Concept

Kernel Regression

Kernel Regression, also known as Nadaraya-Watson estimator, is a Machine Learning concept. The Library holds 6 implementations, each one a working definition you can pull into Quant.

Top Kernel Regression indicators

6 total

Nadaraya-Watson Smoothers

Indicator

Nadaraya-Watson Envelope

Indicator

Machine Learning: Gaussian Process Regression

Indicator

What is Kernel Regression?

Kernel regression is a nonparametric way to draw a smooth curve through noisy data. Instead of assuming a functional form, it estimates the value at each point as a weighted average of nearby observations, with weights supplied by a kernel: a function that is largest at zero distance and decays as observations sit farther away. The classic form is the Nadaraya-Watson estimator, which divides the kernel-weighted sum of observed values by the sum of the kernel weights. On charts the distance is measured in bars, the kernel is commonly Gaussian or rational quadratic, and a bandwidth parameter sets how fast weight decays: small bandwidth hugs price, large bandwidth smooths and flattens.

Traders read the fitted curve as an estimate of the underlying path of price, often wrapped in an envelope built by adding and subtracting a multiple of a deviation measure such as the mean absolute error of the fit. The decisive implementation detail is causality. The textbook estimator is two-sided: each point is estimated using bars on both sides, so every new bar re-weights the window and redraws recent history, making the historical line look far better than anything achievable live. One-sided variants weight only past bars; they do not repaint, and they behave like a weighted moving average with kernel-shaped weights, at the cost of lag.

How to calculate the Nadaraya-Watson estimator

At each bar where the fit is evaluated, the estimate is a normalized, kernel-weighted average of the window.

1Choose a kernel (Gaussian and rational quadratic are common) and a bandwidth controlling how quickly weight decays with distance in bars.
2For the evaluation bar, weight every bar in the window by applying the kernel to the gap between that bar's index and the evaluation index.
3Multiply each bar's price by its weight, sum the products, and divide by the sum of the weights; that ratio is the estimate.
4Repeat across the chart. Weighting bars on both sides yields the smooth two-sided fit that repaints; restricting weights to past bars yields the causal, tradable version.

How traders use it

As a trend estimate: the curve's slope and slope changes are read as direction and turning signals, with the caveat that two-sided fits only confirm turns at the live edge after the fact.
As envelopes: the fit plus and minus a multiple of its typical error forms a channel, the kernel analog of an MA envelope; band touches are treated as stretched prices for mean-reversion setups.
As a denoised input: swing detection, crossovers, or divergence checks run on the smoothed series instead of raw price to cut whipsaw, accepting the added lag.
As a hindsight check: comparing causal and two-sided versions of the same fit shows how much of the visual quality is retrospective; repaint-safe engineering treats only the causal series as tradable.

Kernel Regression vs other smoothers

LOESS Smoothing: LOESS fits a small local polynomial (a line or parabola) inside each window, so it tracks curvature better near turns; Nadaraya-Watson fits a locally constant weighted mean. Both are typically computed two-sided.

Gaussian Process Regression: GPR is a Bayesian model in which the kernel defines a covariance structure, returning a posterior mean plus an uncertainty band; kernel regression returns a point estimate with no probabilistic error model.

Gaussian Filter: A Gaussian filter is a causal linear filter with fixed Gaussian-shaped coefficients applied to past bars only; the common charting form of kernel regression normalizes its weights over the window and looks both directions, which is why it appears smoother in hindsight.

More Kernel Regression implementations

Related concepts · Learned models

Neural Networks3 Gaussian Process Regression2 Kernel Density Estimation2 Support Vector Machines1 Random Forest1 Bayesian Classifiers1 Self-organizing Maps1 Reinforcement-learning Policies1 LSTM / Recurrent Networks0 LLM/transformer-derived Signals0

Concept family

Machine Learning

32 concepts mapped · 21 in the Library

Kernel Regression FAQ

Does Nadaraya-Watson kernel regression repaint?

The standard two-sided version does. Each estimate uses bars on both sides, so new data re-weights the window and shifts recent values of the curve; signals read from the historical line were not available in real time. Causal variants weight only past bars and do not repaint, but they lag like any moving average. Always confirm which form an implementation uses.

What does the bandwidth parameter control in kernel regression?

How quickly kernel weights decay with distance, which sets the smoothness of the fit. A small bandwidth concentrates weight on the nearest bars, so the curve tracks price closely but stays noisy; a large bandwidth spreads weight widely, giving a smoother, flatter, slower curve. There is no correct value: it trades responsiveness against noise, much like a moving-average length.

Is kernel regression better than a moving average?

It is a generalization rather than a rival. A trailing SMA is one-sided kernel regression with a flat, rectangular kernel, so kernel weighting mainly reshapes the smoothing. Two-sided charting versions look dramatically better than moving averages, but that edge is hindsight, not foresight. Compared causally, the difference is a modest change in weight shape, not a different class of tool.

Build Kernel Regression your way.

Quant writes, tests, and refines it with you — then it runs on LuxAlgo charting or ports to TradingView.