# MLMA Also known as: machine learning moving average. A Trend concept (Moving-average lineage) in the LuxAlgo Library, with 1 indicator implementation. ## What is the MLMA? MLMA stands for Machine Learning Moving Average: a smoothing line whose value comes from fitting a statistical learning model to recent prices instead of applying a fixed weighting scheme. Where an [SMA](https://www.luxalgo.com/library/concept/sma/) weights every bar equally and a WMA fixes a linear ramp, an MLMA typically runs kernel regression over the lookback window, often with a radial basis function kernel, a construction closely related to Gaussian process regression, and plots the model's fitted estimate of the underlying value bar by bar. The output behaves like a flexible weighted average: smooth, locally adaptive, and able to bend with curvature that fixed-weight averages round off. Implementations commonly add bands derived from the model's uncertainty or residual dispersion. The underlying mathematics is older than the branding. Kernel-weighted local averaging goes back to the Nadaraya-Watson estimator published in 1964, and Gaussian process regression is a standard tool in modern statistics; what charting communities did was package these fits as a price overlay and name the result a machine learning moving average. Two inputs do most of the work: the window, which sets how much history the model sees, and the kernel bandwidth, which sets how quickly a bar's influence decays, together controlling the smoothness-versus-responsiveness tradeoff the way length does for an [EMA](https://www.luxalgo.com/library/concept/ema/). Viewed this way, familiar averages are just fixed kernels, an EMA being an exponential-decay kernel, and the MLMA generalizes the idea by letting the kernel be chosen and, in some versions, tuned. Machine learning here means a fitted model, not foresight. The line is still a function of past prices, so it lags at turns like any smoother, and the flexibility cuts both ways: a model that bends easily can overfit the window it sees. The other engineering concern is repainting: kernel fits that weight a centered window revise their recent history as new bars arrive, so historical plots can flatter what was actually tradeable. Repaint-safe engineering restricts the fit to one-sided, causal windows and finalizes values on bar close. The name is a community label rather than an academic category, and implementations vary between script authors, so two indicators called MLMA can behave quite differently. Conceptually the MLMA belongs to the adaptive-smoother lineage alongside the [adaptive-lookback MA](https://www.luxalgo.com/library/concept/adaptive-lookback-ma/) and the [Ehlers SuperSmoother](https://www.luxalgo.com/library/concept/ehlers-supersmoother/): all pursue less lag for a given smoothness than fixed-weight averages allow. What the kernel approach adds is a principled uncertainty estimate, which is why MLMA implementations so often ship with bands, effectively a model-driven [MA envelope](https://www.luxalgo.com/library/concept/ma-envelope/) whose width adapts to how noisy the recent fit is. ## How to identify an MLMA on a chart An MLMA plots as a smooth overlay line, usually with optional bands, and reads like any moving average once its parameters are understood. 1. Apply a kernel or machine-learning moving average indicator and locate its two key inputs: window length and kernel bandwidth (Gaussian process variants add a noise term). 2. Read the line like a smoother: slope and color for directional bias, price's side of the line for regime, and curvature for how the trend is bending. 3. If bands are plotted, note what they encode: distance from the fitted mean scaled by model uncertainty or residual dispersion, widening when recent prices fit the model poorly. 4. Test for repainting before trusting history: watch whether the last few plotted values shift as new bars close, or step through bar replay; a causal implementation's past stays fixed. 5. Benchmark against a conventional average of similar smoothness: the kernel fit should track curvature through turns that the fixed-weight line rounds off, with comparable noise rejection. ## How traders use it - In the standard moving-average roles: slope or color for trend bias, price's side of the line for regime, and [crossovers](https://www.luxalgo.com/library/concept/moving-average-crossovers/) against a faster line for timing. - As an adaptive envelope: the model's uncertainty bands frame how stretched price is from the fitted mean, supporting reversion and breakout reads that fixed-percentage envelopes handle crudely. - As a smoothness upgrade in existing systems: swapping a fixed-weight average for a kernel-based one where the goal is less lag for a given smoothness, verified on closed-bar behavior. - As dynamic support and resistance: in a trending regime, pullbacks to the fitted line are watched for reactions, the same playbook as [dynamic S/R via MA](https://www.luxalgo.com/library/concept/dynamic-s-r-via-ma/), with the band edges serving as the deeper zone. - As a regime gate for other systems: the line's slope, thresholded like an [MA slope filter](https://www.luxalgo.com/library/concept/ma-slope-filter/), or price's side of it can switch a strategy between [breakout](https://www.luxalgo.com/library/concept/breakout/) logic and mean-reversion logic. ## MLMA vs other trend smoothers - **EMA** (https://www.luxalgo.com/library/concept/ema/): An EMA is a fixed exponential-decay kernel: one parameter, fully causal, never repaints, and cheap to compute. The MLMA generalizes the weighting and can track curvature better, at the cost of more parameters, heavier computation, and implementation-dependent repainting risk. - **Adaptive-lookback MA** (https://www.luxalgo.com/library/concept/adaptive-lookback-ma/): Adaptive moving averages keep a simple weighting scheme but vary the effective length with market conditions. The MLMA instead changes the weighting function itself via a kernel fit. Both chase the same goal, responsiveness in trends and smoothness in chop, by different routes. - **Ehlers SuperSmoother** (https://www.luxalgo.com/library/concept/ehlers-supersmoother/): The SuperSmoother is a fixed digital filter designed from signal-processing principles to suppress short-wavelength noise with minimal lag. It is deterministic and parameter-light, while the MLMA is a statistical fit whose behavior depends on kernel and window choices. - **Supertrend** (https://www.luxalgo.com/library/concept/supertrend/): Supertrend is a stepped, volatility-offset trailing line built for unambiguous flip signals, not a smoother of price. An MLMA estimates the local mean and bends continuously; Supertrend rides at an ATR distance and only flips sides. They answer different questions and are often layered together. ## FAQ ### Is a Machine Learning Moving Average really machine learning? In the statistical sense, yes. Kernel regression and Gaussian process regression are standard nonparametric learning methods, and the line is a fitted model rather than a fixed formula. What it is not is predictive intelligence: it estimates the local mean of past prices, it does not learn where the market goes next, and it lags at turns like any average. ### Does the MLMA repaint? It depends on construction. A fit that weights bars on both sides of each point revises recent values as new data arrives, which repaints the plot and flatters backtests of its signals. Causal versions that use only bars up to the current one, evaluated on close, do not repaint. Check which type an implementation uses before trusting its historical signals. ### Is an MLMA better than an EMA or SMA? Better is the wrong frame; the tradeoff moves. Kernel fits can follow curvature that fixed weights round off and can carry honest uncertainty bands, but they bring extra parameters, heavier computation, and, in careless implementations, repainting. On many series a well-tuned EMA performs comparably. The only fair comparison is closed-bar behavior on the market and timeframe actually traded. ### What settings does an MLMA have? Nearly all versions expose a window length and a kernel bandwidth; Gaussian process variants add a noise parameter, and many add band-width multipliers. Longer windows and wider bandwidths smooth more and lag more; shorter, narrower settings hug price and risk fitting noise. Defaults differ by author, so settings do not transfer between scripts the way EMA lengths do. ### Can a machine learning moving average predict future price? No. It estimates the current local mean of past prices, sometimes with a short extrapolated tail, and extrapolations from a smoother are notoriously unreliable at turning points. Any practical value comes from the same place as with ordinary averages: trends persisting often enough for trend-following rules built on the line to earn their keep. ### Why do different MLMA indicators give different lines? Because MLMA is a community label, not a specification. Authors choose different kernels, windows, band constructions, and causality conventions, so outputs differ even on identical data. Read the script or its documentation, then verify closed-bar behavior, before assuming results from one MLMA carry over to another. ## Implementations in the Library - MLMA (LuxAlgo): https://www.luxalgo.com/library/indicator/mlma/ ## Related concepts - SMA: https://www.luxalgo.com/library/concept/sma/ - EMA: https://www.luxalgo.com/library/concept/ema/ - Adaptive-lookback MA: https://www.luxalgo.com/library/concept/adaptive-lookback-ma/ - MA Envelope: https://www.luxalgo.com/library/concept/ma-envelope/ - SWMA: https://www.luxalgo.com/library/concept/swma/ - RMA: https://www.luxalgo.com/library/concept/rma/ - HMA: https://www.luxalgo.com/library/concept/hma/ - KAMA: https://www.luxalgo.com/library/concept/kama/ - JMA: https://www.luxalgo.com/library/concept/jma/ - ZLEMA: https://www.luxalgo.com/library/concept/zlema/ --- Source: https://www.luxalgo.com/library/concept/mlma/ (LuxAlgo Library, the encyclopedia of trading & technical analysis). Free to use with attribution: https://www.luxalgo.com/library/license/