Retraining Cadence

Retraining cadence is the second-most-important documentation requirement for ML-augmented EAs (after the training methodology itself). The cadence directly determines how long the model's edge persists in live deployment.

Why ML models decay without retraining:

• **Microstructure evolution**: forex market microstructure shifts continuously as broker execution systems update, liquidity provider panels change, algorithmic competition increases, and macro regimes transition. A model trained on 2022-2024 data may not reflect 2026 microstructure • **Concept drift**: the relationships between input features and target outcomes (trade success/failure) change over time. A feature that was predictive in 2023 (e.g. specific volatility patterns) may be uninformative or anti-predictive by 2025 • **Competitive arbitrage**: when many EAs use similar ML features, the patterns the models exploit are increasingly arbitraged away. Edge from any given feature decays as more participants identify and trade it

Typical retraining cadences and their implications:

• **Weekly**: highest engagement; appropriate for high-frequency scalping where micro-trends matter. Operational overhead is significant — vendor must maintain training infrastructure, validate each retrain run, and version-control deployed models. GoldStrike AI's weekly retraining is the editorial gold standard • **Monthly**: standard for ML-filter products; balances engineering overhead against model freshness. Adequate for most strategy classes; insufficient for the most latency-sensitive scalping • **Quarterly**: lower-engagement cadence; acceptable for trend-following and mean-reversion strategies where microstructure shifts are less consequential. Below the editorial standard for most-profitable-tier deployment • **Annual or none**: insufficient for serious ML-augmented EAs in 2026; signals either poor engineering investment or vendor abandonment risk

Retraining cadence and vendor accountability:

• Vendors should publish their retraining cadence explicitly in product documentation • Vendors should annotate the verified live account with model-version markers showing when retrains were deployed (GoldStrike AI does this on its MyFXBook page) • Buyers should monitor whether the published cadence is actually maintained — vendor commitment to weekly retraining is only valuable if the cadence is actually followed in practice • Vendors who reduce cadence mid-deployment without explanation are signal of either operational issues or vendor disengagement; either way, the EA's edge is at risk

For EA buyer evaluation:

• **Pre-purchase**: confirm the vendor documents retraining cadence in product specifications • **Initial deployment**: verify that recent model updates (within the last cadence window) are visible on the verified account • **Ongoing monitoring**: monthly verification that cadence is being maintained; vendors who skip cadence cycles are at increased decay risk • **Performance attribution**: when EA performance degrades, check whether degradation correlates with reduced retraining cadence; if so, the issue may be model decay rather than market regime

For ML-filter products specifically (Scalperology AI, Smart Robot AI), the underlying rule layer provides some protection against ML decay — when the ML filter degrades, the rule layer still has its native edge (just with lower win rate). For end-to-end ML products (GoldStrike AI), retraining cadence is the single most important operational discipline; without it, the entire edge is at risk.