Extrapolator is an MQL5 indicator that fits a truncated Fourier model to recent price history, then projects the fitted curve into future bars. The forecast is recalculated on each new model bar and the segment at the right edge is redrawn while the current bar is forming, so the line is inherently non-stationary.

Key controls include Npast/Nfut/Nharm to balance stability versus overfit, optional log-price fitting for proportional symmetry, and TrendDamp to reduce the carried-forward slope. Harmonic selection filters out periods below MinPeriod and above MaxPeriodFrac of the window, then stops when incremental variance reduction drops under MinHarmGain.

Visualization adds sigma-based channel bands (+/-1, +/-2) with optional filled zones. Internally, Quinn-Fernandes is used for frequency estimation, results are cached per model timeframe, and then mapped onto char...

👉 Read | Forum | @mql5dev

This article replaces “trust the ONNX class” with a robustness check: how much feature-space noise the input can tolerate before the prediction flips. Using randomized smoothing, the EA runs N Gaussian-perturbed inferences per bar, votes the winning class, and computes a certified radius via a Clopper–Pearson lower bound and the inverse normal CDF using only MQL5’s standard math functions.

The architecture is cleanly split: a feature engine builds a fixed 10-value normalized vector, a certifier owns the ONNX session and certification math, and the EA gates trades on the last closed bar. Shape mismatches are caught early, float32 tensors are handled correctly, and abstains occur when statistical support is weak.

Practically, the radius becomes a trading input: skip or downsize fragile signals, allocate size across symbols by robustness, and log radius history fo...

👉 Read | AppStore | @mql5dev

Tick volume on retail FX charts is mostly a proxy for time of day, and it is only complete at bar close. It also loses resolution when bursts and evenly spaced updates produce the same count. A quote-interval activity gauge addresses this by working on durations between quote changes and normalising out the intraday rhythm.

Quotes are converted into events by folding same-millisecond updates and resetting after long gaps (default 5 minutes) to avoid training on market closures. A 30-minute-bin diurnal profile is fitted and interpolated, then scaled to unit mean so 1.0 represents normal activity for that hour.

Residual clustering is modelled with an Engle-Russell ACD recursion, producing a per-quote activity ratio (mean/psi). Values near 2.0 indicate quotes arriving roughly twice as fast as usual; 0.5 indicates half speed. Additional outputs includ...

👉 Read | VPS | @mql5dev

ATS_RiskLib.mqh is a single include file providing five risk-control classes for MQL5 EAs. The library does not open, modify, or close trades; it only calculates and validates conditions before an order is sent.

CAtsSizer computes volume from stop distance with modes for fixed lot, % balance, % equity, or fixed money risk. Volume is floored to the broker step and never rounded up. If the computed risk is below the broker minimum, Lots() returns 0.0 with a reason via LastReason(). Optional volume cap and free-margin validation use OrderCalcMargin.

CAtsDayGuard enforces daily loss (percent and/or money, using the smaller), optional profit target, and max trades per day. Day start is configurable by server hour and can include open P/L. State is persisted via GlobalVariables, so terminal restarts do not reset limits.

CAtsDrawdownGuard supports static max loss...

👉 Read | AlgoBook | @mql5dev

ATS Prop Rules Monitor aggregates five common prop challenge constraints in a single panel and shows remaining room for each rule, rather than only daily loss. Supported checks: daily loss (percent of day-start balance, optional floating P/L, configurable day-start server hour), max loss (static from initial balance or trailing on equity/balance), profit target, minimum trading days (days with at least one closed trade), and best-day consistency (disabled when not applicable). Status uses green/amber/red plus a summary badge: ALL CLEAR, WARNING, LIMIT BREACHED, TARGET MET.

A “One More Trade” line calculates the largest lot size for the chart symbol that keeps a stop-out within a configurable share of the smaller remaining room. Lot sizing follows ATS Risk Library rules and never rounds up.

Implementation details: panel rendered via CCanvas with DPI scaling...

👉 Read | Docs | @mql5dev

Naive label sampling in ML trading pipelines creates predictable failure modes: class imbalance during trend-heavy periods, feature/label autocorrelation that inflates fit quality, and heavy class overlap that forces unstable decision boundaries.

Dimensionality reduction (PCA) makes these issues visible: loop-shaped structures from autocorrelation and strong overlap between buy/sell regions, consistent with GIGO effects and poor out-of-sample behavior.

A practical mitigation is probabilistic resampling with Gaussian Mixture Models. GMM models density with full covariance, generates synthetic samples with fewer outliers, and reduces serial dependence. Labels are clustered jointly, then re-binarized.

An iterative loop generates GMM samples, trains CatBoost, and ranks runs by out-of-sample R² across a longer test window than the training period. Supporting ch...

👉 Read | Docs | @mql5dev

A chart indicator aimed at studying the ICT/Smart Money sequence on any symbol and timeframe. It is analysis-only, does not place trades, and does not claim performance.

It draws liquidity pools (previous day high/low, Asia session high/low), marks sweeps when price takes a pool and closes back inside, then detects MSS via an ATR-sized displacement candle and highlights the resulting FVG. Signals use WATCHING/CONFIRMED states.

Non-repainting is enforced by creating objects from closed bars only; recalculation rebuilds identical objects. Optional CSV logging writes each signal at creation time.

For confirmed setups it plots entry (FVG midpoint), SL (beyond sweep), TP (risk multiple), estimates lot size for a risk %, and tracks outcomes bar-by-bar (TP/SL/missed; TP+SL same bar counts as SL). A panel aggregates setups, fills, misses, win rate, and n...

👉 Read | NeuroBook | @mql5dev

NewtonianMechanics.mq5 provides bid/ask tick kinematics with explicit uncertainty. Bid and ask are handled as separate streams, each fitted with a local quadratic and reported as position, velocity, acceleration, standard error, and z-score. Forecasts are then checked out of sample with a scoreboard (hit rate and R-squared), including a split by |z| >= 2.

Windows are counted in events rather than seconds, using exponential weighting so evidence stays comparable across quiet and active periods. Time is taken from MqlTick.time_msc, and each side advances only on its own updates; repeated quotes are not treated as new observations.

Error bars apply a random-walk floor to avoid overly optimistic OLS assumptions. A phase label is derived from the slowest window (256 updates): EXPANSION, RETRACEMENT, STRESSED, CONSOLIDATION, UNDETERMINED, WARMING. A spr...

👉 Read | CodeBase | @mql5dev

Chart patterns are visually clear but hard to define in code due to ambiguous start bars, low tolerances, neckline placement, and validation timing. A practical gap is the inability to preserve the exact candle sequence behind a marked formation for later reload, measurement, and comparison.

A six-module MQL5 recorder addresses collection, not detection. The trader manually labels a double-bottom candidate via a chart rectangle. The system extracts candles by time bounds only, derives candle features, normalizes OHLC as percentage displacement from the first close, then persists each sample as CSV and reloads the library after restart.

Architecture splits responsibilities across PatternData, Extractor, Storage, Library, Panel, and a thin EA integration layer. Storage uses strict schema/version headers, UTF-8 quoted fields, non-overwriting writes, and validation...

👉 Read | Freelance | @mql5dev

This article turns two built-in MT5 indicators into a single, rule-driven decision aid by giving Envelopes the missing market-state context from Alligator. It highlights a key pitfall: Envelopes are fixed-percentage bands around an average, not a volatility measure, so boundary touches can mean range noise or trend continuation.

The strategy classifies Alligator line order (Lips/Teeth/Jaws) into neutral, unconfirmed, and confirmed states using persistence across completed bars (InpTrendMinBars). Only confirmed direction or neutral can authorize an Envelope close-cross event; counter-trend crossings are filtered, repeated arrows are throttled via cooldown, and optional alerts trigger on closed bars.

Implementation-wise, Envelope_Alligator_Strategy.mq5 composes four built-in handles (iEnvelopes + three unshifted iMA) to avoid re-implementing formulas, ...

👉 Read | Freelance | @mql5dev

PropFirmAwareSizer adds two modifiers on top of budget-based sizing: derisking_factor and news_window_factor.

derisking_factor reduces exposure as phase_progress approaches the profit target. It stays at 1.0 until 80% progress, then tapers to a floor at 100%. The floor is 0.30 before min trading days are met, and 0.10 after, so day-count compliance remains possible while still cutting risk near a pass.

Both modifiers multiply, not min(). Moderate constraints compound, and identical final sizes can come from different states, so diagnostics like cal_bet_size and derisking are required.

news_window_factor is funded-stage only. It uses a UTC-normalized economic calendar and applies a 5-minute half-window around high-impact events. It is a sizing heuristic, not profit accounting, and should be symbol-aware to avoid false haircuts.

👉 Read | Freelance | @mql5dev

A native MQL5 port of Kronos can match PyTorch outputs and still fail as a forecaster. The article separates verification (tight numeric equivalence) from market evaluation, where misleading “good-looking” results are easy to generate.

KronosForecast EA runs inference once per completed bar (never using the still-forming bar), then renders predicted candles via chart objects. Missing exchange volume/amount is approximated from tick volume and typical-price turnover, introducing a distribution shift. Future timestamps are projected on a fixed grid; weekend crossings are measured and shown not to explain poor accuracy.

Evaluation is built to prevent look-ahead by design: dump deterministic greedy forecasts to file, score them offline, then backtest using only forecast columns. On EURUSD H1 (256-bar context, 16-bar horizon), Kronos underperforms a random-walk ...

👉 Read | NeuroBook | @mql5dev

Deep learning in trading hit a known limit: adding layers can stall or reduce accuracy due to optimization issues. Residual connections (ResNet) addressed this by preserving an identity path and stabilizing gradient flow.

Spiking neural networks (SNNs) better match event-driven market behavior by processing discrete spikes, with potential gains in real-time processing and efficiency. In practice, deep SNNs have been hard to train: identity mapping is non-trivial with spiking neurons, and surrogate gradients amplify vanishing or exploding gradients as depth grows.

SEW-ResNet (Spike-Element-Wise residuals) reframes residual learning for SNNs with element-wise spike-based gating that supports natural identity mapping and stable gradients. Reported results include stable training beyond 100 layers and improved accuracy versus prior deep spiking baselines, enab...

👉 Read | Freelance | @mql5dev

Part 84 adds a structural validator to a MetaTrader 5 double-bottom recorder, turning a visually labeled sample into a testable hypothesis with deterministic rules and explainable measurements.

The validator finds swing lows via pivot-strength windows, consolidates nearby lows, then evaluates every low pair: separation, neckline (highest intervening high), depth, low similarity (depth-relative), and symmetry around the neckline. Accepted pairs are classified by the first decisive close: CANDIDATE, CONFIRMED (breakout), or FAILED (breakdown); otherwise INVALID.

Implementation isolates logic in DoubleBottomValidator.mqh, keeps CSV data unchanged, exposes thresholds as EA inputs, and reports a weighted conformity score plus key indices/prices. Traders get consistent labeling; developers get an inspectable rule engine for testing and extension.

👉 Read | NeuroBook | @mql5dev

Correlation dashboards report a level, but the level alone cannot distinguish sector structure from a single market-wide driver. Ollivier-Ricci curvature adds a geometric signal by comparing neighborhood overlap on a correlation-derived metric graph.

The implementation builds a complete graph using Mantegna distance d=sqrt(2(1-rho)), then assigns non-uniform neighborhood measures via Gaussian decay. The curvature scale requires a structureless reference (“pole”) that depends only on node count, so output is interpreted as a gap from that pole, not a standalone value.

Core dependency is an exact Wasserstein-1 transport solver (min-cost flow with potentials) with certificate checks. Reducing transport to the positive parts of measure differences is exact only under a true metric, making triangle-inequality validation operationally important.

Multi-...

👉 Read | AlgoBook | @mql5dev

Advanced trade-management panel for MetaTrader 5 focused on prop-firm style risk controls. It does not generate signals and does not open trades on live or demo accounts without user action. The panel calculates hidden support/resistance levels from pivots to suggest SL/TP and grid levels, while keeping charts clean.

Risk handling includes daily and maximum drawdown limits with a configurable buffer, optional profit target and max daily profit, and automatic position/pending cleanup plus trading lock when limits are hit. Lot sizing is derived from a risk percentage and validated against leverage and margin rules per asset class, using the stricter of broker or configured limits.

Workflow is split into MAIN (protection/settings), LIVE (market orders with draggable SL/TP), PEND (S/R-based pending orders), and GRID (2–8 level grids, optional breakeven, broker-...

👉 Read | Signals | @mql5dev

Most EAs rely on the broker’s session table via SymbolInfoSessionQuote and SymbolInfoSessionTrade. Time filters, rollover rules, and session boxes assume the table matches observed pricing. This auditor compares the declared sessions with M1 history and reports where they diverge.

Over N whole server days it counts, for each minute-of-week, how often an M1 bar exists on that weekday. It then flags declared minutes that are usually empty, bars occurring outside declared sessions, mismatches between quote and trade tables, and one-off outages (gaps in normally live declared minutes).

Example run: 5 Market Watch symbols over 35 days (2026.09.01–2026.10.06 UTC), single pass in 62 ms. On an Exness demo feed, EURUSD/GBPUSD/USDJPY are declared 00:00–24:00 Mon–Thu, but history frequently has no bars after 21:00, so tester events scheduled there often never trigger....

👉 Read | Calendar | @mql5dev

CajasTF is a lightweight MT5 indicator that adds a real-time mini-chart panel on the right side of the main chart. It supports any timeframe from M1 to MN1 and updates the active candle tick-by-tick, reducing the need to switch charts or open additional windows.

The panel inherits the current chart’s visual settings, including background, grid, labels, candle bodies, wicks, dojis, and the Bid line, and adapts automatically when the template changes. An independent vertical price scale with a 5-level grid keeps proportions stable regardless of the main chart zoom.

Operational features include automatic right-margin shift to prevent overlap with current price action and restoration of original shift settings on removal. Live telemetry includes a candle countdown timer, a current Bid marker, and an OHLC footer with percent change for the active bar.

👉 Read | Freelance | @mql5dev

Session Levels is a chart-only indicator that plots intraday reference levels commonly used for discretionary analysis and execution planning. It does not place orders. Typical output includes Asian range boxes, previous day high/low, daily VWAP, swing highs/lows, and an on-chart info panel.

Usage is straightforward: attach to any symbol and timeframe, with M5–H1 typically preferred. Hover tooltips describe each line or box, while the top-right panel shows current values. Colors switch automatically for dark or light chart themes.

Key inputs cover session hours (broker server time), number of historical days to render, optional London and New York boxes, and extension rules for Asian highs/lows. Level options include PDH/PDL, PWH/PWL, and configurable round-number steps and density around price.

VWAP resets daily using tick volume. Swings confirm only a...

👉 Read | Quotes | @mql5dev

A divergence signal gains relevance when multiple oscillators disagree with the same price swing. This indicator checks ten classic oscillators against recent price pivots on every bar, draws divergences directly on price, and adds a box listing which oscillators triggered plus the total count.

Built-in calculations cover MACD and histogram (12/26/9), RSI 14, Stochastic 14/3, CCI 10, Momentum 10, OBV, VWMACD (12/26), CMF 21, and MFI 14. Each can be disabled independently; volume-based signals use tick volume.

Pivots use 5 bars on each side (close or high/low). Each oscillator compares the signal bar to up to 10 pivots within 100 bars, requiring a clean, unbroken line between points on both oscillator and price. Regular and hidden divergences are supported, with optional bar-close confirmation, confluence filtering by minimum count, and per-side alert...

👉 Read | Forum | @mql5dev

MetaTrader derives TimeTradeServer() and TimeLocal() from the local Windows clock. When that clock drifts, any time-based logic in an EA drifts with it. Order expirations can trigger early or late, session/news filters shift, and end-of-week shutdown logic can miss the intended second. The terminal provides no warning.

A recent VPS case showed the Windows time service stopped and the clock ran 11 seconds slow. Limit orders configured to remain valid for 15 seconds were expiring after 4 seconds.

A script can measure this directly using tick timestamps. Each tick includes the server-stamped time with millisecond precision. The script compares that stamp to the local clock at tick arrival across 40 fresh ticks and uses the smallest non-negative gap, adjusted by half the current ping.

Run on an active symbol during open market hours. Output in Experts log and on...

👉 Read | AlgoBook | @mql5dev