How Elliott Wave counts are generated automatically

Four stages, in this order. Each one narrows the candidate set, and the last one is the only place where anything resembling an opinion enters.

1. Detect swings

Price is reduced to a sequence of turning points using a percentage threshold, so a move counts as a swing only when it exceeds a share of the prior leg. The threshold is the first assumption in the whole chain: a coarser one hides sub-structure, a finer one produces swings that the theory would treat as noise.

2. Fit candidate structures

Every plausible starting point is tried against a library of patterns — impulses, corrections, triangles, partial and still-forming counts, and structures read from the right edge backwards. This is combinatorial and it is why the work is worth automating: the number of anchor combinations on a few hundred bars is far past what anyone labels by hand.

3. Eliminate by rule

Each candidate is checked against the hard rules of the framework: direction alternation between waves, the retracement limit on wave 2, wave 3 not being the shortest impulse leg, wave 4 not overlapping wave 1, and the corrective-structure rules for zigzags, flats and triangles. A candidate that fails a hard rule is discarded rather than penalised.

4. Rank what survives

Survivors are scored — on the share of guidelines they satisfy, on proportion and Fibonacci relationships, on how well the degree fits the timeframe — and a trained model orders competing counts. Ranking is the only inferential step, and it decides presentation, not validity.

The two steps that matter most for reading the output are the first and the last. The swing threshold decides what the engine can see; the ranking decides what you see first. Neither is a fact about the market.

Elliott Wave analysis is contested and has no established statistical validation in the academic literature. The standing objection is structural: the rule set admits enough alternate counts that a labeling can be fitted to nearly any chart after the fact, which makes the framework difficult to falsify. Anyone selling wave software has an interest in not saying that, so it is worth saying plainly on a page that also describes wave software.

Automation does not settle the argument in either direction. What it changes is narrower and real. Proposing counts, testing them against the rules and re-testing after every new swing is mechanical, repetitive work, and a machine does it without preferring the count it would like to be true — which is the specific failure mode hand-labeling is known for. It also makes the alternates visible: an engine that shows three surviving counts with their scores is harder to mistake for a forecast than a single hand-drawn labeling on a chart.

The usable form of the output is therefore a hypothesis with a score and an invalidation level, not a prediction. Read that way it is a structuring device for what price would have to do to prove the reading wrong. Read as a forecast, it inherits every objection above and adds false precision on top.

Automatic labeling is not a rare capability, and any page claiming otherwise is worth distrusting — including earlier copy on this site, which used to claim it was alone in doing this on the web. That was wrong. The differences that actually separate the products are which markets the counts run on, whether wave work sits behind its own price tier, and how much of the reasoning is shown.

A fuller side-by-side of the wave tooling, including what each edition costs, is on the MotiveWave comparison.

Alternate counts do not go away

Several labelings passing every rule on the same swings is the normal case, not an edge case. An engine that presents one count without its alternates is hiding the part of the picture that determines the risk.

The count revises, and that is correct behaviour

A new swing can invalidate the leading count and promote one that ranked lower. Any product whose labels never move is either not re-evaluating or not showing you that it did.

Degree is a heuristic

Assigning a count to a degree is calibration from swing magnitude and timeframe, not a measurement. Two engines can agree on the structure and disagree on its scale, and the scale is what determines the horizon of the claim.

Ranking is not evidence

A model ordering surviving counts is trained on what previously continued, which embeds the same unfalsifiability problem one level up. Treat the top count as the one shown first, not as the one confirmed.

Scores compress information

A count carrying "passed 8 of 10 guidelines" is more legible than a number on its own, and it still says nothing about which two failed or whether they were the two that mattered. Opening the breakdown is the difference between using the score and trusting it.

• The engine fits 24 wave patterns, checks every candidate against the hard rules, scores survivors on the guidelines they satisfy and ranks competing counts by that score; no trained model is involved yet. Labels stay editable by hand afterwards.
• Automatic labeling is free. The scenario inspector, which breaks a count down rule by rule so you can see which ones it failed, and the scanner, which looks for wave setups across the market, are premium at $120 per month.
• Counts run on everything the terminal charts: Binance spot crypto, US stocks and ETFs as Binance perpetual contracts, and forex, metals and indices through a MetaTrader bridge. A stock contract's history starts at its Binance listing, so a long count there lacks its early waves. No listed shares and no exchange-traded futures.
• The count sits on the same chart as the order-flow layers, so a structural hypothesis and the effort behind the current leg are read together rather than in two products.