Confluence Engine Methodology — What It Computes and How the Layers Combine

What does the Confluence Engine compute?

The Confluence Engine computes price bands — we call them confluence zones — at which several independent measurements of the market agree, together with the price that would prove that agreement wrong. It does not forecast. It converts raw exchange data into more than forty analytical layers, normalises those layers into a single event stream, and reports where and when the evidence stacks.

Everything downstream of that — the levels NAIRO draws on air, the zones it defends or abandons, the hold-rate percentages it quotes — is read out of this one structure. The broadcast is a narration layer over a computation layer, and the computation layer is the part that can be audited. That is the whole design intent described on how it works.

A zone is therefore an artefact with four properties: a price band, a direction, a context key recording which families of evidence were present, and an invalidation price. There is no fifth field containing an opinion about what happens next.

Which raw data goes in?

Four streams, all public. Trades and depth snapshots come from exchange websockets. Candles come from exchange REST history plus the live stream. Positioning comes from funding rate and open-interest series, and the options chain used for dealer-gamma work is polled from a public options venue.

Two of those streams are honest about their limits and the engine says so rather than papering over it. Candles have deep history and can be backfilled indefinitely. Resting order-book depth exists only for the window we actually recorded, because a book is not republished historically by any venue. Any layer that consumes book depth — liquidity heatmap, walls, iceberg replenishment, the passive ledger — is faithful on the recorded window and blank before it. The chart shades those gaps instead of interpolating across them.

Trade data sits in between: aggregated trades are publicly retrievable, so footprint and delta layers can be reconstructed for historical bars, while book-derived layers cannot. Explaining that asymmetry is most of what separates a reproducible order-flow read from a plausible-looking one, which is the subject of order flow trading explained.

Why every event carries two timestamps

Every signal fed into the engine carries two timestamps, and they are not interchangeable. The first, t_start, is the drawing anchor: the bar the signal is about. The second, known_at, is the first millisecond the signal was confirmable without seeing a future bar.

The distinction is what makes historical measurement meaningful. A swing pivot is drawn at the pivot candle, as every charting platform draws it, but the engine treats it as unknown until the confirming bars have closed. A prior-day high is drawn at the candle that made the high and becomes knowable at the next day boundary. A volume-profile value area is anchored at the window start and becomes knowable at the window close, because it is an aggregate over the whole window.

Getting this wrong is the standard way backtests lie. Two real leaks were caught in our own code by a dedicated harness: a break of structure that counted as known at its own pivot bar, and a volume-profile level whose price depended on candles after its stated timestamp. Both are fixed, and both are covered by mutation tests that append maximally different future data and assert the historical prefix is unchanged.

How do the layers combine into a single zone?

The combination rule has three stages, and the middle one is the part that most level-stacking tools skip. Clustering groups nearby levels, a latch gate demands a live catalyst, and only then is a band promoted to a zone.

1. Clustering. Events in the LEVEL category that are active at the moment being evaluated are sorted by price and greedily grouped within a relative tolerance of the running cluster anchor.
2. The latch gate. A directional event — a swept liquidity pool, an absorption print, a funding extreme — becomes a latch: a price memory that forms at the event's own known_at and breaks at the first closed candle through a price offset from it by a fraction of the causal ATR. A cluster is promoted to a zone only if at least one latch is still active and in range of the band. No live catalyst, no zone.
3. Emission. Direction is geometric — support if the current price sits above the band's midpoint, resistance if below. The zone's formed_at is the earliest gating latch's formation time, and its invalidation is the tightest break price among the aligned latches, falling back to the band edge padded by a fraction of ATR.

The latch gate replaces a bare "at least N sources" threshold, and it does so for a structural reason: it guarantees every zone spans at least two independent categories of evidence without needing a second counting rule. Zones are ranked by the number of distinct categories present.

What the eleven context bits record

Each zone carries a context key: eleven binary flags describing the conditions under which it formed. The bits are not weights and they do not score anything. They exist so hold rates can be split by context after the fact rather than assumed in advance.

Every bit is evaluated at the zone's own formed_at using lookups that admit only points already knowable at that instant. How the regime-flavoured bits are defined is covered in the market regime index methodology.

How is a zone's hold rate measured?

By replaying the engine, not by inspecting the chart afterwards. A backtest walks every latch formation time, rebuilds the zones exactly as they would have existed at that instant, and scores the forward outcome with a triple-barrier test sized in R from the zone's own invalidation price.

Outcomes are grouped by timeframe, direction and context key. Each group gets a raw win rate, a Wilson interval, and — because a two-sample group reading one hundred percent is not evidence of anything — an empirical-Bayes posterior. A Beta prior is fitted once per sweep by method of moments across the global distribution of group win rates, then each group updates that same prior with its own wins and losses. Thin groups shrink hard toward the global mean; well-sampled groups barely move.

Display prefers the shrunk estimate, and marks it as approximate when it has been materially pulled. This matters more than any single formula: it is the difference between reporting a measurement and reporting a coincidence.

What the engine deliberately does not produce

It produces no positions, no orders, no forecasts and no price objectives. There is no field in a zone that says what will happen. The trade-plan module that once existed for that purpose is deprecated and is not on the broadcast path.

It also refuses to fabricate the data it lacks. Layers with insufficient history return nothing rather than a neutral default — the Hurst estimator returns no value during warm-up instead of a manufactured mid-point, the auction-state machine emits nothing on the first day because there is no prior value area to compare against, and book-derived layers stay blank where nothing was recorded. Where a diagnosis is probabilistic because the book is sampled rather than complete, it is labelled probabilistic.

How to check this methodology

The claims above are structural, so they are checkable without access to our data. Watch the free delayed stream and note that every thesis is stated with its invalidation before the outcome, not after. Check that the invalidation price is a level to watch, never an instruction to act. Check that when a zone fails, the failure is said out loud.

Then check the arithmetic you can do yourself: a hold rate quoted with a sample size, an interval, and a timeframe is a measurement; the same number quoted bare is marketing. The companion pages on how AI analyzes order flow and the order-flow report methodology describe the input layers in the same terms. Coverage per instrument is listed under markets.