SECTION 1 — FEATURES

Everything the product is, module by module, formula by formula, parameter by parameter.

1. Introduction & Philosophy

1.1 What Advanced MACD Pro is

Advanced MACD Pro lives in its own oscillator pane under the price chart. It plots a MACD line, a signal line, a zero reference line and a histogram — the familiar four — and then does four things that the classic indicator does not do:

• It adapts. Every period and window is derived from the resolved bar duration through an explicit scaling law (§3). There is no lookup table of “M1 values” and “H4 values”: one formula, one anchor (H1), one curve. Move the indicator to another timeframe and the whole parameter set re-derives itself, including the divergence geometry, the auxiliary indicator periods, the volatility baseline and the drawing budgets.
• It reads divergences properly. Pivots are prominence-filtered fractals that publish only after their right-hand leg has closed; every admissible pairing inside the distance window is scored and the strongest one wins (not just “compare with the previous pivot”); regular and hidden families are graded A/B/C, confirmed against RSI and the histogram, pushed through eight hard gates, and given a 0–100 quality score built from thirteen weighted components (§6).
• It listens to higher timeframes. A confluence ladder of up to three higher timeframes is built, and each level gets its own fully adaptive profile — the ladder is not the primary setting re-read on a coarser series, it is what the indicator would compute if you had opened that chart yourself (§9).
• It measures itself. Every registered divergence becomes a record with an entry price taken at the confirmation bar (never at the pivot, which would bake in look-ahead), an ATR-denominated target and stop, and a horizon. Outcomes resolve incrementally on closed bars only, are bucketed per signal family, and the measured win rate is fed back as one of the thirteen quality components. The counters persist in the platform's LocalStorage per symbol and profile (§7).
• It manages the trade. Every registered signal becomes an explicit plan — entry at the confirmation bar, a stop from pivot structure or from the ATR ruler, two R-multiple targets — drawn on the price chart and resolved on closed source bars through RUNNING → TP1 HIT → TP2 HIT | STOPPED | STALE, with break-even and chandelier-trailing suggestions and a CSV journal of every event (§13, §16).
• It draws momentum structure. Support and resistance trendlines through the confirmed pivots' own MACD readings, with non-repainting break detection and scale-free gates measured in the oscillator's RMS units — the lead signal the MACD family is oldest known for (§14).

1.2 The philosophy in one paragraph

1.3 Feature summary at a glance

1.4 What the implementation is built on

The source cites its references explicitly; this is what each one contributes.

1.5 What it does not do

2. Architecture & Guarantees

2.1 Two index spaces, and why that matters

The indicator can run on a source timeframe that is not the chart timeframe (Source Timeframe). That creates two index spaces, and confusing them is the classic source of multi-timeframe bugs:

Writing a source index into a chart-indexed IndicatorDataSeries is unsafe whenever the source is finer than the chart (more source bars than chart bars) — it silently writes to the wrong bar or out of range. This indicator therefore keeps the mathematics in its own buffers and only ever writes final, chart-indexed values to the outputs, through the mapping functions of §2.3.

2.2 The computation pipeline, in the exact order it runs

Inside the MACD pipeline, one call to ComputeRange(from, to) performs, in this order:

1. Price input — build the oscillator's source series for the range (close / typical / weighted / median / Heikin-Ashi / volume-price-adjusted, §5.2). Missing components fall back to the close so one bad bar cannot blank the oscillator for a whole window.
2. Efficiency ratio — raw Kaufman ER per bar, then smoothed by an EMA over 2 × ER Period bars. This is the regime weight.
3. MACD line — either one fast/slow pair, or (with the filter bank on) the weighted blend of a responsive and a smooth pair. The blend is defined only when both legs are defined.
4. |MACD| and its rolling mean — Scale, a simple MA over NormWindow = clamp(4×Slow, 40..600) bars. This is the oscillator's own yardstick.
5. Signal line — a moving average of the MACD line, over the resolved signal period, with the kernel chosen by Signal Line MA Type.
6. Histogram = MACD − signal; then |histogram| and its rolling mean (HistScale), which is the unit the early-entry band and the “STRONG/NORMAL/WEAK” panel row are measured in.
7. Cross latch — the sequential ±1 state machine behind every crossover signal (§5.6, §8).

Then, once per chart bar, Calculate(index) runs the outer pipeline:

1. reload / shrink detection → full rebuild if history changed shape (§2.6);
2. Sync(true) on the primary pipeline: append new source bars, re-evaluate the forming one;
3. rewrite the just-closed chart bar with the strict mapping (§2.3);
4. map this chart bar onto the source series and write the outputs;
5. crossover markers (§8);
6. pivots, divergences, momentum trendlines and the plan lifecycle (§6, §13, §14) — deliberately before the confluence / volatility / outcome updates, because the pivot scanner advances those three states itself to their canonical per-confirmation point, which is what makes a bulk history pass bitwise identical to the live pass. The trendline engine and the plan lifecycle resolve on the same closed-bar footing;
7. MTF confluence (§9), volatility baseline, outcome tracking (§7);
8. the draw window is re-enforced between registrations, so a plan survives its record losing a drawing rank (§13.7);
9. on the last chart bar only: alerts (§11), the plan-extension pass (once per closed chart bar, gated on each plan's own source-freshness watermark), the session-marker sweep (§15), panel refresh (§10), and arming of the alert latches at the end of the initial history pass.

2.3 Chart → source mapping

“Closed by T” is answered as “opened at or before T − sourceSpan”, which the platform's time series resolves in O(log n). When the cutoff lands inside a non-trading gap (weekend, session break) the lookup can round up to the next bar, so the mapper walks back to the last bar that really opened at or before the cutoff. The same rule, with the same guard, maps the higher-timeframe ladder (§9.4).

2.4 The guarantees

2.5 What is allowed to move — honestly

“Non-repainting” is only a meaningful claim if it says what it covers. This is the complete list of things that legitimately change while a bar is forming, and everything else is final:

2.6 History reloads, symbol switches, reconnects

The indicator subscribes to the source series' Reloaded and HistoryLoaded events (once per series, de-duplicated by reference, because a confluence level can legitimately point at the very same series as the primary source) and additionally compares the source bar count on every Calculate. Either trigger sets a rebuild flag, and the next pass runs RebuildAfterHistoryChange:

1. Re-derive the timeframe geometry (bar spans), because everything below maps indices through it.
2. Invalidate and re-sync the primary pipeline; invalidate and re-sync every confluence level.
3. Clear pivots, divergences (removing their chart objects first), cross markers, the volatility model, the deduplication slots and every alert latch — including the confluence sign latch and any deferred alert.
4. Reset the learning counters. They are a statistic of the records that were just discarded; keeping them would count the same market events again on every reload and inflate the sample counts.
5. Rewrite the whole output series (RecomputeAllOutputs), because historical chart bars are never visited again by Calculate — merely clearing them would strand every marker older than the live edge.
6. Re-arm the history pass, so a crossover that already happened while loading cannot fire an alert afterwards.

2.7 Cost model

3. Timeframe Resolution & the Adaptive Engine

3.1 How the source timeframe is resolved

Source Timeframe offers CurrentChart, twelve fixed timeframes (M1, M5, M15, M30, H1, H2, H4, H6, H12, D1, W1, MN) and Custom. Every name is resolved through the platform's own parser (TimeFrame.Parse) rather than a hard-coded table, so the set of usable timeframes can never drift from what the running build of cTrader actually supports.

Custom Timeframe Name accepts, in this order of attempts:

1. a canonical platform name — Minute15, Hour4, Renko10, Range50, Tick1000…
2. trader shorthand — M1…M10080, H1…H24, D1…D7, T1…T1000, W/W1, MN/MN1/MO, RENKOn, RANGEn;
3. Heikin-Ashi variants of any of the above — HeikinH4, HeikinAshiMinute15…;
4. the literal words current / chart, which mean “use the chart”.

Case, spaces and punctuation are ignored. A name the platform rejects falls back to the chart timeframe and is reported: WARNING: Custom Timeframe '…' is not a timeframe this platform recognises, with the list of valid forms. With Debug Log on, the exact reason the parser refused is printed too — useful for telling “my spelling was wrong” from “this build has no such timeframe”.

3.2 How the bar duration is measured

The scaling law needs the bar duration in minutes. TimeFrame.Size is internal to the platform API, so the duration is derived from the timeframe's public name, with an empirical fallback:

3.3 The scaling law

One formula, no lookup tables. With D = source bar duration in minutes, D_ref = 60 (H1), ε = Microstructure Noise Horizon (min), δ = the horizon exponent of Timeframe Scaling:

The two terms do different jobs. The noise term keeps the statistical reliability of the smoothed difference constant: per-bar variance is σ²·D + σ²eps, so the number of bars that must be averaged to reach a fixed signal-to-noise ratio grows as sqrt(1 + eps/D). The horizon term moves the indicator between “constant bar count” (δ = 0) and “constant wall-clock horizon” (δ = 1). The default δ = 0.15 exists because the noise term alone under-corrects in practice on fast charts.

A period is then scaled with a per-parameter exponent, because not every parameter is horizon-driven: ScaleBars(ref, scale, exponent, min, max, gain) computes round(ref × scaleexponent × gain) and clamps to [min, max]. The gain is 1 / Global Sensitivity, which is why sensitivity divides periods: 2.0 gives half the periods (twice as fast), 0.5 doubles them, and the label agrees with the effect on every timeframe.

Two thresholds are scaled as values rather than bar counts:

The first follows the diffusion of ATR-as-a-percent-of-price with the square root of bar duration; the second keeps the slope gate statistically comparable when the number of bars it is measured over changes.

3.4 Wall-clock caps (automatic mode only)

Bar counts are the natural unit for oscillator periods, but windows that describe a market phenomenon must not grow without bound on slow charts — a 60-bar outcome horizon on a monthly chart is five years, which no trader will ever see resolved. Each is capped by BarsForDays(days, barMinutes, floor):

3.5 Derived quantities you cannot set directly

3.6 What the defaults actually resolve to

Computed with the shipped defaults (Adaptation = Automatic, Timeframe Scaling = Balanced, ε = 1.0 min, Global Sensitivity = 1.0, High Timeframe Policy = Canonical, Filter Bank Spread = 0.35, reference set 12/26/9). This is the same arithmetic the indicator performs at start-up and prints to the log — reproduced here so you can see the curve rather than trust it.

3.7 Invariants that hold in both modes

3.8 Manual mode

Adaptation Mode = Manual honours every dialog value verbatim, clamped only to the range the parameter itself declares. Nothing is scaled, no wall-clock cap applies, and the quality thresholds are taken as written (the scale-free bias of §3.5 is an automatic-mode feature). The derived quantities of §3.5 are still derived — they are not parameters — and the invariants of §3.7 still hold, because they protect against configurations that cannot mean what they say.

4. Trading Styles

Trading Style is the master switch and sits first in the dialog on purpose. It is the orthogonal axis to timeframe: it sets everything a timeframe cannot tell the engine, and it deliberately does not touch the timeframe. That is what makes one style work on M1 and on H1 alike — the adaptive engine keeps doing its job underneath.

4.1 Universal (default)

Writes nothing. Every parameter is whatever you set, and the start-up note about a style being active is not printed.

4.2 Scalping

Writes 134 of the 162 parameters, in one pass, before anything is resolved — 47 of them differ from the Universal defaults, and §32.8 lists every one of them. The design intent, group by group:

4.3 Swing

The balanced classic configuration for H4–D1 work: the Appel 12/26/9 triple on the close, the filter bank off (a swing horizon does not need the responsive/smooth blend), a slightly longer lambda (0.96) so the early-entry band is a nudge rather than a trigger, structure stops and the default 1R / 2R plan, and a session gate taken from the symbol's own calendar. It pins 134 parameters, of which 17 differ from the Universal defaults:

4.4 Position

The slow, maximally selective configuration for W1–MN work. It moves the references themselves (18/39/12, RSI 20, ATR 20, trend EMA 100), switches the scaling law to TimeHorizon with HorizonMatched high-timeframe policy so a weekly chart measures a comparable wall-clock horizon, keeps regular divergences only (a hidden continuation pattern on a monthly chart is a different trade than the one this style is for), demands a bigger price leg (1.5 ATR) and a longer outcome horizon (120 bars, 2.0 ATR target and stop), and turns the confluence ladder into a hard gate with three rungs at Min Confluence Score 45. Plans are wide (1.5R / 3.0R, buffer 0.75 ATR, trail 4.0 ATR, at most 2 visible), crossover markers are off — on a weekly chart a cross is noise next to a grade-A divergence — and so are the session markers. It pins 134 parameters, of which 61 differ from the Universal defaults:

5. The MACD Core

5.1 Structure

The core is a single self-contained class instantiated once for the source timeframe and once per confluence level, so the indicator behaves identically on every timeframe — each instance with its own automatically derived profile. Every kernel in it is a strictly causal recursion: the value at index i depends only on inputs at ≤ i and on already-computed outputs at < i. That property is what makes an exact incremental update possible, so re-evaluating the forming bar on every tick reproduces exactly what a full recomputation would produce — no drift, no O(history) rescan.

5.2 Price inputs

The volume-price-adjusted input, step by step:

This is the scale-preserving form of the paper's idea: σ and the body ratio are normalised against their own rolling means and used as conviction weights, not as multiplicative price factors. The adjusted series therefore stays in price units and the resulting MACD remains directly comparable to the classic one — which matters, because every threshold in the divergence engine is denominated in ATR or in the oscillator's own RMS. Robustness details: a bar with no usable price contributes nothing rather than poisoning the average; volume that is missing, zero or negative is folded onto the neutral 1.0; and the whole scan is memoised on the forming bar's inputs, so a stream of ticks that change nothing costs O(1).

5.3 The kernel library

Ten moving-average kernels, each implemented from its definition (composite kernels are compositions of primitive ones, which keeps them provably consistent). “Warm-up” is the number of bars before the first non-NaN output — the same quantity EstimateWarmupBars uses to warn you about insufficient history.

Two implementation details worth knowing, because they are the difference between an indicator that survives a bad bar and one that goes blank:

• Missing components are substituted, not propagated. Every price component is read through a guard that replaces NaN/∞ with the close, so one broken bar cannot blank a window-kernel oscillator for n bars (and cannot freeze a recursive kernel at a stale level, which is the quieter and worse failure).
• A NaN inside a window aborts that bar rather than being skipped: an efficiency ratio or a weighted mean computed over a truncated path is not a slightly-wrong number, it is a wrong number in the direction of “more confident”.
• Trendline breaks and plan lifecycles confirm on closed source bars only. A forming bar can neither break a momentum line nor touch a plan level, which is why the trendline engine needs no “closed bars” toggle. The start-up line states the whole contract: ready on … Non-repainting core: divergences, trendline breaks and plan lifecycles confirm on closed bars only.
• A bar that touches both the stop and a target fills the stop first. OHLC cannot reveal the intrabar order, so the pessimistic fill is the honest one; the plan engine and the learning engine apply the same rule (§13.5, §7).
• A plan survives the drawing budgets. Losing a Max Drawn Divergences rank or falling out of the draw-recent window removes a record's graphics only; its plan keeps being managed, alerted and journalled. Max Visible Trade Plans is the one knob that governs plan visibility (§13.7).

5.4 The signal line's own kernel

The signal line is a moving average of the MACD line, and its kernel is a separate choice (Signal Line MA Type):

5.5 The regime-adaptive filter bank

With Regime-Adaptive Filter Bank = true (default) and a non-zero Filter Bank Spread (default 0.35), the MACD line is not one oscillator but a bank of two, mixed by the market's own regime:

w → 1 means an efficient, trending regime, where the responsive leg should dominate; w → 0 means chop, where the smooth leg suppresses the whipsaw. The raw ER is far too twitchy to drive this on its own, which is why it is smoothed over 2 × ER Period bars (a derived quantity, §3.5).

The confluence ladder runs with the bank off at every level, by design: the ladder reports direction, not responsiveness, and a regime-mixed score would be hard to compare across timeframes.

5.6 Signal Sensitivity Lambda and the crossover latch

Signal Sensitivity Lambda ∈ [0.80, 1.00] controls how early a crossover is announced. At λ = 1 (default) the band collapses to zero and the latch reproduces the classic crossing exactly. Below 1, the trigger is advanced to the bar on which the histogram climbs out of the band on its way to a crossing:

The unit matters. Measuring the band in the oscillator's own mean absolute histogram is what makes λ mean the same thing on M1 and on W1, on EURUSD and on Bitcoin: λ = 0.92 always means “trigger once the histogram has closed 8% of its typical distance to zero”. This is also why λ is deliberately not part of the adaptive profile — scaling it on top of an already scale-free quantity would adapt it twice.

Two rules keep the advance honest:

1. Direction. An early trigger may only fire in the direction the histogram is leaving, never the one it currently occupies — otherwise the latch would flip back on the next bar and emit BUY/BUY or SELL/SELL pairs out of pure noise inside the band.
2. Consumption. An early trigger is consumed by the crossing it anticipated, so the crossing itself does not emit a second one.

The result is exactly one trigger per oscillator excursion. That cannot be guaranteed by a local two-bar test once the band is wider than zero: the bullish band edge (−band) then sits below the bearish one (+band), so the two conditions overlap and a difference that dwells inside the band and then crosses zero satisfies neither. The latch is a strict causal recursion, so re-evaluating the forming bar reproduces exactly what a full recomputation would produce.

Three further details of the latch, each of which prevents a specific class of phantom signal:

• Numerical noise floor. Differences within 1e-12 × |price| are snapped to exact zero. On a dead-flat tape the two MA chains accumulate independent rounding error and the histogram jitters around zero by a few ulps; without the floor, strict inequalities would fabricate crossovers — phantom markers and alerts on a completely flat market. The floor is derived from the price scale, never from an absolute constant, so it behaves identically on an instrument quoted at 1e-6 and one quoted at 1e6. It deliberately has no absolute component: an absolute 1e-9 floor is larger than the entire histogram of any instrument quoted below ~1e-3, and would silence the indicator permanently on crypto crosses, spread and rate instruments.
• Hysteresis at exactly zero. A bar that lands on zero inherits the previous side, so a flat tape cannot fabricate crossings.
• First bar with a histogram adopts the side the oscillator is already on and emits no trigger: there is no previous state to have moved away from.

5.7 Output scale and gain

Output Scale chooses the units the plotted series is expressed in; Output Gain (0.05–50) is a plain multiplier applied on top. Every plotted value — MACD, signal, all four histogram buckets, both trigger series and both divergence series — goes through the same conversion, so the pane stays self-consistent.

5.8 The four histogram buckets

The histogram is not one series but four, so that momentum quality is visible as colour rather than as a number you have to compute in your head. A bar is strong when the histogram is both expanding (|H_i| > |H_{i-1}|) and at or above its own average size (|H_i| ≥ HistScale_i):

All four buckets are cleared before every write, so a live bar whose classification flips between ticks can never leave two buckets populated for the same bar. The panel's Momentum (BUILDING / FADING / STEADY) and Strength (STRONG / NORMAL / WEAK) rows are the same two tests, spelled out.

6. The Divergence Engine

6.1 The path from a bar to a drawn signal

1. Confirm a pivot. Every source bar that has just become confirmable is tested for a significant fractal pivot (§6.2). “Confirmable” means its right-hand window consists only of closed bars.
2. Search the pool. The new pivot is paired against every older pivot of the same side inside [MinPivotDistance, MaxPivotDistance], up to Max Pivot Candidates pairings, and each pairing is scored. The strongest one wins — this is the single largest accuracy improvement available to a pivot-based detector, because divergences that span an intermediate pivot are no longer missed (§6.3).
3. Classify and grade. Regular or hidden, bullish or bearish, A/B/C, with the zero-line rule and the two confirmations applied on the way (§6.4–6.5).
4. Run the hard gates at the confirmation bar (§6.6). A rejection is logged with its reason when Debug Log is on.
5. Score the quality — thirteen weighted components, 0–100 (§6.7). Below Min Quality Score the candidate is dropped and the drop is logged.
6. Register an immutable record, claim the de-duplication slot, draw it if it is inside the drawing window, place the trigger marker at the confirmation bar, mark the panel dirty and raise the alert if the quality clears the alert bar (§6.8).
7. Track the outcome on closed bars only, and feed the family's win rate back into future quality scores (§7).

6.2 Pivot detection

A pivot at bar c is a fractal extremum over Pivot Lookback bars on each side, filtered by prominence.

Each pivot record carries its index, open time, extreme price, MACD, histogram, RSI and prominence in ATR units. Confirmed pivots are kept per side in a pool whose capacity follows the resolved profile:

It must comfortably exceed the maximum pivot separation, otherwise the multi-candidate search would silently lose its oldest candidates — which is why it is derived rather than a fixed constant.

ATR for prominence comes from the platform's own ATR indicator on the source series; while that is still warming up, the mean true range of the bars that do exist is used instead, so pivots found early are still measurable rather than being silently skipped.

6.3 Multi-candidate search and the pairing score

The pool is walked from the newest pivot backwards. A pairing is skipped when the distance is below MinPivotDistance; the walk stops when it exceeds MaxPivotDistance; at most Max Pivot Candidates pairings are evaluated. Every surviving pairing is assessed, and the one with the highest Score is taken forward. The score is a weighted sum, deliberately not normalised (it only ever compares candidates against each other):

DistanceFit is a trapezoidal membership function of the pivot separation — very tight pairs are usually one swing split in two, very wide pairs lose causal connection:

The confirmation bar of a candidate is min(pivot + PivotLookback, lastSourceBar) — the first bar on which the pivot could have been known. Everything below (gates, quality, entry price, alerts) is evaluated at that bar, never at the pivot.

6.4 Regular, hidden, and the A/B/C grade

A low pivot always produces a bullish signal and a high pivot a bearish one — for regular (reversal) and hidden (continuation) families alike. What differs is which leg moved against which:

Divergence Kinds selects which families are allowed (RegularOnly, HiddenOnly, both by default).

The zero-line rule (Appel / Brown): with Require Zero-Line Match on, a regular bearish divergence must be built entirely above zero and a regular bullish one entirely below zero. Hidden divergences are exempt — the rule is about exhaustion at an extreme, and a continuation pattern is not one. The depth inside the correct zone is measured as ZoneDepth (in units of the oscillator's own scale, clamped to 0…3) and contributes to the quality score.

The grade separates genuine momentum exhaustion from a shallow diagonal:

6.5 The two confirmations

6.6 The hard gates, in the order they run

6.7 The quality score

A weighted, bounded 0–100 score. Every component is normalised against a scale-free reference (ATR for price, the rolling mean absolute MACD for the oscillator), so the score means the same thing on EURUSD M1 and on XAUUSD Weekly. Components that cannot be evaluated are dropped from both the numerator and the denominator rather than being replaced by a neutral value, so an unavailable input can never inflate or deflate the result: quality = 100 × Σ(w·v) / Σw over the available components.

Two thresholds act on the score: below Min Quality Score the candidate is dropped entirely; below Weak Signal Threshold it is drawn in the weak colour (amber by default) instead of its family colour. Alert Min Quality is a third, higher bar that only gates alerts (§11). In automatic mode all three carry the ±10% timeframe bias of §3.5.

6.8 What a registered divergence remembers

A pivot that predates the loaded chart history cannot be anchored, so its chart index stays −1. The record is still registered — outcomes, learning, the panel census and the alert all key off source indices — and drawing simply refuses to paint anything with a negative anchor. Dropping the record instead would silently delete those signals too. (An earlier version clamped the index to 0, which drew the divergence from the very first bar of the chart: a line spanning the whole pane, anchored on a bar that has nothing to do with the signal, and impossible to distinguish from a real one.)

6.9 Drawing

7. Outcomes & Adaptive Learning

7.1 What is measured

Every registered divergence becomes a measured trade hypothesis, denominated in ATR so that one set of numbers means the same thing on every instrument and timeframe:

• Intrabar extremes, not closes. A target touched by a wick counts as touched, because that is what a real limit order would have done.
• Closed bars only. The scan runs to currentSourceIndex − 1; the forming bar is never evidence.
• Latched and incremental. Each record remembers ScannedTo, so the total cost over the whole history is O(bars) instead of O(bars × horizon), and a later wick can never un-resolve an outcome.
• Target wins ties. When a single bar touches both levels, the favourable one is credited if it is at least as large — the optimistic reading, stated here rather than buried.
• Unmeasurable records are marked expired (no ATR, no entry price) rather than being left pending forever or being counted as a win.
• The scan is skipped entirely unless a new source bar has closed since the last one, so a chart finer than the source pays nothing per tick.

7.2 The four buckets

Outcomes are accumulated per signal family, because a regular bullish divergence and a hidden bearish one are not the same hypothesis and should not share a track record:

Each bucket stores wins, losses and expired counts; the win rate is wins / (wins + losses) — expired signals are neither wins nor losses, they are signals that did not resolve, and folding them into either side would invent a result.

7.3 How it feeds back

When Adaptive Learning is on and a bucket has at least Min Learning Samples resolved outcomes, that bucket's win rate enters the quality score as component 13 with weight 0.08 (§6.7). Below that sample count the component is dropped from numerator and denominator, not replaced by 0.5: a track record of two trades is not evidence, and treating it as neutral-but-counted would let a lucky pair drag every subsequent score.

7.4 Persistence

8. Crossover Signals & Markers

8.1 What a crossover signal is

A trigger is the transition of the latch described in §5.6: +1 on a bullish flip, −1 on a bearish one, read from the sequential state machine rather than from a two-bar comparison. DetectSignalCross(s) answers whether the latch flipped at source bar s. With λ = 1 this is exactly the classic MACD/signal crossing; below 1 it is the advanced trigger.

A separate detector watches the zero line (DetectZeroCross), with the same noise floor and the same hysteresis at exactly zero, so a flat tape cannot fabricate zero crosses out of ulps. Zero-line markers and zero-line alerts are both off by default.

8.2 What gets drawn

Zero-line crosses use the neutral wording ABOVE 0 / BELOW 0 so they can never be mistaken for a crossover signal. Since v2.0 the family the markers follow is a choice, not a checkbox: Cross Mode = SignalLine (default), ZeroLine or SignalAndZero (§17). The trigger outputs stay bound to the signal-line latch whichever mode is selected, and the two alert toggles of group 10 stay independent of it. The hover tooltip of every marker carries the full detail: signal kind (and the λ in force when the band is open), the MTF confluence score, the oscillator / signal / histogram values and the source bar's UTC timestamp.

Label placement is not a fixed offset. Clearance is computed in the oscillator's own units (the rolling mean absolute histogram, falling back to the mean absolute MACD and then to |MACD| itself), multiplied by Marker Label Offset, converted into whatever scale the outputs are plotted in, then limited to 40% of the visible band and stepped toward the zero axis. Stepping toward zero is guaranteed to stay inside the band the pane displays; a “bullish up / bearish down” rule collapses onto the marker point whenever the cross sits on the same side of zero as its direction — the common case in trends — where the label ends up painted on the arrow inside a same-colour histogram cluster and reads as a missing label.

8.3 Bookkeeping rules

• Keyed by source index, so a cross can never be marked twice and can never be missed when several chart bars share one source bar.
• Deterministic anchor: the marker always sits on the chart bar that contains its source bar's open time — identical live, on a replay and after a reload. Anchoring the forming bar to the “current” chart bar instead made the position depend on when the cross first became visible intrabar: a replay and the live pass disagreed about the newest marker, and on charts finer than the source the marker slid rightwards while the bar formed.
• Closed bars are back-filled exactly once, in order, and the back-fill is bounded by DrawRecentBars + 2 × MaxCrossMarkers bars — anything older would be trimmed immediately anyway, and on a deep history an unbounded walk meant a multi-second stall with every object alive at once. Trimming also runs during the walk, so peak memory stays flat.
• The forming bar is provisional: re-evaluated on every tick, and its marker is withdrawn if the cross does not survive to the close.
• Bounded by count (Max Cross Markers) and by the recent-history window.
• Switching the markers off does not lose history: the back-fill watermark is parked at “nothing built”, so turning them back on rebuilds the visible window instead of waiting for a history reload.
• After a reload everything is rebuilt from scratch, so markers can never be stranded or lost.

8.4 The trigger outputs

Bull Trigger and Bear Trigger are plotted at the MACD value of the triggering bar, so the dot sits exactly on the oscillator, and they are exposed as outputs so cBots can consume them (§12).

9. Multi-Timeframe Confluence Ladder

9.1 How the ladder is built

Two ways, and an explicit list always wins:

9.2 The state score

Each level (and the primary) is reduced to a single number in −100…+100 by three signed tests:

The ladder is then a weighted mean, with the primary timeframe carrying the most weight:

A level that has fewer than three bars, no mapping, or no oscillator value yet is marked unavailable and contributes nothing — including nothing to the denominator, so a missing level cannot drag the score toward zero.

Textual states, used by the panel and the alert: BULL at ≥ +33, BEAR at ≤ −33, otherwise FLAT / NEUTRAL.

9.3 Mapping a source bar onto a higher timeframe

The same closed-bar rule as the primary mapping: “the last HTF bar that had closed by sourceOpenTime + sourceSpan”, answered as “opened at or before asOf − htfSpan”, with a walk-back guard for cutoffs that land inside a non-trading gap (the platform's lookup can round up there, and that bar had not closed).

Strict Non-Repaint MTF Mapping (default on) additionally excludes the newest HTF bar entirely. This matters on the live bar, where asOf is a moment in the future: without the exclusion the ladder quotes a higher-timeframe bar that is still forming, so the confluence score moves while the source bar is open — and after a history reload the same divergence would be re-scored against a now-closed HTF bar, i.e. a registered signal's quality would change retrospectively. With it, every quoted value is final and a replay reproduces it exactly. Turn it off if you want the ladder to follow the higher timeframe tick by tick and accept both consequences.

9.4 Where the confluence score is used

10. The Info Panel

10.1 Why it is built from real controls

A single DrawStaticText block is why info panels collide with the plot: it is one opaque string anchored to a corner whose height is unbounded, so it grows over the oscillator and cannot be clipped by anything. This panel is built from real chart controls — Border cards containing StackPanel rows of TextBlock cells — pinned to a corner of the indicator window, with a static-text fallback for builds that cannot host controls.

10.2 The four layouts

Panel Side chooses left / right / centre; Vertical Edge chooses which vertical edge is “primary” (the other section goes to the opposite edge in split mode).

10.3 Geometry — the guarantee that it cannot cover the oscillator

• Rows that fit are computed from the pane's real height minus margins, padding, border and the separator, divided by the row height implied by Panel Font Size. Zero is a legal answer: when even one row cannot fit, the card is hidden rather than allowed to spill outside the pane.
• Columns that fit are computed from 78% of the pane width and the cell width, clamped to the built cell count.
• In merged mode the row budget is shared between the two sections in proportion to what each actually wants to show, guaranteeing at least one row to whichever section has content, and hard-clamped so that no section can be asked to fill more rows than row controls exist (that would take the whole render pass down with an index error).
• Priority selection: every item carries a priority (0 = must show, 1 = should, 2 = nice to have); the lowest priorities win, ties keep natural reading order, and the rest are dropped for this render. Shrinking the pane therefore removes the least important cells instead of overflowing.
• The panel re-fits on the pane's SizeChanged event and refreshes no more often than Panel Refresh (ms), and only when something is actually dirty.
• If native controls are unavailable or fail, everything degrades to two corner-anchored ChartStaticText objects with the same content and the same priority selection — the separator is then inserted between the two selected blocks, and carries the lowest possible priority so it can never win a slot ahead of real content.

10.4 The five sections

Every row is explained line by line in §21.

11. Alerts

11.1 What an alert is

Three independent channels, each with its own switch. cTrader gives an indicator no modal dialog of its own, so these are all the routes that exist:

Enable Alerts is the master switch for all three.

11.2 The six triggers

11.3 Latching, arming and anti-spam

• Nothing alerts during the initial history pass. Alerts are armed only when the first pass over the loaded history reaches the live edge, and the guard indices are seeded at that moment — so a crossover that happened while the chart was loading cannot fire an alert afterwards. A history reload re-arms the whole mechanism (§2.6).
• One alert per event. Crossover and zero-line alerts are latched per evaluated source index; divergence alerts per end-pivot index; confluence flips per regime change (the sign latch), with at most one confluence alert per source bar.
• A deferred flip is not a dropped flip. If the ladder flips twice inside one source bar, the second flip is announced on the next bar rather than silently discarded — and it is announced with the regime the sign latch actually tracks, so it can never announce a state that is no longer current.
• Divergences must be live. A divergence alert only fires when the confirmation bar is within two bars of the live edge, so re-registering an old signal after a rebuild cannot produce a stale alert.
• Divergences must clear the alert bar. Alert Min Quality is a separate, higher threshold than Min Quality Score: signals between the two are drawn and counted but stay silent. If your alert bar ends up below the quality gate, every passing signal alerts — and that is reported at start-up.
• Closed-bar mode. Only On Closed Source Bar (default on) evaluates the last closed source bar, which makes alerts strictly non-repainting. Turn it off to be told inside the bar — the Scalping style does, because a scalper acts inside the bar — and accept that an intrabar alert can be withdrawn by the close.
• Cooldown per family and direction. Alert Cooldown (source bars, 0 = off) enforces a minimum number of source bars between two alerts of the same family and the same direction; 0 (default) disables it. The suppression is final for the event that tripped it — the scan watermark advances regardless — which is what “minimum bars between alerts” means: the next event alerts normally once the cooldown has elapsed.
• Freshness is ratio-scaled. With a source finer than the chart, one Calculate sweep closes several source bars at once, so every freshness window in the alert path is multiplied by the source / chart bar ratio and floored at a 512-bar scan window. Without it a fine source on a coarse chart would silently swallow the very events whose markers, trigger outputs and journal lines were still produced.
• Plan and trendline alerts are history-gated. Like the journal, they fire on live events only: a start-up or reload rescan re-resolves every historical plan and break without re-announcing any of them (§13.9, §16.2).

12. Outputs & Programmatic Use

12.1 The twelve data series

12.2 Consuming them from a cBot or another indicator

// In a cBot or a plugin - the platform's standard way to instantiate another indicator.
// Overloads of Create<T> let you pass parameter values; the parameterless one uses defaults.
var amp = Indicators.Create<cAlgo.Indicators.AdvancedMacdPro>();

double macd   = amp.MacdLine[index];                 // MACD
double signal = amp.SignalLine[index];               // Signal
double hist   = macd - signal;                       // the four buckets are mutually exclusive
bool bullTrigger = !double.IsNaN(amp.BullTrigger[index]);           // crossover latch flipped bullish
bool bearTrigger = !double.IsNaN(amp.BearTrigger[index]);
bool bullDiv     = !double.IsNaN(amp.BullDivergenceOutput[index]);  // divergence confirmed on THIS bar
bool bearDiv     = !double.IsNaN(amp.BearDivergenceOutput[index]);

• All outputs are chart-indexed, i.e. indexed like the chart you attached the indicator to, even when the source timeframe is different. The mapping is the strict one of §2.3.
• NaN means “not available”: warm-up, an empty row because no source bar had closed, or “no event on this bar” for the four event series. Test with double.IsNaN, never against zero.
• Values are in display units: whatever Output Scale and Output Gain say. For a robot that compares magnitudes, use Raw and gain 1.0, or read the sign and the crossings rather than the level.
• The histogram is split across four series; to reconstruct it, take the first non-NaN of the four (they are mutually exclusive per bar).
• The four event series are the machine-readable channel: triggers are the latch transitions of §5.6 (so they respect λ), divergence dots are placed at the confirmation bar, which is the first bar on which the signal could have been acted upon — never at the pivot.

13. Trade Management — the Projection Engine

A registered divergence is a trade idea, not a picture. Version 2.0 closes the loop: at the moment a signal is registered the engine resolves an explicit plan — Entry / SL / TP1 / TP2 — draws it as labelled horizontal levels on the price chart, and then manages its lifecycle bar by bar on closed source bars only. The indicator still places no orders: a plan is a drawing, a set of alerts and a set of journal lines, and nothing else.

13.1 What a plan is, and when it is built

The plan is computed once, at registration, from the confirmation bar — never from the pivot, which would bake in look-ahead. Everything a later consumer needs (the drawing, the lifecycle, the alerts, the panel, the journal) reads the resolved levels from the record:

A plan is refused — silently, with no half-drawn levels — when the projection engine is off, when the ATR at entry is not a usable positive number, or when the entry price itself is not usable. “Usable” means finite: an infinite entry or ATR would otherwise produce infinite levels that the panel, the alerts and the lifecycle would all treat as a real plan.

13.2 The two stop rulers, and the band between them

Whatever the source, the plan risk is then clamped into a sanity band:

13.3 Targets: R-multiples of the plan risk

Because both targets are multiples of the plan's own risk, a plan whose reward is below its risk cannot be produced by the default settings, and the reward:risk reading is identical on EURUSD M1 and on XAUUSD W1.

13.4 Drawing: levels, labels, and the HIT / VOID / STALE styling

The four levels are drawn from the confirmation bar forward for Projection Length chart bars (default 120), each with its own label when Show Projection Labels is on. Colours: entry and stop from Entry Line Color / Stop Line Color, targets from the signal's direction colour. A live plan draws solid lines; the moment the plan leaves RUNNING the whole group is re-rendered dotted, with a suffix on every level, so a finished plan can never masquerade as a live one:

13.5 The lifecycle: five states and the conservative same-bar rule

Resolution runs on closed source bars only, and the intrabar-order problem is answered the pessimistic way: a bar that touches both the stop and a target fills the stop first. OHLC cannot reveal the order of touches inside a bar, so the conservative fill is the honest one — the same rule the outcome engine applies when it grades a learning sample (§7).

TP1 HIT is not terminal. The plan continues to TP2, to the stop, or to staleness, and the panel keeps showing that half the position already paid: a plan that hit TP1 and then stopped does not read as a plain STOPPED.

13.6 Break-even and the chandelier trail

With Show Break-Even & Trailing Hints on (default), a managed plan gains up to two extra lines:

13.7 Budgets, and why a plan survives them

Max Visible Trade Plans (default 3, range 1–10) caps how many plans are on the chart at once; the oldest lose their graphics first. A plan that loses its slot loses only its graphics — its lifecycle, its alerts and its journal lines continue, because visibility and management are deliberately separate budgets. The same is true of the divergence drawing budgets: when a record loses its Max Drawn Divergences rank or falls out of Draw Only Most Recent N Bars, only its oscillator graphics are removed; the price-chart plan stays managed. The one full-teardown path is the hard record cap.

13.8 Alerts, panel rows and the journal

13.9 The guarantees

• Non-repainting. A transition is decided once, on the closed bar that caused it, from levels fixed at registration. A forming bar can neither touch a level nor expire a plan.
• Reload-stable. A history reload re-resolves the same records to the same states; it never re-announces them. Alerts and AMP-PLAN lines are history-gated exactly like AMP-OUTCOME.
• Conservative. The same-bar rule fills the stop first, so a plan can never be reported as a winner on ambiguous data.
• Self-announcing. Contradictory settings are reconciled at start-up and reported: NOTE: Plan TP2 (x R) is not above Plan TP1 (y R), and NOTE: trade plans are generated from divergence signals; with Show Divergences off … (§18.2).

13.10 Parameters — group 12, Trade Management

All twelve parameters, with their defaults, are in the reference table (§17).

14. Momentum Trendlines

Two automatic lines in the oscillator pane: support and resistance, drawn through the momentum readings of the very pivots the divergence engine confirms. A momentum trendline breaking before price does is one of the oldest lead signals in the MACD family (Appel; Brown's momentum trendlines) — and the oscillator, not the price chart, is where that reading belongs.

14.1 What is drawn, and from which pivots

There is exactly one pivot truth in this indicator: the trendline engine consumes the confirmed pivot pools the divergence engine already publishes (§6.2), so a line never rests on a pivot that the prominence filter, the warm-up guard or the right-leg-closed rule would have rejected.

14.2 Scale-free gates: why everything is in “x RMS”

A MACD lives in price units, so “points per bar” would mean something different on every symbol. Both gates are therefore multiples of the oscillator's own yardstick — the rolling mean absolute MACD, the same “x RMS” stick the Min Oscillator Leg gate measures with — resolved at the anchor being tested:

The result behaves identically on EURUSD M1 and on XAUUSD W1, which is the whole point of the adaptive engine.

14.3 When the lines move — and when they never move again

• The lines extend toward the newest closed bar on a redraw throttle of max(10, 2 × Trendline Min Pivot Gap) closed bars.
• They redraw immediately on every re-anchor (a newer qualifying pivot replaces an anchor) and on every break.
• Break detection itself runs on closed source bars only, which is why the engine needs no “mark only closed bars” toggle: a break that is drawn has already happened on a bar that can no longer change.
• Each line generation announces itself once: the Announced flag is cleared only when the line re-anchors, so a bar-by-bar re-test cannot machine-gun the same break.

14.4 What you see and hear

14.5 Reading one break in five seconds

1. Which line broke? Support breaking is bearish momentum; resistance breaking is bullish.
2. Look at the histogram buckets on the same bar: a support break printed while the histogram is already bright red is a confirmation; the same break on a tired, muted histogram is the lead signal this engine exists to give you.
3. Check the ladder rows (§21.3): a momentum break against the higher timeframes is a warning, not an entry.
4. Only then look for the divergence or crossover that would actually trigger a plan (§13).

15. Session Markers

Thin vertical lines on the price chart at configurable UTC times — the London and New York opens by default. They are pure context: no gate, no score, no alert and no output reads them, ever. The signal logic and the session filter (§6.6, gate 7) are separate mechanisms and stay separate.

15.1 Syntax

Marker Times (UTC, HH:MM list) accepts a list separated by commas, semicolons or spaces; each entry is HH:MM or a bare hour. Default 07:00,13:30. Entries the parser cannot use are skipped individually rather than aborting the list; the parsed set is sorted ascending, so markers inside one wide bar are emitted (and trimmed) in chronological order.

15.2 Wide bars carry every occurrence

A Weekly or Monthly bar can contain several configured times. The object key is

so one wide bar draws one line per occurrence instead of silently dropping all but the first, while key / 1 000 000 is still exactly the chart index for the rolling-window trim. The key space is collision-free for any chart length the platform can hold, and needs no overflow guard.

15.3 Window and object cap

Because the sweep only ever advances over newly closed chart bars and re-runs from the window edge after a history rebuild, the layer costs nothing per tick and cannot leave stale lines behind when you scroll, switch symbol or change timeframe.

16. The CSV Signal Journal

Signal Journal (CSV lines to log) (group 11, default on) writes Excel-ready, semicolon-separated, invariant-culture lines into the platform log. Three streams, one join key: the signal's id is the first column of every line, so a signal, its lifecycle and its measured verdict can be joined in a spreadsheet without guessing. The header is printed once per instance, so a captured log is self-describing.

16.1 The three streams

16.2 History-gating: why a reload never re-emits

Only live events are journaled. A start-up or reload rescan re-resolves the whole history to the same records but emits nothing, so the CSV stays duplicate-free across restarts and a Monday-morning log is not a flood of last month's events. The same gate applies to the plan-level and divergence alerts (§11.3). The consequence to remember: the journal describes what happened while you were watching, not the whole chart history.

16.3 Getting it into Excel

1. Open the platform log, select the lines, copy.
2. Paste into a text file (or straight into a sheet) and split on ; — Text → Columns in Excel, or Data → Split text to columns in Sheets.
3. Keep the AMP-JOURNAL-HEADER row as your header; filter the rest by stream prefix.
4. Join AMP-PLAN and AMP-OUTCOME onto AMP-JOURNAL by id: you now have a per-signal blotter — the idea, its grade and quality, its plan, what the plan did, and what the learning ruler measured.

17. Complete Parameter Reference — 162 Parameters, 15 Groups

Every parameter, with its default and what it actually does. Defaults in bold, dialog-enforced ranges in parentheses. Names ending in “(reference / manual)” or “(reference)” are reference values anchored at H1 that automatic mode rescales (§3.3) — the numbers actually in force are printed at start-up and shown in the panel. All of them are cross-validated when the indicator initialises: contradictory combinations are reconciled to the safe side and reported (§18.2), and in manual mode any value that does not survive the profile invariants verbatim is reported as well.

Trading Style — 1 parameter

Timeframe & Source — 10 parameters

MACD Core — 11 parameters

Automatic Adaptation — 9 parameters

Divergence Detection — 16 parameters

Signal Filters — 21 parameters

Quality & Learning — 7 parameters

Signal Markers — 10 parameters

Divergence Appearance — 17 parameters

Info Panel — 19 parameters

Alerts — 15 parameters

Advanced — 3 parameters

Trade Management — 12 parameters

MACD Trendlines — 7 parameters

Session Markers — 4 parameters

18. Diagnostics & the Log

18.1 What is printed at start-up

With Log Resolved Profile On Start (default on) you get the resolved configuration — the numbers the indicator is actually running, read from the resolved settings object rather than from the dialog:

[Advanced MACD Pro] style=UNIVERSAL | source=h1 (60.00 min/bar) | mode=Automatic | scale=1.000
[Advanced MACD Pro] MACD 12/26/9 kernel=Exponential signal kernel=Exponential input=Close bank=on (+/-35%)
[Advanced MACD Pro] filter bank: responsive 8/17  smooth 16/35  spread=35%
[Advanced MACD Pro] early entry: lambda=1.00 (band 0% of the mean |histogram|)
[Advanced MACD Pro] pivots: lookback=5 prominence=0.35ATR distance=[10..100]
[Advanced MACD Pro] aux: RSI=14 ATR=14 ADX=14 TrendEMA=50 slopeBars=5 ER=10
[Advanced MACD Pro] gates: ATR=adaptive p20 = 0.0311% maxCounterSlope=0.150 minQ=50.0 weakQ=65.0
[Advanced MACD Pro] session=off | follow-through>=0.50 | volatility window=240 bars
[Advanced MACD Pro] windows: normalisation=104 stats=100 drawRecent=1500 bars | warm-up ~78 bars
[Advanced MACD Pro] outcome horizon=60 bars (target 1.0ATR / stop 1.5ATR)
[Advanced MACD Pro] trade plans: on | stop source: pivot structure | TP1 1.0R / TP2 2.0R | risk band 0.5-6 ATR | max visible 3
[Advanced MACD Pro] MACD trendlines: on | min gap 5 | max slope 0.25 RMS/bar | break threshold 0.10 RMS
[Advanced MACD Pro] plan lifecycle: alerts on | mgmt hints on (trail 3.0 ATR)
[Advanced MACD Pro] session markers: on (2 times, window 300 bars)
[Advanced MACD Pro] signal journal: on (AMP-JOURNAL / AMP-PLAN / AMP-OUTCOME lines)
[Advanced MACD Pro] MTF ladder: h2 h3

The warm-up line counts the deeper filter-bank leg (35) plus the signal kernel (9) plus a quarter of the normalisation window (26) plus 8. The ladder line lists the derived levels for an h1 source — the next two rungs up are h2 and h3; if you pin Confluence Timeframes = H4,D1 it lists those instead.

18.2 Every warning and note, and what it means

18.3 Debug Log, and what is not in it

Debug Log = true adds three things, and only three — it is not a firehose:

• One line per rejected candidate, with the reason. The format is [Skip {Bull|Bear}{Hidden|Regular}] {reason} q={quality} @ {bar time UTC}; a hard-gate rejection reports q=0.0 because the candidate never reached the scoring stage. Real examples: [Skip BullHidden] ATR% 0.0091 below floor 0.0311 q=0.0 @ 2026-10-10 14:00, [Skip BearRegular] trend too strong to fade (-0.412) q=0.0 @ …, [Skip BullRegular] quality below threshold q=47.3 @ …, [Skip BullRegular] clustered with a previous signal q=0.0 @ …, [Skip BearHidden] outside the trading session (2026-10-10 22:15 UTC) q=0.0 @ …, [Skip …] duplicate end pivot q=0.0, [Skip …] follow-through 0.31 < 0.55 q=0.0, [Skip …] ADX gate (14.2) q=0.0, [Skip …] MTF gate (-41) q=0.0, [Skip …] hidden vs trend (-0.183) q=0.0. This is the answer to “why hasn't it fired in three days?”.
• One line per registered divergence: [Div BullRegular A] q=82.4 priceLeg=1.31ATR oscLeg=0.27RMS dist=34 @ ….
• Learning restore and the timeframe parser's refusal reason (§3.1).

SECTION 2 — HOW TO USE

The working trader's guide: every glyph, every panel row, every decision.

19. Installation & First Launch

1. Get the source into cTrader. Open Automate → New → Indicator, paste the whole of AdvancedMacdPro.cs, and press Build. It compiles against cAlgo.API with no dependencies and no external packages. (A pre-built .algo can be installed by double-clicking it or by dragging it onto the chart.)
2. Attach it to a chart. It is an oscillator, so it opens in its own pane under the price chart — it is not an overlay (IsOverlay = false), and it declares AutoRescale = true and TimeZone = UTC.
3. Read the log. The first thing it prints is the resolved profile (§18.1) — the actual periods, windows and gates in force, not the dialog's reference values. If a parameter combination was contradictory you get a WARNING or a NOTE saying what was reconciled and to what. If history is too short for the resolved configuration, it says so and names the three fixes.
4. Give the pane some height. The panel fits itself to the pane: a tall pane shows more rows, a short one hides the low-priority cells and, if even one row cannot fit, hides the card rather than covering the plot.
5. Check the panel's TF row. It shows the resolved source timeframe and whether the profile is auto or manual. Everything else in the panel is derived from those two facts.
6. Let the warm-up finish. The oscillator needs its slow leg, the signal line and part of the normalisation window before it plots; the start-up log prints the estimate. The adaptive ATR gate stays inert (permissive) until its volatility model has enough samples — the panel's ATR row reads WARM until then.

20. Reading the Pane — Every Glyph

21. The Info Panel, Line by Line

Cells are colour-coded: bright = a verdict (bullish/bearish, OK/LOW, OPEN/CLOSED), muted = context with no verdict (labels, timestamps, counts). Rows that cannot be computed print - rather than a guess.

21.1 Configuration section

21.2 Market section

21.3 MTF section

21.4 Signals and statistics sections

22. The Parameter Dialog & Presets

22.1 The order to change things in

1. Trading Style — decide whether you want a hand-built configuration or the scalping one. Everything below is overwritten by a style, so choose it first.
2. Source Timeframe — the single most consequential choice. Leave it on CurrentChart unless you specifically want an oscillator computed on another series.
3. Adaptation Mode — keep Automatic unless you are deliberately testing fixed periods. In automatic mode treat every “(reference)” number as an anchor, not as a setting.
4. What you want to see — groups 4 (divergences), 7 (markers) and 8 (appearance). These are the ones you will actually tune.
5. How picky it is — group 5. Loosen here if you see nothing, tighten here if you see too much (§27).
6. Only then the oscillator internals (group 2) and the adaptive engine (group 3).

22.2 Saving what you built

Use the platform's own template mechanism (Save as template in the dialog) to keep a configuration, and remember two things: a template stores the values as shown, so if a style is active it stores the style's values — and a style re-applies itself on every initialisation regardless of what the dialog shows. To keep a tuned variant of the scalping setup, set Trading Style = Universal first, then save.

22.3 The Lines tab

All twelve outputs can be recoloured, restyled, thickened or hidden individually there (§12.1). Hiding Hist Bull Weak and Hist Bear Weak, for example, leaves only the strong buckets — a quick way to see momentum that is actually building. Hiding Zero Line removes the dotted reference without affecting any detection.

23. Kernel Gallery — the Same Chart, Nine Kernels

Each screenshot below was taken with a different Moving Average Type; everything else is the chart's own configuration. Compare three things: how far the MACD line sits from the turns in price, how much noise survives in the histogram, and how early or late the BUY/SELL markers land relative to the swing. The table columns are the kernel's definition, its warm-up and what it is good for — the same facts as §5.3, so you can match a picture to a formula.

24. Trading Divergences

24.1 Reading one signal in five seconds

1. Colour: which family? Green/red = regular (reversal). Teal/orange = hidden (continuation). Amber = quality below your weak threshold, whatever the family. If you only trade reversals, you have just filtered half the labels without reading anything.
2. Label: how good? The percentage is the 13-component score. Treat it as a ranking, not a probability: 90% means every component that could be evaluated came back strong.
3. Grade (if shown): what kind of disagreement? A = the oscillator turned back through the zero line, or the family-aware delta is ≥ 0.60 of the oscillator's scale — genuine exhaustion. B = a clear but shallower disagreement. C = a diagonal, the weakest form.
4. Position: where in the move? A regular bullish divergence wants price at a lower low after a downtrend, with the trend row reading DOWN or FLAT and momentum FADING. The same label printed halfway up a strong rally is a different (worse) trade.
5. Context: what do the higher timeframes say? The MTF row. A bullish divergence against MTF −62 BEAR is a counter-trend scalp; with +62 BULL it is a pullback entry.

24.2 The two playbooks

24.3 When to ignore a divergence that was drawn

• It is amber. Below your weak threshold. The engine registered it because it cleared the floor; the colour says it is the weakest thing on the chart.
• It is old. The panel's Age row is there for this. A divergence from 200 bars ago is history, not a signal — and the alert logic agrees: it only fires for signals whose confirmation bar is within two bars of the live edge.
• The ATR row says WARM. The volatility gate has not armed yet, so the adaptive floor is not protecting you from dead-market pivots. On a freshly loaded chart, wait.
• The family bucket's win rate is small-sample. Hid Bear 57% (7) is not information; the panel prints the sample size next to the percentage for exactly this reason.
• Price is inside a spike. The volatility-regime component already marked it down (0.25 for a spike), but a spike also makes the pivots themselves less trustworthy than their prominence suggests.

25. Trading Crossovers

25.1 What you are being told

A BUY/SELL marker is a transition of the crossover latch (§5.6). With Signal Sensitivity Lambda = 1.00 that is exactly the classic MACD/signal crossing. Below 1.00 it is an advanced trigger: the histogram has climbed out of a band measured in its own average size, on its way to a crossing that has not happened yet.

25.2 The two things that make the advance safe

• An early trigger may only fire in the direction the histogram is leaving, never the one it occupies — otherwise the latch would flip back on the next bar and print BUY/BUY pairs out of noise inside the band.
• An early trigger is consumed by the crossing it anticipated, so the crossing itself does not print a second one. Exactly one trigger per oscillator excursion.

25.3 Live vs closed-bar marking

Mark Only Closed Bars = false (default) evaluates the forming source bar: you see the cross the moment it happens, and the marker is withdrawn if the cross does not survive to the close. Set it to true for a strictly non-repainting marker trail — the price of that is one bar of latency, which on H4 is four hours. Only On Closed Source Bar does the same job for alerts, independently.

25.4 Combining the two signal types

26. Multi-Timeframe Workflow

1. Decide what the ladder is for. Directional context (keep SoftPenalty) or a hard veto (HardGate). Soft is the better default: a divergence against the higher timeframe is not worthless, it just has to earn a higher quality score.
2. Let it derive, or pin it. Empty Confluence Timeframes takes the next Confluence Levels timeframes above the source from the standard ladder. Pin it when you have a view — an H1 trader who cares about H4 and D1 and nothing else should say H4,D1.
3. Check the ladder actually resolved. The log prints MTF ladder: H4 D1, and the panel shows one row per available level. A missing row means that series has too few bars or no closed mapping yet — not an error, and it contributes nothing rather than dragging the score toward zero.
4. Read the score, not the state. BULL/BEAR/NEUTRAL are the ±33 bands; the number tells you how much of the ladder agrees. +100 means every level is above zero, above its signal and rising.
5. Keep strict mapping on unless you specifically want the intrabar view. With it on, the ladder only ever quotes higher-timeframe bars that have really closed, so a registered signal's confluence — and therefore its quality — cannot change after a history reload.
6. Remember each level is independently adaptive. An H4 level under an M5 primary is not your M5 settings read on H4 data; it is the profile the indicator would resolve if you opened H4 directly. That is why the ladder stays meaningful across a wide timeframe span.

27. The Gates in Practice — Loosening and Tightening

27.1 The order to investigate in

1. Turn on Debug Log. Every rejection prints its reason and, where it is numeric, the value against the threshold: [Skip BullHidden] ATR% 0.0091 below floor 0.0311 q=0.0 @ …, [Skip BearRegular] trend too strong to fade (-0.412) q=0.0 @ …, [Skip BullRegular] quality below threshold q=47.3 @ ….
2. Read the reasons for the last dozen rejections. If they are all the same gate, that gate is your binding constraint — change it and nothing else.
3. Change one thing, then re-read. Every gate interacts with the quality score, so changing three at once tells you nothing.
4. Check the panel's Avg qual row before and after: it is the fastest summary of how selective your configuration is.

27.2 Which knob for which symptom

28. Alerts Workflow

1. Choose the channels. Write Alerts To Log, Show Alert Popup, Play Sound — three independent switches under the Enable Alerts master. The popup is the one that reaches you when the terminal is in the background; it is colour-coded green for bullish and red for bearish.
2. Choose the triggers. Crossovers (on by default) are the frequent ones; divergences (on) are the selective ones; zero-line crosses and confluence flips are off by default and both are noisy on fast charts.
3. Set the alert bar. Alert Min Quality (default 70) is deliberately above Min Quality Score (default 50): you want to see more signals on the chart than you want to be interrupted about. The start-up log tells you which side of the quality gate your alert bar ended up on.
4. Decide about intrabar alerts. Only On Closed Source Bar (default on) makes alerts strictly non-repainting. Off is earlier and can be contradicted by the close.
5. Know the arming rule. Nothing alerts during the initial history pass — otherwise attaching the indicator would replay every historical crossover at you. Alerts arm when the first pass reaches the live edge, and the guard indices are seeded at that moment.

29. Reading Your Own Statistics

• Always read the sample size with the percentage. 58% (43) is a fact about your feed; 75% (4) is a coincidence. Min Learning Samples exists to keep the second kind out of the quality score, but the panel shows both so you can judge for yourself.
• Compare the four buckets, not the average. The families are different hypotheses. A configuration that is profitable on Reg Bull and hopeless on Hid Bear should be traded with Divergence Kinds narrowed — that is what the parameter is for.
• Expired is not a loss. A signal that reached neither target nor stop inside the horizon is unresolved. The win rate is wins / (wins + losses); expired signals are excluded, because folding them into either side would invent a result.
• The default target/stop is deliberately conservative. 1.0 ATR against 1.5 ATR means the measured win rate is the one a tight-target/loose-stop trader experiences. If you trade the other way round, set Outcome Target and Outcome Stop to your own ratio before you start judging the numbers.
• Statistics reset on a history reload. They are a statistic of the records in the current window; keeping them across a reload would count the same market events twice and inflate the sample counts.
• They persist per symbol and per timeframe. AMP|EURUSD|h1 and AMP|EURUSD|m15 are different stores, so switching timeframe does not mix two populations.
• Avg qual is your selectivity gauge. If it sits just above Min Quality Score, the floor is your binding constraint and most candidates are being rejected at the last step.

30. Timeframe & Instrument Guide

30.1 What the defaults resolve to per timeframe

The same table as §3.6, kept here because this is the one you will actually consult. Defaults: Balanced scaling, ε = 1.0 min, sensitivity 1.0, Canonical high-timeframe policy, spread 0.35, reference set 12/26/9.

30.2 Instrument notes

SECTION 3 — TIPS

The habits that make the tool pay for itself, and the mistakes that make people abandon it.

31. The Golden Rules

1. Read the start-up log before you read the chart. It prints the configuration actually in force — resolved periods, both kernels, gates, windows, warm-up estimate — and every contradiction it reconciled. Two minutes there saves a week of guessing why the indicator “ignores” a setting.
2. Treat “(reference)” parameters as anchors. In automatic mode you are moving a curve, not a number. If you want literal periods, switch Adaptation Mode to Manual — and then read the drift warnings, because manual mode promises to honour what you typed.
3. Never change two gates at once. The quality score is a weighted mean; changing two inputs tells you nothing about either. Change one, look at Avg qual and the per-family win rates, then decide.
4. Sample size beats percentage. Every win rate in this product is printed with its n. Respect it.
5. Family matters. A regular bullish and a hidden bearish divergence are different hypotheses with different track records. Narrow Divergence Kinds to what your bucket statistics actually support.
6. Quality is a ranking, not a probability. 92% means “every component that could be evaluated came back strong”. It does not mean 92 chances in 100.
7. Use the confirmation bar as your reference. That is where the indicator measures entry, ATR, target and stop. Comparing your fills to the pivot price compares yourself to a trade nobody can take.
8. Give the pane room. The panel fits itself to the pane height and hides low-priority cells first; a two-row-tall pane will show you almost nothing and that is by design, not a bug.
9. Keep strict MTF mapping on unless you have a specific reason. It is what makes a registered signal's quality stable across history reloads.
10. Let the warm-up finish before judging. The oscillator, the ATR gate and the learning buckets all need bars. A chart attached ten minutes ago on M1 knows nothing yet, and the panel says so (WARM, collecting).
11. Turn learning off when comparing configurations. Adaptive Learning = false makes the detector a pure function of parameters and data, which is the only fair way to A/B two setups.
12. If you see nothing, read the skip reasons. Debug Log = true prints one line per rejected candidate with the number that rejected it. “No signals” is always a specific gate, and it is named.

32. Recommended Presets by Trading Style

Only the parameters that differ from the shipped defaults are listed. Everything not mentioned stays as it is.

32.1 Scalping, M1–M5

Use the built-in style: Trading Style = Scalping. It writes 95 parameters (§4.2), including the Hull kernel on the volume-price-adjusted input, λ = 0.92, every precision gate armed, follow-through required, and quality floors of 60/72/76. Then consider:

32.2 Intraday, M15–H1 (the default configuration, lightly adjusted)

32.3 Swing, H4–D1

32.4 Position, W1–MN

32.5 Crypto, 24/7

32.6 Divergence-only, minimal chrome

32.7 The “plain MACD” baseline (for comparison)

32.8 The three style presets, value by value

§32.1–§32.7 above are manual configurations you assemble yourself. Since v2.0 three of them are also one click away: Trading Style = Scalping, Swing or Position. Every preset writes the same complete surface — 134 of the 162 parameters — so two charts on one style can never disagree about a gate, a window or a plan level because one of them kept a template value. What deliberately stays yours in every style: colours, icons, sounds, Z-index, line style, the session-marker times / window / colour, the source timeframe, and the two start-up diagnostics toggles.

The tables below list only the values that differ from the Universal defaults — everything else the preset writes is the default, pinned so a saved template cannot leak into it. 47 values differ for Scalping, 17 for Swing, 61 for Position.

Scalping — what it changes

Swing — what it changes

Position — what it changes

33. What NOT to Do

34. Performance & Chart Hygiene

• The core is incremental. A tick re-evaluates one source bar. Simple-MA windows are O(1) per tick (the cache remembers the value it holds at the forming index and swaps it, re-deriving exactly every 1024 steps); recursive kernels are O(1) by construction; composite kernels pay for their cascade.
• The heaviest choices, in order: Price Input = VolumePriceAdjusted (an O(window) scan per moving tick, memoised when nothing moved), long Slow Period with Triangular or Hull kernels, a large NormWindow (derived: 4×Slow), and three confluence levels (each a complete pipeline).
• Drawing budgets are the thing to tune on deep histories. Draw Only Most Recent N Bars and Max Drawn Divergences bound divergence objects; Max Cross Markers and the same recent-bars window bound markers. The marker back-fill after a reload is bounded by DrawRecentBars + 2×MaxCrossMarkers, and trimming runs during the walk, so a deep history cannot stall the first paint or hold thousands of objects alive.
• Object names are stable and namespaced per instance. No flicker, no leaks, and two copies of the indicator on one chart never fight over the same object.
• Records outlive their drawings on purpose. A divergence whose chart objects were evicted keeps its record, because outcomes, learning, the panel census and the alerts all key off source indices. The list is capped at max(MaxDivergences, LearningMemory) + 64, and a record with no objects holds plain values only.
• Panel refresh is throttled (Panel Refresh (ms), default 250) and skipped entirely when nothing is dirty. Raising it to 500–1000 on a busy multi-chart workspace is free readability.
• Turn off what you do not read. Every section switch (Show Configuration/Market/MTF/Signals/Statistics Section) removes work as well as ink; Show Divergences = false stops the pivot scan entirely.

35. Verifying the Non-Repaint Claim Yourself

The claim is specific, so the test can be too. Do this once on a symbol you care about; it takes five minutes.

1. Record the state. On a chart with a decent history, note the three most recent divergence labels (family, quality, position) and the last few BUY/SELL markers. A screenshot is enough.
2. Force a rebuild. Scroll the chart far to the left so the platform downloads deeper history, then back to the live edge. The indicator detects the change and rebuilds everything from scratch (§2.6).
3. Compare. The same divergences must be there, with the same quality numbers, the same colours and the same anchors. If a quality number changed, Strict Non-Repaint MTF Mapping is off and the confluence component was re-scored against a now-closed higher-timeframe bar — which is exactly what the switch is for.
4. Switch timeframe and back. Same result; the rebuild path is the same one.
5. Watch the forming bar. A crossover marker on the live bar may appear and then disappear if the cross does not survive to the close. That is the documented provisional behaviour, not repainting of a confirmed signal: closed bars are back-filled exactly once and never revisited.
6. Check a pivot's publication time. A pivot at bar c can only be acted upon from bar c + Pivot Lookback onwards, and the divergence dot sits at that confirmation bar — never at the pivot. Count the bars between the pivot and the dot: it is the lookback, every time.
7. Replay it if your build has bar-by-bar replay. Step through the last 200 bars. Signals must appear at the same bars with the same labels as the static chart shows — the pivot scanner advances confluence, volatility and outcomes to their per-confirmation state precisely so that a bulk pass and a live pass agree.

36. FAQ

37. Troubleshooting

38. Glossary

APPENDICES

Appendix A — Screenshot Index & Naming

Every illustration in this manual lives in the images/ folder next to this file, under the name the referencing section uses. Numbering is intentional and forms blocks: 01–09 the general views (terminal, panel, and the parameter dialog by group block), 10–18 the kernel gallery of §23 in the dialog order of Moving Average Type, 20–29 trade plans, 30–39 momentum trendlines, 40–49 core and markers, 50–59 session markers, 60–69 diagnostics and the journal, 70–79 trading styles.

Appendix B — Source-Code Audit (v2.0, 2026-10-23)

This revision of the manual was re-verified against the shipped source (AdvancedMacdPro.cs, header “ADVANCED MACD PRO 2.0”, 14 245 lines). The comparison was mechanical wherever possible: every [Parameter], [Output] and public enum declaration was extracted from the source and diffed against the tables of Sections 1–3.

Advanced MACD Pro — Official User Manual · v2.0 · © 2026. Trading involves substantial risk of loss. Past measurements on one feed do not guarantee future results; this manual describes tooling, not investment advice.