Auto ZigZag Fibonacci Golden Zone is an MT5 indicator that derives pullback entry zones from the latest confirmed swing using an internal ZigZag. It scans history to lock the most recent swing high/low, then draws Fibonacci retracements at 50%, 61.8%, and 78.6%.

The 61.8%–78.6% band is marked as the Golden Zone and extended a configurable number of bars. Optional chart labels can trigger when a candle closes inside the zone, with separate handling for uptrend and downtrend measurements.

A compact on-chart panel reports current price, swing points, exact fib levels, and usage steps. Key inputs include ZigZag Length (default 13), Golden Zone Length (15), swing price labels, and signal labels. Calculations are based on closed candles and remain fixed until a new swing is confirmed.

👉 Read | VPS | @mql5dev

Market Structure Shift (CHoCH) is used to confirm a directional reversal.

Bullish CHoCH occurs when price breaks above the last confirmed Lower High (LH), shifting bias from bearish to bullish. Bearish CHoCH occurs when price breaks below the last confirmed Higher Low (HL), shifting bias from bullish to bearish.

After confirmation, Areas of Interest (AOI) are mapped for execution. Common zones include Fibonacci Equilibrium anchored from the origin swing, and a post-CHoCH Fair Value Gap (FVG) formed within the first three candles of the breakout impulse, treated as an imbalance-based entry area.

Risk and targets are defined structurally. For buys, stop loss is placed below the origin swing low (0% Fibonacci anchor). For sells, stop loss is placed above the origin swing high (0% anchor). TP1 is set at the breakout structure level (100%). TP2 targets t...

👉 Read | NeuroBook | @mql5dev

Dynamic flip zones update automatically, converting broken Support into SBR and broken Resistance into RBS on confirmed closes. Optional behavior allows broken zones to be removed instead of flipped.

Zones are drawn as normalized rectangles rather than thin lines. Width is adjusted using average historical volatility, reducing oversized areas and expanding narrow ones to keep sizing consistent across market regimes.

Repaint risk is reduced via confirmation bars that validate pivot points before zones are plotted. Pivot detection is based on ZigZag swing highs and lows, with additional filters to keep levels relevant.

Alerts are gated by a minimum move-away requirement. Price must travel a defined distance from a new zone before a retest can trigger notifications, cutting noise in ranges.

Overlap handling includes Most Extreme and Newest modes to ma...

👉 Read | NeuroBook | @mql5dev

Many indicator specs claim “non-repainting” without a measurable definition. A testable invariant is stricter: once a bar is closed and processed, any drawn object on that bar must never change, move, recolor, change text, or vanish.

A script operationalizes this by recording every object after an initial pass, then appending more bars and forcing a full recalculation so the indicator rebuilds from a longer history. Objects on already-closed bars are matched and compared field by field (anchor times/prices, color, text). This targets failures caused by object names tied to bar index rather than bar time.

Changes and disappearances are tracked separately; only changes falsify the claim. Empty comparisons are reported as inconclusive. Results are written to CSV (one row per symbol/timeframe/step) including compared/changed/vanished counts and the first o...

👉 Read | VPS | @mql5dev

This article builds a practical case for a Cairo-style 2D renderer in pure MQL5 to draw modern UI elements (rounded rectangles, rings, gradients, translucent layers, arbitrary polygons) as a single OBJ_BITMAP_LABEL, avoiding the performance and feature limits of native chart objects and CCanvas.

The core model separates geometry (paths) from paint (sources), then converts paths into a per-pixel coverage mask for true anti-aliasing. Rendering becomes compositing: source through mask onto destination, so new paints (solid, gradients, images) automatically work with all shapes.

Part 1 focuses on the foundation: a strict ARGB uint color pipeline compatible with ResourceCreate, and a reusable pixel surface bound to one bitmap object—keeping later improvements isolated to masking and painting logic.

👉 Read | Calendar | @mql5dev

Manually scrolling to a specific candle in MT5 becomes impractical on lower timeframes and deep history. This History Navigator EA solves it with a small dialog where day/month/year/hour/minute are entered, then the chart jumps to the correct historical area and can return to the live market view in one click.

The design separates concerns: lifecycle code in HistoryNavigator.mq5, and UI + logic in a CNavigatorDialog class built on the Standard Library (CAppDialog, event map, controls). Inputs are validated in two stages: range checks plus real calendar validation with leap-year handling, then converted via MqlDateTime + StructToTime().

Bar location uses CopyTime() and a binary search over available history, selecting the latest bar open not exceeding the requested timestamp. Chart positioning disables auto-scroll/shift, centers the target using CHART_VISIBLE_...

👉 Read | Forum | @mql5dev

Momentum oscillators often become noisy when raw price differences are used directly. A Hull Moving Average layer can smooth the series while staying responsive to direction changes.

An MQL5 implementation combines classic momentum (Close[i] minus Close[i+Length]) with HMA built from multiple WMAs: a fast WMA on half period, a slow WMA on full period, then a final WMA on the rounded square-root period.

The design relies on separate buffers for raw momentum, intermediate HMA values, final output, color indices, a zero reference, and a fixed gray fill between the line and zero.

Interpretation is based on position versus the zero line; crossings indicate state change but degrade in flat regimes due to frequent flips and reduced signal quality.

👉 Read | Forum | @mql5dev

Clustered feature importance depends on a correlation matrix that is both denoised and detoned. Estimation noise inflates spurious correlations, and a dominant first eigenvector from shared regime exposure makes unrelated feature families look similar. Both effects break clustering and bias MDI/MDA via substitution.

Noise is bounded using a Marcenko–Pastur fit, with q = T/N and sigma^2 fit to the empirical eigenvalue density. Two common silent failures are inverting q and using an incompatible KDE bandwidth definition, both yielding plausible but incorrect ceilings.

For serially correlated bars, raw T overstates information. An AR(1)-style effective sample size can shift lambda_max enough to change factor retention near the margin. After denoising (constant residual eigenvalues) and detoning (remove top eigenvector), ONC/K-means clustering recovers...

👉 Read | AlgoBook | @mql5dev

MetaTrader 5 file access stays inside the terminal sandbox, so identical file APIs can yield different internal layouts while producing the same terminal output.

Directory traversal in MQL5 splits into two paths. Interactive selection uses FileSelectDialog with a fixed root, constrained filters, and flags; results come back via a dynamic string array and are typically passed into FileOpen.

Non-interactive enumeration relies on FileFindFirst/FileFindNext. An empty sandbox returns an invalid handle and is not a code error. Basic enumeration lists only the current directory; recursive traversal requires passing subdirectory paths. A trailing slash in returned names can be used to detect directories without FileIsExist, but output may lose full path context unless it is assembled explicitly.

👉 Read | AlgoBook | @mql5dev

A losing EURUSD short exposed a deeper issue: stacking RSI/MACD/Stochastic (and even a neural net trained on them) still reduces a nonlinear, chaotic market into mostly linear price transforms, so the model misses context.

LLMs can discuss that context, but they are nondeterministic: small prompt or temperature changes flip conclusions, which makes probability outputs unusable for trading decisions.

A more reliable path combined CatBoost with quantum feature extraction. Plain indicator features stalled near 59% accuracy; the breakthrough was quantum encoding with 8 qubits plus CZ entanglement, producing a non-uniform state histogram that captured higher-order relationships across indicators.

Four derived quantum statistics (entropy, dominant-state strength, state-count complexity, distribution variance) became numeric features. Adding them to 33 tec...

👉 Read | Calendar | @mql5dev

Backtests often look clean because the tester uses a “current” spread, while the terminal stores a spread value per M1 bar that rarely gets reviewed.

SpreadAudit reads M1 history for a configurable lookback and computes median, p90, and p99 spreads in price units and pips. The same percentiles are also normalized by ATR(H1) to make cost comparisons consistent across instruments. It also aggregates average and maximum spread by hour and can export the hourly table to CSV in MQL5\Files.

This matters when edge is small. A setup averaging 0.28 price units per trade fails if the median spread is 0.37, even if the backtest equity curve looks stable without realistic costs. Hourly stats typically show widening at rollover and near session opens, where a spread filter expressed as a fraction of ATR remains valid as volatility shifts.

Caveat: the recorded fi...

👉 Read | AppStore | @mql5dev

EA initialization requires attaching it to a standard symbol chart (for example, BTCUSD). A custom replay symbol is created automatically and should be opened as a separate chart.

Replay start is controlled from the panel in the top-left. Use SELECT to place a yellow vertical marker, move it to the target start date, then run LOAD DATA and press PLAY.

Playback speed can be adjusted from 0.5x to 20x. Controls include pause and step/skip actions (+10, -10, <<, >>) for fast repositioning.

Trade setups are created from the bottom-right drawing toolbar. Choose LONG or SHORT, click to set entry, move the cursor to define stop distance, then click again to lock; take profit is calculated automatically. Entry, stop, and take-profit lines remain draggable for later adjustments.

For reliable deep history loading, set Tools > Options > Charts and increase Max bars i...

👉 Read | Quotes | @mql5dev

MetaTrader 5 indicators cannot change orders or positions on the trading server. Only Expert Advisors can send trade requests, so UI actions in an indicator must be forwarded to an EA via custom chart events.

A position indicator was extended with a clickable OBJ_BITMAP_LABEL (16x16 resource). On click, it emits a dedicated close-position event with the position ticket in sparam. The EA intercepts it and closes only the targeted position, avoiding NETTING vs HEDGING edge cases tied to “close all” semantics.

The indicator also tracks chart scale changes to recompute bitmap coordinates using price-to-screen conversion, keeping the control aligned with price lines. Next steps generalize the same event pattern for Stop Loss and Take Profit line actions without code duplication.

👉 Read | Calendar | @mql5dev

Dandelion Optimizer (DO) is a 2022 metaheuristic that models how dandelion seeds explore wide areas, then converge where conditions look best. Each seed represents a candidate parameter set; “soil quality” is the objective value.

DO updates a population in three phases: Rising for broad exploration (mostly spiral/vortex moves, sometimes linear drift toward the search-space center), Decline for coordination via the population mean, and Landing for exploitation around the current elite using Lévy-flight steps to keep occasional long jumps.

The MT5-style implementation wraps this into a C_AO_DO class with Init/Moving/Revision, explicit boundary reflection, and adaptive control: step intensity decays over epochs while an attraction ratio increases, shifting from global search to fine-tuning. This fits strategy optimization where stable convergence and ...

👉 Read | VPS | @mql5dev

RL trading often fails for non-algorithmic reasons: weak input signal, leakage, unstable validation, or execution mismatch. A fast supervised baseline should be the first gate: if features cannot rank direction above chance, more complex RL adds cost without information.

A LightGBM check with triple-barrier labels and purged walk-forward CV is a practical filter. On ~144k M15 XAUUSD bars, AUC stayed near 0.50 across horizons and added cross-asset features, aligning with low explained variance and weak live stats.

Engineering controls that reduce train-to-live failure include embargoed walk-forward splits, multi-seed promotion gates, evaluating realized equity/trades (not shaped rewards), and deployment contracts: saved normalization, pinned versions, dataset snapshots, warm-up rules, and broker reconciliation.

👉 Read | Calendar | @mql5dev

Exit logic can be treated as a survival problem, not a fixed rule. Once a position is open, state changes each bar, so the probability of TP before SL should be conditional on survival and updated with time-varying covariates.

A discrete-time competing-risks model fits this: TP and SL are mutually exclusive events, while time-stop creates censoring. A person-period table (one row per trade per bar) supports proper estimation; open-at-time-stop rows are not “losses”.

A native MQL5 implementation is described: no DLLs, no ONNX, no ALGLIB. Two cause-specific hazards are fit via penalised maximum likelihood, converted to cumulative incidence, then to a bar-by-bar hold/close decision while keeping hard TP/SL orders.

Results: out-of-sample log-loss improves 20–22% vs time-only baseline across EURUSD, GBPUSD, USDJPY, XAUUSD, but fixed decision thresholds...

👉 Read | VPS | @mql5dev

MetaTrader 5 sandbox navigation can be implemented via OS dialogs for convenience or via code to enumerate files and directories. The code-driven path becomes necessary when output must be filtered and ordered by attributes such as creation or modification time.

MQL5 provides ArraySort for numeric arrays, but it does not handle strings. A custom bubble-sort style routine can generalize sorting, yet string ordering requires explicit comparison rules.

Using StringCompare alone produces lexicographic ordering, which misplaces numeric strings. Adding length checks fixes numeric strings but can break dictionary word ordering. A revised comparator that switches strategy restores expected results for both cases.

Bubble sort remains O(n²) and is suitable mainly for small datasets or demonstration, not large directory listings.

👉 Read | Freelance | @mql5dev

SCNN reframes multivariate forecasting as an explicit decomposition problem, addressing non-stationarity and regime shifts that make generic neural models brittle and hard to interpret.

The architecture splits each series into long-term, seasonal, short-term, and cross-series co-evolving components, plus a residual. Each component is normalized with additive and multiplicative terms, then handled by a specialized subnetwork matched to its dynamics.

Decomposition and reconstruction are embedded inside the network, enabling information flow between components. A two-branch module adapts parameters online to changing autocorrelation, while structural regularization discourages noise-driven features.

For trading systems, this yields more stable forecasts during distribution shifts and anomalies, with practical computational cost and clearer diagnos...

👉 Read | CodeBase | @mql5dev

This update refactors the MT5 position indicator to support removing stop-loss and take-profit directly from the chart, alongside the existing “close position” interaction.

Core changes focus on decoupling UI elements into a reusable C_ElementsTrade class. C_IndicatorPosition now holds three pointers (open, SL, TP), initializes them to NULL, allocates with new only when needed, and reliably frees memory with delete while resetting pointers to avoid invalid dereferences.

Event handling is hardened: DispatchMessage first validates pointers, then verifies ticket ownership and server-side position existence. Protective levels are created, updated, or destroyed based on whether the server reports a nonzero SL/TP, keeping the chart synchronized with real trade state.

C_ElementsTrade centralizes creation of the price line and an action button, relying on M...

👉 Read | NeuroBook | @mql5dev

Random-walk MCMC methods degrade in high-dimensional models due to small effective step sizes, high autocorrelation, and low throughput. Hamiltonian Monte Carlo addresses this by using gradients of the log target density to generate long, directed trajectories.

HMC augments parameters z with momentum r and defines a Hamiltonian H(z,r)=E(z)+K(r), with E(z)=-log p(z). Leapfrog integration provides reversible, volume-preserving proposals, with Metropolis correction based on energy error.

A 100D correlated normal test case is used, including adaptive tuning of step size epsilon and a diagonal mass matrix. Warm-up is accelerated via MAP initialization using L-BFGS.

Sampling quality is tracked using Rhat, ESS, MCSE, summary stats, and runtime monitoring of log density stability, acceptance ratio, step size, and divergences.

👉 Read | Quotes | @mql5dev