Discussing the article: "Market Microstructure in MQL5 (Part 7): Regime Classification"

[MetaQuotes]

Check out the new article: Market Microstructure in MQL5 (Part 7): Regime Classification.

We integrate eleven one-minute microstructure measurements from Parts 2–6 into a composite regime label with confidence and direction. A rule-based RegimeClassifier() assigns one of six regimes—Normal, Stressed, Noisy, Informed, Trending, Mean-Reverting—using empirically derived thresholds from 514 NQ M1 sessions (May 2024–May 2026). The deliverable includes MARKET_REGIME, RegimeAnalysis, and PopulateRegimeAnalysis(), enabling position sizing, stop placement, and signal filtering from a single call.

Part 1 built the defensive foundation. Parts 2 and 3 measured whether price has memory. Part 4 measured what volatility is doing and whether it persists. Part 5 decomposed price variation into noise and signal. Part 6 measured the direction of informed flow. Parts 2-6 together produce eleven distinct measurements of the market environment on the one-minute timeframe.

Each measurement is individually informative. Together they are correlated, partially redundant, and impossible to act on simultaneously in real time. A position-sizing system cannot respond to eleven separate signals. What is needed is a reduction: a single composite label that summarises the current session environment, and a confidence score that reflects how reliably the available measurements agree.

Part 7 implements this reduction. RegimeClassifier() takes the outputs of Parts 2-6 and assigns one of six regime labels (Normal, Stressed, Noisy, Informed, Trending, Mean-Reverting). It also returns a confidence score in [0, 1] and a composite directional score in [-1, +1]. The classification boundaries are derived empirically from 514 NQ M1 sessions, and the six labels correspond to distinct trading environments that warrant different responses in position sizing, stop placement, and signal filtering.

The classifier uses a priority-ordered rule set rather than a statistical model. This is a deliberate design choice. A machine learning classifier (k-means, random forest, hidden Markov model) would require out-of-sample validation, which is not feasible within a single article. A rule-based classifier is fully transparent, directly connected to the empirical thresholds derived in Parts 4-6, and immediately actionable without a training step. The rules are documented in full, and readers are encouraged to recalibrate the thresholds on their own instruments and sample periods.

Author: Max Brown