Lorentzian Classification Pine Script: Complete TradingView Guide

What Is Lorentzian Classification?

Lorentzian Classification is a machine learning indicator that implements a K-Nearest Neighbors (KNN) classifier using Lorentzian Distance as its similarity metric. It predicts the directional move of price over the next 4 bars and is designed to overlay directly on the TradingView price chart. Traders use it to get a data-driven directional bias without manually scanning multiple indicators.

History and Inventor

This indicator was built by the TradingView community member Jdehorty, who created the MLExtensions and KernelFunctions library ecosystem. The script was published in 2021 on TradingView and gained traction because it brought a real machine learning pipeline (feature extraction, distance-based classification, and kernel smoothing) into Pine Script for the first time. It is not based on a single academic paper but rather adapts the KNN algorithm with Lorentzian Distance to the specific challenges of financial time series. The Lorentzian Distance component borrows from classification methods used in physics and signal processing, where handling outliers reliably is critical.

How It Works

The algorithm works in three layers. First, it extracts five feature series from price: RSI (14), WaveTrend (10, 11), CCI (20), ADX (20), and a shorter RSI (9). Second, for each bar, the algorithm computes the Lorentzian Distance between the current feature vector and every past feature vector within the lookback window (default 2000 bars). It picks the 8 nearest neighbors and records what their direction labels were. Third, the sum of those labels produces a prediction: positive means bullish, negative means bearish. The Nadaraya-Watson kernel then adds a trend-smoothing layer on top.

What Markets It Suits

Stocks and ETFs: Works cleanly on SPY, QQQ, and individual large caps. The feature distributions are stable enough that the KNN finds repeatable patterns.
Crypto: Viable on BTC and ETH at 1H or higher. Lower timeframes produce too many false signals. I get roughly 42% win rate on BTC 15M with default settings.
Forex: Works on major pairs like EUR/USD but the ADX filter becomes critical because forex can sit in low-volatility ranges for weeks.
Futures: ES and NQ perform well thanks to high liquidity. The backtest table on ES daily shows about 55-60% win rate in trending months.

Best Timeframes

1H and 4H are the sweet spots. On 1H you get enough data points for the KNN to find meaningful neighbors without excessive noise. On 4H the signal quality improves but you get fewer trades. About 1-2 per day. Daily works but expect 1-3 signals per week. The indicator explicitly targets a 4-bar prediction horizon, so matching the timeframe to your holding period matters. On 1M charts, the noise rate is too high and the KNN struggles to find stable neighbors within 2000 bars of lookback.

Parameters and Tuning Guide

Lorentzian Classification has a larger parameter set than most indicators (25+ configurable inputs). The table below covers the most impactful ones.

Tuning Scenarios

Neighbors Count has the biggest impact on classification stability. A lower value (4-5) reacts faster but picks up more noise. A higher value (12-15) produces fewer, more filtered signals but can miss trend changes by 2-3 bars. I usually keep it at 8 and adjust the filters instead. The Regime Threshold is the second most influential parameter.

Lorentzian Classification Trading Strategies

These three strategies use Lorentzian Classification as the primary signal generator, with different filter configurations for different market environments.

Common Mistakes and Limitations

These are the most frequent errors I see traders make when using Lorentzian Classification.

1. 1. Ignoring the filters and taking every signal.

   The indicator produces many raw KNN predictions, but only those passing the volatility, regime, and ADX filters are shown as labels. If you disable all filters to get more signals, the false signal rate jumps from ~35% to ~65% on 1H charts. Keep at least the volatility and regime filters on.

2. 2. Using the same Neighbors Count on all timeframes.

   8 neighbors works well for 1H and 4H. On 5M charts, the feature space is noisier and 8 neighbors introduces too much averaging. The signal arrives 2-3 bars late. Drop to 5 on lower timeframes, raise to 12 on daily or above.

3. 3. Treating the kernel line color change as an exit signal.

   The kernel line changes color when the smoothed trend direction flips. This is not the same as an exit label. Exiting at every kernel flip doubles your trade count but the average win drops from 55% to around 40%. Wait for an exit label or a confirmed opposite-direction entry label.

4. 4. Overfitting the Feature Count setting.

   Five features gives the KNN a richer feature space, but at 2000 bars of lookback you risk fitting to noise rather than signal. I have seen people set Feature Count to 5 on 500 bars of lookback and wonder why the indicator overreacts. Match the lookback to the feature count. 5 features need at least 1000 bars for stable neighbors.

5. 5. Enabling both EMA and SMA filters simultaneously without reason.

   The EMA and SMA filters do similar things. Having both active makes the indicator enter fewer trades but the reduction is almost entirely in good trades, not bad ones. Pick one (EMA is faster) and leave the other off unless you are testing a specific condition.

6. 6. Relying on the indicator alone in low-volume or exotic assets.

   Lorentzian Classification uses technical features that assume stable data. On low-cap altcoins or thin forex crosses, the RSI, CCI, and ADX calculations can produce erratic values. Stick to liquid markets with at least 500k daily volume for reliable KNN neighbors.

Lorentzian Classification FAQ