Transaction Timing, Prentice-Hall, 1970, Appendix VI, pp. 215–217 — the complete historical procedure: generalized least squares, frequencies, amplitudes, and phases.
- NIST/SEMATECH, Least Squares — the least-squares criterion, residuals, and the distinction between estimates and true parameter values.
- B. W. Rust, NIST, Fitting Nature's Basic Functions Part III: Exponentials, Sinusoids and Nonlinear Least Squares, 2002 — sinusoidal fitting and statistical diagnostics.
- F. Marvasti et al., A unified approach to sparse signal processing, 2012, Section 5.1 — model order, the Prony algorithm, and sensitivity to noise.
Links
Parabolic interpolation · Ormsby filters · Spectral analysis · Hurst's appendices