Cleaning Up the Noise of the NZX 50
As the academic rigour of Semester 1 accelerates, the collaborative “Signal Processing & Econometrics Group” at Parvis School of Economics and Music has presented its first controversial paper of the year: “Spectral Subtraction in High-Frequency Trading: A Dangerous Smoothing?”
The research, presented this week at the Inverlochy Place Lecture Theatre, represents a daring methodological crossover. Third-year students from the School of Sonic Arts took algorithms standardly used for restoring vintage vinyl recordings (such as De-Clicking and De-Crackling filters) and applied them to the raw, noisy tick-data of the New Zealand Exchange (NZX 50).
The hypothesis was elegant in its simplicity: specific patterns of market volatility behave mathematically like “Pink Noise” (1/f noise) found in acoustics. By applying audio mastering filters, the students aimed to strip away the “hiss” of inconsequential algorithmic trading to reveal the “true melody” of the market trend.
The Methodology: Treating Money as a Waveform
Under the supervision of Dr. Elara Vance (Music) and Ms. Yuna Kim (Econometrics), the cohort converted three years of NZX trading data into .wav audio files. Instead of listening to them, however, they processed the data through MATLAB’s digital signal processing (DSP) toolbox.
“We utilised a Fast Fourier Transform (FFT) to break the market data down into its constituent sine waves,” explained James Tautari, a dual-major student leading the coding team. “We then applied a high-pass filter to remove long-term macro trends, leaving us with the high-frequency ‘jitter’ of daily trading. We effectively tried to ‘master’ the stock market like a pop song.”
The initial results were visually seductive. The processed graphs showed remarkably smooth trend lines, removing the erratic spikes that often confuse traditional moving-average models.
The “Transient Smearing” Failure
However, in the spirit of academic inquiry, the presentation focused heavily on where the model failed. While the audio filters successfully removed “noise,” they also inadvertently removed “risk.”
The critique came from the Economics faculty. In audio engineering, a sudden, sharp spike in volume (a “transient”) is often a mistake—a pop from a microphone or a scratch on a record—and should be removed. In financial markets, however, a sudden sharp spike is often a crash or a major news event (a “Black Swan”).
“The algorithm worked too well,” noted Ms. Kim during the post-presentation Q&A. “It identified the market crash of early 2024 as an ‘audio artefact’—essentially treating a financial crisis as a ‘click’ on a vinyl record—and smoothed it out completely. The model predicted a stable market when, in reality, investors would have lost everything.”
The Problem of “Look-Ahead” Bias
The team also identified a latency issue known in audio circles as “look-ahead.” High-quality audio limiters need to see a few milliseconds into the future to prevent distortion. When applied to real-time finance, this created a theoretical impossibility—the model required data that hadn’t happened yet to smooth the curve effectively.
“It is a classic case of overfitting,” admitted Tautari. “We created a model that looks beautiful in hindsight (backtesting) but is functionally blind in real-time.”
Future Iterations
Despite the critical flaw regarding transient preservation, the research has opened a new avenue for inquiry. The team is now pivoting to use “Compressors” (dynamic range compression) rather than “Denoising” tools to see if they can model market liquidity constraints.
The full paper, including the raw MATLAB code and the “remastered” financial charts, is available on the internal Parvis server for peer review. It serves as a potent reminder that while the mathematics of music and money are similar, their consequences are vastly different.
Leave a Reply