Causal Filtering Options for Nonperiodic Financial Time Series

Summary

The discussion concerns filtering a stationary but nonperiodic time series in a causal way, a challenge for Fourier methods because they do not localize frequency changes in time. The questioner has tried empirical mode decomposition and intrinsic time-scale decomposition, but considers them unsuitable for causal use, and asks what alternatives exist beyond wavelets. The replies mention wavelets and Kalman filtering as possibilities, while noting that wavelet boundary effects can distort results.

Another response recommends Jurik’s JMA filter and claims it performs better than Kalman and Volterra filters, but provides no details of its algorithm or supporting comparison. The recommendation is explicitly uncertain because JMA is treated as a black box. Overall, the exchange is a brief list of candidate techniques and a warning about boundary artifacts, not a tested comparison or a prescription for trading. It leaves open how causality, delay, noise reduction, and performance should be evaluated for a particular signal.

Key ideas

• Fourier methods can be poorly suited to nonperiodic signals because they are not local in frequency.
• Wavelets and Kalman filtering are suggested as causal signal-processing candidates.
• Wavelet analysis can suffer from boundary distortions that affect interpretation.
• JMA is recommended by one respondent, but its internal method and claimed performance are not substantiated.

Tags

# DSP: stationary non-periodic signal: what's the best causal technique?


# DSP: stationary non-periodic signal: what's the best causal technique?












This is a bit DSP-related: so if you turn your non-stationary time series into a stationary process, you'll probably see that it is not periodic.. This is an issue for Fourier-based techniques because they are not local in frequency. Now, besides wavelets (some types are causal btw), which other causal techniques can you use? (and ARMA is not it). I tried Empirical Mode Decomposition (HHT), but that's not causal; I tried Intrinsic Time-scale Decomposition: not causal either. Wavelets are pretty old and I would think something better would have been "discovered" by now? Does anyone know of a good causal signal processing technique that deals well with non-periodicity? Thanks!!

## Answer by Igor (score 3)

https://quant.stackexchange.com/a/1950

I know only that Jurik's JMA is good causal filter, better than Kalman and Volterra filters, but I don't know for sure what algorithm inside - it's black box. Does anybody know better causal filter?

## Answer by RockScience (score 2)

https://quant.stackexchange.com/a/833

The issue with wavelets is that you'll have some boundary distortions so be careful when exploiting the results.

## Answer by Tangurena (score 2)

https://quant.stackexchange.com/a/1084

Wavelets and Kalman filtering.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.