Data-to-Noise Ratio in Financial Series

• Data-to-Noise Ratio (DNR) is a metric that decomposes observed price data into a smooth signal representing tradable structure and noise from random fluctuations.
• It employs techniques like non-local total variation smoothing, Fourier-based band-pass filtering, and accelerated denoising diffusion to effectively extract underlying trends.
• Empirical evaluations reveal that DNR improvements lead to enhanced predictive metrics, such as Return Ratio and Information Coefficient, across multiple market regimes.

Data-to-Noise Ratio (DNR) is a domain-specific metric for quantifying the relative strength of signal to noise in financial time series, designed as an analog of the classical signal-to-noise ratio (SNR) in engineering but tailored to the peculiarities of financial price data. By formalizing the decomposition of observed price sequences into smooth, tradable “signal” and unpredictable “noise,” DNR provides an interpretable, quantitative measure of data quality that directly correlates with the efficacy of predictive modeling and trading strategies (Che, 2024).

1. Formal Definition and Mathematical Basis

Given an observed price sequence $x_t$ in a financial time series, the data is decomposed additively into a signal component $s_t$ (representing the underlying tradable structure) and a noise component $n_t$ (representing stochastic or adversarial fluctuations):

$x_t = s_t + n_t$

DNR is defined as the decibel-scaled power ratio:

$$\mathrm{DNR} = 10 \cdot \log_{10}\left(\frac{\mathbb{E}[s_t^2]}{\mathbb{E}[n_t^2]}\right)$$

where $\mathbb{E}$ denotes expectation over the sample window. In practical implementation, $s_t$ is estimated by a model’s own denoising output $\hat{x}_t$, leading to the concrete discrete form:

$$\mathrm{DNR} = 10 \log_{10} \frac{\sum_{t=1}^T \hat{x}_t^2}{\sum_{t=1}^T (x_t - \hat{x}_t)^2}$$

Unlike generic SNR definitions, this operationalizes DNR as a model-dependent metric whose reference “signal” is the algorithmically denoised series itself. Empirical studies show that increases in DNR are aligned with improvements in key predictive metrics, notably Return Ratio (RR), Information Coefficient (IC), and Rank IC (Che, 2024).

2. Sector-Based Synthesis for DNR Enhancement

The sector-based synthesis approach, referred to as CS-Diffusion, augments historical A-share market data at the sector and board (market segment) level, targeting both noise reduction and maintenance of signal structure. Its key components are:

a. Approximate Non-Local Total Variation (ANTV) Smoothing

ANTV extends traditional Total Variation by weighting inter-point differences over non-local neighborhoods using a similarity kernel,

$$L_{\mathrm{ANTV}}(\mathbf{x}) = \alpha \sum_{i=1}^n \sum_{j \in w(i)} |x_j - x_i| \; \omega(i,j)$$

with

$s_t$0

Iterative gradient steps on $s_t$1 smooth the series, reducing spurious volatility while preserving underlying trends.

b. Fourier-Based Band-Pass Filtering

A band-pass filter is imposed in the frequency domain:

$s_t$2

where $s_t$3 is the FFT and $s_t$4 denote the empirically determined signal band. Retaining mid-frequency energy and suppressing noise frequencies further increases DNR.

c. Accelerated Sampling via Denoising Diffusion Implicit Models (DDIM)

DDIM is used for fast, partially deterministic sampling, with substantial reductions in computational steps and residual noise compared to standard DDPM. The process maps $s_t$5 in $s_t$6 steps, enabling both efficient data synthesis and higher DNR (Che, 2024).

3. Recursive Synthesis and Sub-Time-Level Augmentation

For stocks with limited historical data or few comparable peers, a recursive synthesis methodology is employed:

a. Pattern Recognition with Markov Modeling

Common motifs (short-term price patterns) are extracted and clustered. Empirical motif transitions are modeled as a first-order Markov chain, enabling the synthesis of sequences by concatenating motifs with learned transition probabilities.

b. Sub-Time-Level Data Augmentation

Each daily price bar is subdivided into multiple sub-bars by injecting volume-weighted micro-jitters and Gaussian noise. This increases the granularity of the data, resulting in statistically richer datasets and improved separation of true signal from noise. For example, converting a 60-day sequence into 300 sub-daily points yields datasets doubly robust to noise, measurably raising DNR (Che, 2024).

4. Empirical Evaluation and Measured Impact

Experiments on A-share (Main Board, STAR, ChiNext, BJSE) and major U.S. markets (NASDAQ, NYSE, AMEX) demonstrate consistent enhancement of DNR by 1–2 dB, corresponding to roughly 50% reduction in residual variance post-synthesis. For the Main Board, DNR improvement of +1.8 dB was observed. Performance metrics for a Transformer predictor under a simple long-only trading rule are summarized as follows:

Board	RR (Orig → Synth)	IC (Orig → Synth)	Rank IC (Orig → Synth)
Main Board	0.1834 → 0.2471	0.0345 → 0.0372	0.0523 → 0.0587
ChiNext/STAR	↑8–12%	↑5–11%	↑3–6%
U.S. (all)	↑7–12%	↑7–9%	↑3–5%

Models previously discounted in noisy environments (e.g., ALSTM, HIST) exhibit substantial restoration of predictive power as a direct consequence of DNR elevation (Che, 2024).

5. Comparative Frameworks and Broader Relevance

Unlike generic SNR, DNR in this context is explicitly model-dependent and reflects the effectiveness of denoising mechanisms specific to financial time series synthesis. The sector-based, non-local smoothing, frequency-domain bandpass, and accelerated sampling techniques provide modular means to systematically target noise modes while preserving investment-relevant structure. The flexibility of these components allows direct transfer to other equity universes, commodity, forex, and macroeconomic time series where the need to reliably partition signal and noise is paramount.

6. Implications and Generalization

Enhancing DNR through such generative modeling techniques leads to demonstrable improvements in forecasting accuracy, backtested profitability, and robustness of trading strategies. The sub-time-level augmentation provides a canonical solution to data scarcity in newly listed or thinly traded securities, establishing a generalized blueprint for new-issue markets. The empirical validation across multiple market regimes and geographies suggests that DNR-focused preprocessing can serve as a universal enhancement layer in deep learning pipelines for noisy sequential data.

7. Summary and Research Directions

Systematic elevation of DNR by methods combining ANTV smoothing, spectral band-pass filtering, and accelerated diffusion has proven sufficient to yield double-digit improvements in both traditional and deep learning-based financial predictors. Recursive, motif-based generation further overcomes data scarcity barriers, particularly in short-lived assets. Widespread application of these approaches portends further advances in the extraction of actionable signals from noisy financial ecosystems, supporting improved scenario analysis and more secure model deployment in risk-sensitive domains (Che, 2024).