16 November 2026 · New coverage since the 9 November edition
A trading robot can pass a local check and still lose control of a position when the server responds late. This week's MetaTrader reading examines that gap from several directions: request validation, input timing, trading friction, forecast comparisons and the evidence left by connected services. The useful question is what each test actually proves before its result becomes permission to trade.
TL;DR
No new MT4 or MT5 desktop release is listed. A new preflight study reduced rejected requests but exposed the cost of an extra safety buffer. A forecast-comparison study found every tested model less accurate than its random-walk benchmark over the full sample. New timing, drift and tracing tools make failures easier to inspect, with explicit gaps between local checks and live evidence.
This report was prepared with AI from the linked public sources. Verify the original material and test implementations independently. It is not a recommendation to buy, sell or trade any investment.
Get MetaTrader updates in your inbox
MetaTrader and MQL research, testing notes and platform changes, alongside TradingView in the shared Trading Development newsletter.
Select Trading Development in the free signup flow. Already a member? Sign in and update your email preferences. Unsubscribe anytime.
At a glance
| Subject | Why it matters |
|---|---|
| MT4 and MT5 release histories | No newly listed desktop release; September and October notes are earlier context. |
| Execution delay · 10 November | A passed validation still left position-management failures. |
| Trade preflight · 12 November | Measure rejected requests and acceptable requests blocked by the buffer. |
| Temporal audit · 12 November | A PASS applies only to supplied timing metadata. |
| Friction Guard · 13 November | Quote-side filters leave latency, slippage and cost assumptions to validate. |
| Forecast comparison · 13 November | Separate prediction accuracy from exposure and net trading results. |
| Drift watchdog · 10 November | Investigate distribution changes before deciding whether to retrain. |
| Distributed tracing · 13 November | Distinguish local serialization from actual network delivery. |
| Native S3 client · 12 November | Tiny-file round trips do not prove full-scale model synchronization. |
| MQL4, MQL5 and Strategy Tester | Use distinct platform result contracts and separate network proof. |
In this edition: Platforms · Execution · Timing · Friction · Forecasts · Drift · Tracing · Artifacts · Implementation and testing
MT4, MT5 and language releases: no new announcement in this window
The official English release histories checked on 16 November still lead with MT5 build 6230, dated 24 September 2026, and MT4 build 1490, dated 2 October. Neither history lists a new desktop release since our 9 November research cutoff. This is a statement about published release notes, not the newest binary distributed through every broker or update channel. MT5 release history · MT4 release history
Earlier MT5/MQL5 context: build 6230 expanded the AI Assistant's ability to prepare Expert Advisors for tests, work with programs on charts and retrieve Economic Calendar data. Its numerical additions included vector/matrix sorting, matrix printing, more complex-number operations and CoreCLR 8+ support. Build 6180, from 3 September, already included tester-report, optimization, configuration and log tools, plus a visual-testing keyboard-event fix. These remain earlier changes, not this week's news.
Earlier MT4/MQL4 context: build 1490 describes security and stability improvements and changes to shared stop-loss/take-profit handling in FIFO mode. The note does not announce a new MQL4 language feature or identify particular vulnerabilities. Existing positions and new orders can share protection levels under that mode; account-specific assumptions need checking.
Our development priority: preserve the exact build, broker, account mode and symbol specification with a reproducible example. When two terminals behave differently, compare those facts before attributing the difference to a release. For this week, the new engineering material below is more actionable than presenting September or October functionality as newly shipped.
Execution: fewer rejections can still leave unresolved exposure
10 November: validation passed, delayed execution still failed
Ali Akbar's execution-delay investigation revisits a scalper that passed CodeBase validation but still suffered rejected modifications. A custom-symbol weekend case also directed modifications at the wrong position. In a fixed five-second-delay ablation, adding a burst-based distance reduced EURUSD rejected modifications from 37 to six; all five changes reduced them to four. The weekend case did not recur in the controlled runs, so the session buffer and deal-based lookup were not independently validated by those results.
The article's remaining defect is consequential: after failed protection and closure attempts, the EA can still lose control of an open position. Its improvements are debugging evidence, not a production-ready trading system.
Our proposed test: start with a deliberately failed protection change. Require a durable unresolved-position record, a visible reason and a subsequent reconciliation against the actual account. Repeat after a restart and a delayed response. A lower error count is insufficient if the remaining error leaves exposure unmanaged. Record position identity separately from the latest signal.
12 November: preflight checks and the price of a buffer
Akbar's standalone trade-request preflight class checks volume, protective levels, margin, sessions and request results. Its 52-check script verifies local rules. In three EURUSD random-delay stress runs, the unbuffered gate still allowed 271–277 server rejections. The first run's ten-point shadow buffer would have blocked all 275 rejections, but also 1,056 accepted requests. A stronger filter therefore trades rejection avoidance against lost opportunities.
The author did not log the server-side validation price; price movement is a supported hypothesis, not a measured cause. Nonzero stop/freeze limits were tested locally rather than confirmed on the demonstrated server. Repairs can change the requested trade, and successful preflight never guarantees a fill.
Our proposed test: run a candidate buffer in observation-only mode on the same requests as the baseline. Record blocked-and-rejected, blocked-and-accepted, passed-and-rejected and passed-and-accepted separately. Define acceptable changes to size and protection before enabling repairs. Evaluate request identity and outcome alongside counts, retaining partial or unresolved results instead of folding them into a success total.
12 November: timing metadata is evidence only when inputs are complete
Stanley Kimathi Kibaara's temporal-causality audit applies eight ordered checks to recorded dependencies. It distinguishes closed-bar and intrabar policies and returns named violations. The prototype trusts caller-supplied bar boundaries and availability timestamps: it does not establish that a source did not repaint, validate unrecorded inputs or trace dependency ancestry. An empty audit can report PASS. Twelve tests cover five of seven violation types; invalid-shift and timestamp-mismatch paths lack isolated tests.
A forming bar is not automatically wrong. Its acceptability depends on the declared decision policy and the value available at that moment. A warning remains unresolved timing uncertainty.
Our proposed test: inventory every input used by one decision, then remove a dependency intentionally and require the surrounding system to reject the incomplete record. Compare a captured decision with a later replay using the exact saved values. Test boundary timestamps, stale cross-symbol observations and a contradictory closed flag. Preserve both the value and its provenance so a plausible timestamp cannot substitute for evidence of what was actually read.
13 November: a quote-side friction filter needs a cost model
Kelvin Muturi Muigua's Friction Guard tutorial combines recent tick velocity, tick arrival frequency, adaptive spread statistics and an estimated friction ratio. It produces four market-condition classifications and reasons for permitting or blocking an entry. The supplied diagnostic observes a GBPUSD chart, but its exact observation date, server timestamp and commission schedule are undisclosed. It is an operational demonstration, not a reproducible performance benchmark.
The implementation does not explicitly model network transit latency or full order-book depth; slippage is estimated. Its hypothetical cost illustration should not be read as measured trading results. Nor does passing the filter establish positive expected returns.
Our proposed test: capture a representative quote stream and replay fixed entry candidates against several predeclared thresholds. Preserve the candidates rejected by every variant. Compare estimated costs with independently observed costs in a controlled execution study before assigning the filter financial value. Include quiet periods, sudden spread changes and insufficient warm-up data; publish the rate of unavailable estimates rather than treating absence as cheap execution.
13 November: forecast skill, predictive content and trading profit differ
Adewumi Babatunde Gbadebo's Diebold–Mariano implementation compares three-bar forecasts with a zero-return random walk, accounting for overlapping errors and small samples. A Clark–West adjustment addresses the nested autoregressive model. These tests ask different questions: predictive content is not identical to lower realized forecast error.
Across six market/timeframe cases, every candidate had higher full-sample RMSE than the benchmark. Gated runs traded far less often; their balance advantage cannot simply be credited to forecast skill. The reported open-versus-closed-bar trade comparison had p = 0.07. No commission was charged in those runs, and the author confirmed that squared-error loss measures accuracy rather than profit after costs.
Our proposed test: freeze the forecast, horizon, loss function and benchmark before evaluation. Save each prediction before its outcome arrives. Compare a statistical gate with a control that takes a similar number of trades, while keeping execution and costs identical. Report forecast accuracy, selection frequency and net results separately. If the conclusion changes when the loss function changes, explain which decision each test was designed to support.
Build on this week’s MetaTrader research
Follow the next practical development report, with source links, implementation limits and ideas to test. MetaTrader and TradingView share the Trading Development newsletter.
Select Trading Development in the free signup flow. Already a member? Sign in and update your email preferences. Unsubscribe anytime.
10 November: a drift alert is a prompt to investigate
Olamide Daniel Adebayo's PSI watchdog compares rolling observations with a captured reference, logs every reading and sends webhooks on state changes. Its demonstration monitors normalized RSI and transformed EMA spread. PSI detects distribution change; it does not measure model accuracy, identify the cause of change or establish that retraining will help.
The familiar 0.10/0.25 thresholds are heuristics, not validated FX boundaries. A fifty-observation window over ten buckets can be noisy. The article separately checks the mathematics, historical monitor behavior and delivery of a sample webhook. It does not demonstrate an actual running-EA transition followed through to receipt, and the false-alarm rate remains unvalidated.
Our proposed test: establish an alert budget on untouched reference periods and test several known disturbances without selecting thresholds retrospectively. Treat alerts as investigation records with explicit acknowledged, unresolved and closed states. Exercise the full trigger-to-receiver path, including duplicate messages, receiver downtime and recovery. Keep a model unchanged until a separate validation process supports replacement; a distribution alarm should not silently become a deployment instruction.
13 November: distributed traces can explain a missing trade
Adesewa Tolulope Gbadebo's native OTLP exporter links MQL5 spans and a Python service using HTTP JSON and a traceparent header. Finished spans enter a bounded queue with batched delivery and backoff. The demonstrated tester runs produced identical trades with tracing on and off, but the tester sink counted serialized spans without sending them. Its microsecond flush figures are not network-latency measurements.
A separate demo-chart example showed a cross-process trace, and a collector interruption exercised recovery. These observations do not prove zero overhead for every deployment. WebRequest remains synchronous even when invoked from a timer; the article's rough CPU timing is expressly not a controlled benchmark.
Our proposed test: define a complete trace for both a trade and a deliberate non-trade. Add bounded tests for slow collectors, a full queue and shutdown with unsent spans. Compare trading decisions and event-processing delays under the same inputs with instrumentation enabled and disabled. Document sampling and dropped-span counts so a missing trace cannot be mistaken for proof that an event never happened.
12 November: artifact storage needs verification beyond HTTP success
Gbadebo's native S3 client implements Signature V4 and wrappers for upload, download and listing, without DLL imports. The demonstrated round trips used files of 42 and 58 bytes and compared SHA-256 hashes. Listing was implemented but not exercised in that run. The evidence therefore supports tiny-file round trips, not large-model throughput or complete S3 compatibility.
The client uses fixed five-second request timeouts and specialized canonicalization. Empty-upload handling needs testing, listing returns raw XML, and credentials appear in script inputs and memory. Removing the DLL dependency does not remove terminal networking or permission constraints.
Our proposed test: define a versioned artifact record containing size, content hash, feature schema and model identity. Download into a staging location, verify it and only then make it eligible for use. Exercise truncation, a valid but wrong model, missing metadata and interrupted transfer. Keep activation separate from transport success and retain the last verified artifact. Any production integration also needs scoped access and a deliberate secrets-handling design.
MQL4, MQL5 and Strategy Tester: turn the reading into bounded experiments
The references here are existing platform documentation checked for this edition, not fresh language announcements. They help define what the new examples can demonstrate.
Read the result contract before designing the interface
OrderCheck populates a result structure; its Boolean return alone is insufficient to establish that an operation will succeed. A valid preliminary check does not guarantee subsequent execution. Keep the check result and the later submission result as distinct records. MQL5 OrderCheck documentation
For an MT4 port, OrderSend returns an order ticket or -1, with errors available through GetLastError. A zero reported stop level may coexist with dynamic server restrictions and an invalid-stops error. The new MQL5 helpers cannot be assumed to implement this different contract unchanged. MQL4 OrderSend documentation
Our proposed portability check: write equivalent observable outcomes for both platforms before sharing higher-level logic. Include invalid volume, a rejected protective level, insufficient funds, a stale proposal and a successful order. Each platform adapter should preserve its native identifiers and errors. Test the signal calculation separately from execution so a mismatched indicator value does not become an apparent broker failure.
Separate deterministic trading tests from network checks
MQL5 WebRequest is synchronous, requires allowed server URLs and is unavailable in the Strategy Tester. Moving a call into OnTimer does not make the HTTP operation asynchronous. The official testing documentation also describes limitations on external-interaction functions and differences between tick-generation modes. A simulated sink or printed delivery message cannot establish that a remote recipient received data. WebRequest reference · Strategy Tester documentation
Our proposed proof split: first run trading and state-transition logic against fixed inputs without external dependencies. Then exercise network serialization, timeouts and receiver acknowledgements from a controlled terminal environment. Finally compare the integrated behavior with the isolated baseline. Record which layers actually ran; passing one does not confer evidence on the others.
A development plan with explicit stop conditions
- Choose the failure to investigate. Start from an observed weakness: unexplained rejections, uncertain input timing, unexpected costs or missing operational evidence. Write one question and a result that would disprove the proposed improvement. Avoid introducing several new filters at once.
- Preserve the baseline. Save code and artifact hashes, platform build, symbol specification, broker/account mode, data interval, every input, tick model and cost assumptions. Retain unsuccessful runs as carefully as the attractive ones. Record why a run is not comparable instead of silently dropping it.
- Inject the failure before tuning performance. Force stale data, rejected requests and unavailable external services in a controlled environment. The immediate acceptance criterion is an understandable, recoverable state. Prevent an uncertain outcome from triggering another action before its identity and result are reconciled.
- Measure benefit and exclusion cost together. Count prevented failures and otherwise valid decisions that the new rule would block. Use the same candidate set where possible. Compare like-for-like exposure and avoid presenting fewer trades as proof of improved selection.
- Reserve an untouched evaluation. Choose settings before inspecting its outcomes. If you revise the design after viewing the result, mark that dataset as development evidence and obtain a new evaluation. Separate performance claims from correctness claims in the final decision.
- Finish with keep, revise or reject. List what passed, what failed and what remains unknown. A useful research week can end by rejecting an attractive idea. Keep the evidence sufficient for another developer to reproduce that decision without needing your terminal's hidden state.
Scope, source dates and limits
This edition continues our 9 November research, which closed at 06:07:55 UTC that day. Newly selected articles are dated 10–13 November; the current source check is 16 November. The selected articles add to the five applied studies reviewed last week. For earlier context on signals, sequence risk and connected execution, see the 5 November full report.
Official release histories, original author articles on MQL5 and the platform documentation were consulted directly, including methods, results and limitations rather than relying on index summaries. Article dates follow the publisher's display; an unstated publication timezone is not inferred. The eight selected articles are an editorial selection, not an exhaustive inventory of every forum message, language edition or broker build.
Article implementations and trading experiments were not independently executed for this report. Their findings remain author-reported observations with the stated settings and limits. Proposed experiments above are our analysis, not completed tests or instructions to deploy a strategy. No technique reviewed here establishes future profitability. The practical objective is evidence that distinguishes an unavailable result, a failed check and a confirmed outcome.
Forwarded this MetaTrader edition?
Get your own copy through Trading Development, our shared newsletter for MetaTrader, MQL and TradingView.
Select Trading Development in the free signup flow. Already a member? Sign in and update your email preferences. Unsubscribe anytime.