Lorentzian Classification: The Complete Machine-Learning Guide

What Lorentzian Classification is

At its heart, Lorentzian Classification is a k-nearest-neighbours (k-NN) classifier applied to price. k-NN is one of the oldest and most intuitive machine-learning methods: to classify a new example, you find the k most similar past examples and let them vote. Applied to trading, each "example" is a snapshot of market conditions on a bar — described by a handful of indicator values — and the label is what price did next. The current bar is classified by finding the most similar historical bars and seeing whether they were mostly followed by up moves or down moves.

What makes this particular indicator distinctive is the choice of distance. To decide which past bars are "most similar," you need a way to measure the gap between two market snapshots. Most methods use Euclidean distance — ordinary straight-line geometry. Lorentzian Classification instead uses Lorentzian distance, a measure borrowed from the mathematics of spacetime, which compresses the effect of large differences. In practice this means extreme, volatile readings distort the similarity search less, so the neighbours it finds tend to be more genuinely comparable in turbulent markets.

QuantumBot

Automate your trades. Let Quantum Algo trade for you.

Every signal executed on your own account — 24/7, hands-off.

Automate now →$199/mo · your keys, your funds

Verified track recordTradingView

75%▲ +92R

Win rate · 140 trades · public ledger

5

Exchanges

400+

Pairs

<50ms

Fills

How k-nearest-neighbours works here

The mechanics are easier to follow as a sequence. Each historical bar is stored as a point in a "feature space" whose axes are the indicator readings on that bar. When a new bar forms, the model places it in that same space and measures the Lorentzian distance to every stored point. It keeps the k closest — the nearest neighbours — and looks at how price resolved after each of them. If most of those similar past situations were followed by an up move, the current bar is classified bullish; if most led to down moves, bearish.

The number k matters. A small k makes the model sensitive and reactive but noisy; a large k smooths the vote but can blur genuine signals. This is exactly why adaptive implementations, like AI-clustered Supertrend, adjust k to conditions rather than fixing it — the Quantum ML Engine in the chart above shows an "Adaptive K" of 6/8, meaning it is currently using six of a possible eight neighbours based on the market state. The classification you see is the aggregated vote of those historical analogues.

Why Lorentzian distance instead of Euclidean

The distance metric is the heart of what makes this indicator different, so it is worth understanding intuitively. Imagine two market snapshots that are similar in most respects but differ sharply on one volatile reading — say a momentum spike. Euclidean distance squares differences, so that single large gap dominates the calculation and pushes the two snapshots far apart, even though they are alike in every other way. The model would then overlook a genuinely comparable historical bar.

Lorentzian distance grows only logarithmically with large differences, so a single extreme reading does not blow up the similarity measure. Similar-but-volatile bars stay close together in the feature space, complementing tools like the Gaussian Channel, which means the nearest-neighbour search finds more truly analogous situations when markets are turbulent — exactly when reliable analogues are most valuable. This robustness to volatility and outliers is the whole reason the Lorentzian metric was chosen, and it is why the indicator often behaves more sensibly through chaotic price action than a plain-geometry classifier would.

The features: what the model actually "sees"

A k-NN classifier is only as good as the features that describe each bar. Lorentzian Classification typically describes market conditions with a small set of normalised technical readings — commonly momentum and trend oscillators such as RSI, a Wave Trend or CCI-style measure, and ADX for trend strength. Each bar becomes a short list of these values, and that list is its coordinates in the feature space where similarity is measured.

The choice and number of features shape everything downstream. Too few and the model cannot distinguish meaningfully different situations; too many and the "curse of dimensionality" makes every bar look roughly equidistant, weakening the neighbour search. Good implementations keep the feature set compact, normalised, and complementary — each feature capturing a different aspect of the market — so that "nearest" in the feature space genuinely means "most similar" in market terms.

Reading the Quantum ML Engine dashboard

The chart above is Quantum Algo's Adaptive Lorentzian implementation, and its dashboard makes the model's internal state visible — which is a good way to learn what the indicator is doing. Each row is a live output of the classifier and its supporting logic, so you can see not just a signal but the context behind it.

The most important habit is to read the signal in context. A short signal with low confidence in a contracting-volatility, trending regime is a very different proposition from a high-confidence signal as volatility expands. Notably, the Quantum ML Engine labels its win-rate figure as a short "calibration, not a backtest" — an honest reminder that a recent hit-rate over a small window is a health check on the model's current calibration, not a promise of future performance. Treating every readout as probabilistic context rather than a guarantee is the correct mindset for any ML indicator.

Strengths and honest limitations

Used well, Lorentzian Classification has real strengths. It adapts to the actual data rather than assuming a fixed model of the market, its Lorentzian metric makes it robust in exactly the volatile conditions where simpler tools break, unlike a moving-average tool, and a good implementation gives a nuanced, probabilistic read complete with confidence and regime context. It is a genuinely different lens on the market from the usual moving-average toolkit.

But it is not magic, and honesty about its limits is what separates good users from disappointed ones. It is still pattern-matching on past data: if the future stops resembling the past — a genuine regime change with no historical analogue — its neighbours become misleading. It does not know why anything happens, only what tended to follow similar readings. It can be curve-fit through over-tuning — which is why honest backtesting matters, and like any indicator it can produce losing streaks. It is a high-quality probabilistic signal to combine with structure, context, and strict risk management — not a standalone oracle.

How to use it in a real strategy

The productive way to use Lorentzian Classification is as a high-quality directional filter and context engine, not as a blind buy/sell trigger. Let the classifier tell you which side has the historical edge right now and how confident it is, then require your own confirmation before acting — a market-structure shift in the same direction, a key level, or a separate entry signal. The classification sets the bias; your structure and risk plan set the trade.

It also pairs naturally with the same discipline any indicator demands: weight higher-confidence signals more, respect the regime and volatility context the dashboard provides, and size every trade with strict risk management so no single misclassification hurts. Because the Quantum ML Engine exposes confidence, regime, and volatility explicitly, you can build simple rules — for example, only taking signals above a confidence threshold in a trending regime — that turn a raw classifier into a disciplined, filtered system.