LOESS Smoothing

What is LOESS Smoothing?

LOESS (locally estimated scatterplot smoothing) fits a smooth curve through noisy data one point at a time. For each evaluation point it takes a neighborhood of nearby observations, fits a low-degree polynomial (locally linear in the original LOWESS, often quadratic in LOESS) by weighted least squares with nearer points weighted more heavily (a tricube weighting is the classic choice), and records the fitted value at that point. Repeating this across the series traces a flexible curve with no assumed global shape. The method comes from statistician William Cleveland; its span parameter, the fraction of data used in each local fit, controls smoothness.

On a chart, LOESS behaves like a trend curve that bends with the data, where a single linear regression forces one straight line; kernel regression is its close cousin, built from weighted averages rather than local fits. The catch is the right edge: the newest bars have no future neighbors, so their fit is one-sided and gets revised as bars arrive. Historical values look excellent precisely because they were smoothed with hindsight.

The parameters are few and consequential. The span, commonly between a fifth and a half of the data in chart applications, is the bias-variance dial: small spans follow every wiggle, large ones flatten genuine bends. Degree matters less, locally linear for stability, quadratic where the curve must track curvature through turns. Cleveland's robust variant adds iterations that reweight by residual size, so outliers progressively lose influence and one wild bar cannot drag the local fit, a genuinely useful upgrade on spike-prone market data. The cost of all this fitting is computational: every point is its own regression, which is why LOESS lives in analysis tools more comfortably than in per-tick engines.

Chart implementations exercise both its strengths and its known weaknesses. Channel builds wrap the curve in residual-scaled envelopes, framing stretch around a local fit rather than a rigid average; extrapolation builds project the fitted curve forward, which is the method's weakest deployment, since a local smoother has no model beyond its last neighborhood; and monotone-fit relatives like isotonic regression solve a cousin problem, best fit under an order constraint rather than a smoothness one. Across all of them the right-edge honesty rule applies: the newest stretch of any LOESS curve is provisional by construction.

How to identify LOESS smoothing

Local fits, distance weights, connected values: the construction explains both the beauty and the edge problem.

1. 1Choose the span, the fraction of data each local fit uses, and the degree, linear or quadratic; these two settings are most of the method.
2. 2For each evaluation point, weight neighbors by distance, the classic tricube giving smooth influence that vanishes at the neighborhood edge.
3. 3Fit the local polynomial by weighted least squares and record its value at the evaluation point.
4. 4Optionally run robustness iterations: reweight by residuals so outliers lose influence, then refit.
5. 5Connect the fitted values into the curve, and treat the newest segment as provisional, since one-sided neighborhoods revise as bars arrive.

How traders use it

• As a flexible trend baseline: the curve's direction and bend give a visual trend read, and subtracting it from price yields a detrended oscillator of local deviations.
• As the centerline of a channel: offsetting the curve by a multiple of the residual standard deviation frames stretch and reversion around the local fit.
• As a smoother for indicator inputs where fixed-length moving averages feel too rigid, accepting the edge instability that comes with recomputation.
• As a detrender before cycle work: removing the LOESS trend isolates the oscillation for dominant cycle measurement without imposing a global trend shape the data never had.
• As projected scenario paths: extrapolation builds extend the fitted curve forward as a sketch of current curvature persisting, honestly the method's weakest use, and read as a visual hypothesis rather than a forecast.

LOESS vs related fitting methods