Magister Mentis
How It Really Works
There’s a basic distinction between constructing a buying and selling robotic and constructing a call engine.
Most Knowledgeable Advisors are procedural. They await a situation — an indicator cross, a threshold breach, a sample completion — and so they execute a predefined response. They’re deterministic. They react to guidelines.
Magister Mentis was not designed as a rule executor. It was designed as a probabilistic inference system.
As a substitute of asking, “Did RSI cross 30?”, it asks:
Given the present multidimensional state of the market, what’s the likelihood distribution of future directional bias?
Every part that follows stems from that single query.
The Market as a Classification Drawback
At its core, Magister Mentis treats every closed candle as a characteristic vector. The mannequin doesn’t try and forecast worth magnitude. It performs classification.
For every determination cycle, it outputs three values:
The impartial class shouldn’t be ornamental. It’s vital. It acts as a structural uncertainty absorber. Markets are usually not binary methods, and forcing them into up/down logic typically results in overtrading in ambiguous zones.
The system trades solely when one directional class meaningfully dominates the distribution.
The 17-Dimensional Characteristic House
Each inference begins with a 17-feature vector derived from the final closed candle.
These options are usually not uncooked values. Nearly all are volatility-normalized to cut back regime bias. This was a deliberate design determination throughout mannequin coaching to mitigate overfitting to particular volatility environments.
The characteristic teams embrace:
Candle construction asymmetry (wick imbalance)
EMA directional relationship (50 vs 200), normalized by ATR
EMA displacement from worth
RSI
ADX
ATR
Greenback index delta (actual image if accessible, inverse proxy in any other case)
Quantity delta normalized
Structural stability rating derived from Singular Worth Decomposition
Quick-term normalized returns (1, 3, 6 bars)
Normalized candle vary
Cyclical encodings of hour and weekday utilizing sine/cosine transforms
The SVD-based stability metric deserves particular point out. Quite than relying solely on momentum or oscillation, the system analyzes latest candle construction coherence. A steady singular worth distribution signifies structural consistency; instability signifies regime fragmentation.
This characteristic engineering layer is the place most robustness is created. The mannequin is barely pretty much as good because the house it observes.
Regime Detection: The Macro Filter
Earlier than the system decides path, it decides context.
A secondary mannequin evaluates whether or not the present market is trending or ranging. It outputs a likelihood:
P(pattern)
This isn’t used instantly as a buying and selling sign. As a substitute, it determines routing.
To stop flip-flopping between regimes on marginal likelihood shifts, hysteresis thresholds are utilized:
This ensures structural stability in mannequin choice.
Specialist Routing
As soon as regime is set, inference could also be routed to a specialist:
Every mannequin has its personal calibration fixed. This structure permits specialization with out fragmenting execution logic.
AUTO mode routes dynamically. Guide override is feasible.
Likelihood Calibration
Uncooked mannequin outputs are hardly ever well-calibrated. Machine studying classifiers are usually overconfident.
Magister Mentis applies temperature scaling:
Every mind can use a distinct temperature fixed decided throughout validation.
This improves threshold reliability and prevents systematic bias from overconfident outputs.
This calibration step is without doubt one of the causes the EA behaves extra persistently throughout datasets.
Entry Logic: Dominance and Thresholds
A commerce shouldn’t be triggered just because a likelihood is excessive.
Two situations have to be glad:
The directional likelihood should exceed its confidence threshold.
It should dominate competing lessons by not less than a configurable margin.
Formally:
Purchase requires:
Promote requires the symmetric situation.
This eliminates ambiguous distributions reminiscent of:
P(Up)=0.62, P(Down)=0.59, P(Impartial)=0.05
Which could look sturdy however are structurally unstable.
Moreover, just one motion is permitted per candle. This prevents intra-bar churn and overreaction.
Danger as a Operate of Confidence
Place sizing can function in two modes:
In risk-based mode, lot dimension is derived from:
The scaler will increase place dimension reasonably if likelihood exceeds the edge considerably. It’s bounded and capped by:
Dealer quantity limits
Free margin constraints
A tough most lot cap
There isn’t a martingale logic. There isn’t a geometric publicity improve. Loss doesn’t set off dimension enlargement.
Adaptive Cease Logic
Stops and targets could also be ATR-based:
SL = ATR × multiplier
TP = ATR × multiplier
This ensures volatility-adjusted threat management, significantly necessary for devices reminiscent of XAUUSD the place static stops turn out to be structurally inconsistent throughout regimes.
Commerce Administration Layer
Unbiased of entry logic, the system can apply:
These mechanisms function after commerce entry and are decoupled from Sensible Exit (RSI exhaustion).
Institutional Shields
Two laborious safety methods exist:
If both is breached:
All positions are closed
The EA locks till reset
Delicate filters embrace:
Unfold filter
Latency filter
Buying and selling hour window
Optionally available Friday shut
These filters forestall structurally poor execution environments from degrading long-term efficiency.
Coaching and Overfitting Management
The mannequin pipeline consists of:
Multi-year information
Separate coaching, validation, and holdout take a look at units
Optuna-based hyperparameter optimization
Likelihood calibration on validation
Characteristic normalization
Regime specialization
The target was not most backtest revenue. The target was steady likelihood habits throughout unseen information.
Overfitting was addressed by way of:
No grid logic was launched to artificially clean fairness curves.
Execution Mannequin
All inference is executed regionally inside MT5 through ONNX runtime.
There’s:
The EA operates autonomously as soon as connected.
Beta Tester Program
Magister Mentis is presently in staged analysis.
For the primary two analysis phases, a structured beta testing window shall be accessible.
On roughly the twenty sixth day of every month, one designated “Beta Testing Day” shall be introduced.
For that day solely:
The worth shall be diminished by 50%.
Entry is granted in change for structured suggestions.
Beta contributors are anticipated to:
Ahead take a look at responsibly.
Present written suggestions.
Go away an trustworthy evaluate reflecting actual expertise.
Participation is proscribed and manually dealt with.
merchants might ship a non-public message on that day to use for the beta window.
This strategy permits real-world suggestions with out compromising long-term product positioning.
Magister Mentis shouldn’t be designed to commerce always.
It’s designed to guage, filter, and act solely when likelihood alignment and structural affirmation intersect.
It waits.
Then it executes.
