Which Time‑Series Smoothing Method Is Right for Your Data? A Deep Dive into Six Techniques

Overview

Time‑series data are frequently corrupted by sensor defects, measurement errors, or intrinsic statistical fluctuations, which mask underlying trends. Six widely used smoothing techniques are examined with respect to their mathematical principle, key parameters, performance characteristics, and typical application scenarios.

1. Moving Average (Rolling Mean)

The moving average replaces each observation with the arithmetic mean of a symmetric window of window size points surrounding it.

Key parameter: window size (number of points).

Advantages: extremely simple, O(N) computational cost, effective at attenuating short‑term noise while preserving long‑term trends.

Limitations: symmetric window introduces a phase delay proportional to half the window length; sharp spikes or step changes are smoothed out; assumes zero‑mean stationary noise.

2. Exponential Moving Average (EMA)

EMA computes a weighted average where the weight of each past observation decays exponentially. The recursive update is: EMA_t = α·x_t + (1‑α)·EMA_{t‑1} Key parameter: smoothing factor α (0 < α ≤ 1). Larger α gives more emphasis to recent data, reducing lag but also reducing smoothing.

Advantages: causal (uses only current and past data), low memory footprint, fast response suitable for streaming.

Limitations: still exhibits lag for small α; performance degrades when the underlying signal changes faster than the filter can track.

3. Savitzky‑Golay Filter

This filter fits a low‑order polynomial to the data inside a moving window and replaces the central point with the polynomial’s value, preserving higher‑order moments such as slope and curvature.

Key parameters: odd‑sized window length (number of points) and polynomial degree . Larger windows increase smoothing; higher degree captures more complex local shapes.

Advantages: retains peak height, slope, and curvature; ideal for signals with structured features.

Limitations: over‑fitting can occur if the window is too small or the degree too high, potentially amplifying noise.

4. LOESS (Locally Estimated Scatterplot Smoothing)

LOESS performs a weighted local regression for each point. For a target point, a fraction frac of the nearest observations is selected, weights are assigned by distance, and a low‑order polynomial (usually linear) is fitted.

Key parameter: frac – proportion of data used in each local fit. Smaller values follow the data more closely.

Advantages: highly adaptive, handles non‑linear trends and irregular sampling without assuming equally spaced points.

Limitations: computationally intensive (O(N²) in naïve implementations); very small frac may over‑fit; edge estimates can be unstable due to fewer neighbours.

5. Gaussian Filter

The Gaussian filter applies a weighted moving average where the weights follow a Gaussian distribution centred on the target point. The weight for offset k is exp(-k²/(2σ²)), where σ is the standard deviation.

Key parameter: σ (standard deviation). Smaller σ preserves fine detail; larger σ yields stronger smoothing.

Advantages: produces smooth results with minimal edge artifacts; single‑parameter tuning.

Limitations: applies uniform smoothing across the entire series, which can blur sharp edges or peaks in rapidly changing regions.

6. Kalman Filter

The Kalman filter is a recursive Bayesian estimator for linear‑Gaussian systems. At each time step it predicts the state using a process model, then updates the prediction with the new observation, weighting each by its estimated covariance.

Key parameters: process‑noise standard deviation ( transition_std) and observation‑noise standard deviation ( observation_std). These control the relative trust in the model versus the measurements.

Advantages: optimal (minimum‑variance) under linear‑Gaussian assumptions; handles time‑varying noise, missing data, and real‑time streaming.

Limitations: higher computational cost; requires an accurate linear process model; diverges for strongly non‑linear or non‑Gaussian dynamics unless extended/unscented variants are used.

Practical Challenges in Real‑World Smoothing

Real data often exhibit irregular sampling, missing values, and mixed change rates. Many smoothing algorithms assume regularly spaced, complete data, so preprocessing (e.g., interpolation) may be required. Causal filters (EMA, Kalman) are preferable for streaming scenarios because they do not rely on future observations.

Choosing appropriate parameters is non‑trivial; automated validation is difficult, and visual inspection remains a reliable sanity check.

Roughness Preservation Ratio (RPR) for Quantitative Evaluation

RPR quantifies how much short‑term variation (roughness) remains after smoothing:

RPR = Σ |ŷ_{i+1} – ŷ_i|  /  Σ |y_{i+1} – y_i|

where y is the original series and ŷ the smoothed series. Values near 1 indicate mild smoothing, values near 0 indicate strong smoothing, and values > 1 imply the method introduced additional high‑frequency variation. RPR measures only roughness reduction, not overall shape preservation.

Implementation Reference

Interactive visualizer and reference implementations are available at:

https://github.com/dbolotov/ts_smoothing_visualizer