Princeec13/mql5

Bifurcar 1

Princeec13 0b557e494f

2026-02-05 01:22:42 -05:00

16 KiB

Original Enlace permanente Blame Histórico

Policy & Exploration Guide — DualEA System

ML policy gating, fallback modes, and exploration system

Policy System Overview
Policy Structure
Policy Gating Logic
Policy Fallback Modes
Policy Scaling
Exploration Mode
Exploration Caps
Troubleshooting

Policy System Overview

Purpose

The policy system allows ML models to influence trading decisions through:

Confidence gating: Block low-confidence predictions
Parameter scaling: Adjust SL/TP/lots based on model output
Risk management: Dynamic position sizing based on predicted outcomes

Files

policy.json: ML-generated trading policy

Path: Common/Files/DualEA/policy.json
Generated by: ML/policy_export.py
Structure: Per-slice (strategy|symbol|timeframe) probabilities and scaling multipliers

policy.backup.json: Last-known-good policy

Used for rollback on parse/plausibility failures
Updated on successful policy load

policy.reload: Trigger file for hot-reload

PaperEA_v2: Hot-reload supported via policy.reload + HTTP polling
LiveEA: Load on init only (restart required for updates)

Policy Structure

JSON Schema

{
  "version": "1.0",
  "generated_at": "2025-01-15T10:00:00Z",
  "model_hash": "abc123...",
  "train_window": {"start": "2024-01-01", "end": "2025-01-15"},
  "metrics": {
    "roc_auc": 0.72,
    "brier_score": 0.18,
    "expected_r": 0.65
  },
  "min_confidence": 0.55,
  "slices": [
    {
      "strategy": "ADXStrategy",
      "symbol": "EURUSD",
      "timeframe": 60,
      "probability": 0.75,
      "sl_mult": 1.0,
      "tp_mult": 1.2,
      "lot_mult": 1.0,
      "trail_mult": 1.0
    },
    ...
  ]
}

Fields

Global:

version: Schema version
generated_at: Timestamp of policy generation
model_hash: Trained model identifier for provenance
train_window: Data range used for training
metrics: Model performance metrics
min_confidence: Global minimum confidence threshold

Per-Slice (as implemented in LiveEA; PaperEA_v2 uses minimal parsing):

strategy: Strategy name (e.g., "ADXStrategy")
symbol: Trading symbol (e.g., "EURUSD")
timeframe: Timeframe in minutes (e.g., 60 for H1)
p_win: ML model confidence (0.0-1.0)
sl_scale: Stop loss scaling multiplier (optional)
tp_scale: Take profit scaling multiplier (optional)
trail_atr_mult: Trailing stop ATR multiplier (optional)
confidence: Alternative confidence field (optional)

Policy Gating Logic

Configuration

input bool     UsePolicyGating = true;           // Enable ML policy gating
input bool     DefaultPolicyFallback = true;     // Enable fallback modes
input bool     FallbackDemoOnly = true;          // Restrict fallback to demo
input bool     FallbackWhenNoPolicy = true;      // Fallback if policy not loaded
input bool     FallbackWhenSliceMissing = true;  // Fallback if slice missing

Decision Flow

Is UsePolicyGating=true?
  No  → PASS (no policy gating)
  Yes → Continue

Is policy loaded (slices > 0)?
  No  → Check FallbackWhenNoPolicy
         Yes → Check FallbackDemoOnly
                Demo account → FALLBACK (neutral scaling)
                Live account → BLOCK
         No  → BLOCK (reason: "no policy loaded")
  Yes → Continue

Does exact slice exist (strategy|symbol|TF)?
  No  → Check FallbackWhenSliceMissing
         Yes → Check FallbackDemoOnly
                Demo account → FALLBACK (neutral scaling)
                Live account → BLOCK
         No  → BLOCK (reason: "policy slice missing")
  Yes → Continue

Is slice.probability >= policy.min_confidence?
  No  → BLOCK (reason: "confidence too low")
  Yes → PASS + Apply scaling

Code Example

bool ApplyPolicyGating(SignalData &signal, string &reason) {
    if(!UsePolicyGating) return true;
    
    // Check if policy loaded
    if(policyData.slices == 0) {
        return HandleFallback("no_policy", signal, reason);
    }
    
    // Find policy slice
    PolicySlice slice = policyData.Find(
        signal.strategy_name, 
        symbol, 
        timeframe
    );
    
    if(slice.probability < 0) {  // Slice not found
        return HandleFallback("slice_missing", signal, reason);
    }
    
    // Check confidence threshold
    if(slice.probability < policyData.min_confidence) {
        reason = StringFormat("confidence %.2f < %.2f",
                             slice.probability, policyData.min_confidence);
        return false;
    }
    
    // Apply policy scaling
    ApplyPolicyScaling(signal, slice);
    
    return true;
}

Policy Fallback Modes

Fallback: No Policy

Triggers when:

UsePolicyGating=true
Policy file not loaded OR policy.slices == 0
FallbackWhenNoPolicy=true

Behavior:

Check FallbackDemoOnly:
- If true and demo account → ALLOW with neutral scaling
- If true and live account → BLOCK
- If false → ALLOW with neutral scaling (any account)
Bypasses insights thresholds
Bypasses exploration caps
No SL/TP/lot/trail multipliers applied

Log Example:

FALLBACK: no policy loaded -> neutral scaling used for ADXStrategy on EURUSD/H1 demo=true

Telemetry:

policy_fallback,EURUSD,60,ADXStrategy,no_policy,true,true

Fallback: Slice Missing

Triggers when:

UsePolicyGating=true
Policy IS loaded (policy.slices > 0)
Exact strategy|symbol|timeframe slice not found in policy
FallbackWhenSliceMissing=true

Behavior:

Same as "No Policy" fallback
Allows trade with neutral scaling
Demo-only restriction if FallbackDemoOnly=true

Log Example:

FALLBACK: policy slice missing -> neutral scaling used for BollAverages on GBPUSD/H1 demo=true

Neutral Scaling

Definition: No adjustments applied, use base parameters.

sl_mult = 1.0
tp_mult = 1.0
lot_mult = 1.0
trail_mult = 1.0

Trade proceeds with:

Original SL/TP from strategy
Original lot size from risk calculation
Original trailing stop settings (if enabled)

Fallback Safety

Why demo-only by default?

Fallback is permissive (allows trades without ML validation)
Intended for data collection, not production live trading
In live, you want explicit policy coverage for all slices

When to disable FallbackDemoOnly?

After verifying policy coverage is comprehensive
When transitioning from demo to live with same strategy set
With explicit risk acceptance of trading without ML guidance

Best Practice:

Keep FallbackDemoOnly=true until policy proven
Monitor policy_fallback telemetry events
Aim to eliminate fallbacks by improving policy coverage

Policy Scaling

Policy Scaling (LiveEA Only)

Note: Full policy scaling with per-slice multipliers is implemented in LiveEA. PaperEA_v2 has minimal policy parsing (checks min_confidence only).

Scaling Application (LiveEA):

void ApplyPolicyScaling(SignalData &signal, PolicySlice &slice) {
    // SL scaling
    double slDistance = MathAbs(signal.entry_price - signal.stop_loss);
    if(signal.direction == 1) {  // Buy
        signal.stop_loss = signal.entry_price - (slDistance * slice.sl_scale);
    } else {  // Sell
        signal.stop_loss = signal.entry_price + (slDistance * slice.sl_scale);
    }
    
    // TP scaling
    double tpDistance = MathAbs(signal.take_profit - signal.entry_price);
    if(signal.direction == 1) {  // Buy
        signal.take_profit = signal.entry_price + (tpDistance * slice.tp_scale);
    } else {  // Sell
        signal.take_profit = signal.entry_price - (tpDistance * slice.tp_scale);
    }
    
    // Trailing scaling (if enabled)
    if(TrailEnabled && slice.trail_atr_mult > 0) {
        // Apply trail ATR multiplier
    }
}

Scaling Examples

Conservative Scaling (low confidence):

{
  "probability": 0.60,
  "sl_mult": 0.8,   // Tighter stop
  "tp_mult": 1.5,   // Wider target (better RR)
  "lot_mult": 0.5,  // Smaller position
  "trail_mult": 0.9 // Tighter trail
}

Aggressive Scaling (high confidence):

{
  "probability": 0.85,
  "sl_mult": 1.2,   // Wider stop (more breathing room)
  "tp_mult": 0.8,   // Closer target (take profits faster)
  "lot_mult": 1.5,  // Larger position
  "trail_mult": 1.0 // Normal trail
}

Neutral Scaling (fallback):

{
  "probability": 0.0,
  "sl_mult": 1.0,
  "tp_mult": 1.0,
  "lot_mult": 1.0,
  "trail_mult": 1.0
}

Exploration Mode

Purpose

Bootstrap insights for strategy|symbol|timeframe slices that lack sufficient historical data.

Without exploration:

New slices blocked by insights gating (no data → can't pass thresholds)
Chicken-and-egg problem: can't trade → can't collect data → can't build insights

With exploration:

Limited trades allowed for no-data slices
Caps prevent excessive exposure
Data collected → insights built → insights gating takes over

When Exploration Triggers

Conditions:

UseExploration=true
Insights gating is enabled (UseInsightsGating=true)
No slice exists in insights.json for this strategy|symbol|timeframe
Exploration caps not exceeded

Important: Exploration bypass ONLY when slice truly missing. If slice exists but fails thresholds (low win rate), it is BLOCKED (no bypass).

Configuration

input bool     ExploreOnNoSlice = true;          // Enable exploration when no slice exists
input int      ExploreMaxPerSlicePerDay = 100;   // Daily cap per slice (default 100)
input int      ExploreMaxPerSlice = 100;         // Weekly cap per slice (default 100)

Note: Previous documentation listed defaults of 2/3. Actual code defaults are 100/100, effectively unlimited for most practical purposes. The UseExploration input does not exist; exploration is controlled via ExploreOnNoSlice.

Exploration Caps

Cap Types

Daily Cap: ExploreMaxPerSlicePerDay

Resets at midnight (00:00 server time)
Per-slice basis (each strategy|symbol|TF tracked separately)
Default: 100 trades/day/slice

Weekly Cap: ExploreMaxPerSlice

Resets on Monday
Week bucket = Monday of the week (yyyymmdd format)
Default: 100 trades/week/slice

Counter Persistence

Files:

Common/Files/DualEA/explore_counts_day.csv
Common/Files/DualEA/explore_counts.csv

Format:

key,date_yyyymmdd,count
ADXStrategy|EURUSD|60,20250115,2
BollAverages|GBPUSD|60,20250115,1

For weekly:

key,week_monday_yyyymmdd,count
ADXStrategy|EURUSD|60,20250113,3

Cap Checking

bool CheckExplorationCaps(string strategy, string symbol, int tf, string &reason) {
    string sliceKey = strategy + "|" + symbol + "|" + IntegerToString(tf);
    
    // Load counters
    int dayCount = LoadExploreCountDay(sliceKey);
    int weekCount = LoadExploreCountWeek(sliceKey);
    
    // Check daily cap
    if(ExploreMaxPerSlicePerDay > 0 && dayCount >= ExploreMaxPerSlicePerDay) {
        reason = StringFormat("explore_cap_day (day=%d/%d, week=%d/%d)",
                             dayCount, ExploreMaxPerSlicePerDay,
                             weekCount, ExploreMaxPerSlice);
        telemetry.Event("explore_block_day", sliceKey, dayCount, weekCount);
        return false;
    }
    
    // Check weekly cap
    if(ExploreMaxPerSlice > 0 && weekCount >= ExploreMaxPerSlice) {
        reason = StringFormat("explore_cap_week (day=%d/%d, week=%d/%d)",
                             dayCount, ExploreMaxPerSlicePerDay,
                             weekCount, ExploreMaxPerSlice);
        telemetry.Event("explore_block_week", sliceKey, dayCount, weekCount);
        return false;
    }
    
    // Increment counters
    IncrementExploreCountDay(sliceKey);
    IncrementExploreCountWeek(sliceKey);
    
    // Allow exploration
    Log(StringFormat("GATE: explore allow %s on %s/%d (day=%d/%d, week=%d/%d)",
        strategy, symbol, tf,
        dayCount+1, ExploreMaxPerSlicePerDay,
        weekCount+1, ExploreMaxPerSlice));
    
    telemetry.Event("explore_allow", sliceKey, dayCount+1, weekCount+1);
    
    return true;
}

Resetting Caps

Manual Reset: Delete counter files:

Remove-Item "C:\Users\<you>\AppData\Roaming\MetaQuotes\Terminal\Common\Files\DualEA\explore_counts*.csv"

Automatic Reset:

Daily: Midnight (00:00 server time)
Weekly: Monday 00:00

Interaction with NoConstraintsMode

When NoConstraintsMode=true:

Insights gating: BYPASSED
Exploration caps: BYPASSED (trades allowed regardless of caps)
Exploration counters: Still incremented for telemetry

Troubleshooting

Issue: Policy fallback always triggering

Symptoms:

FALLBACK: no policy loaded -> neutral scaling for ADXStrategy on EURUSD/H1 demo=true

Diagnosis:

Check if policy.json exists:

Test-Path "C:\Users\<you>\AppData\Roaming\MetaQuotes\Terminal\Common\Files\DualEA\policy.json"

Check policy load logs (enable DebugPolicy=true):
```
Policy loaded: 0 slices, min_confidence=0.00
```

Verify policy.json content:

Get-Content "...\DualEA\policy.json" | ConvertFrom-Json | Select-Object -ExpandProperty slices | Measure-Object

Solutions:

Run ML/policy_export.py to generate policy
Copy policy.json to Common Files
Verify JSON is well-formed (no parse errors)

Issue: Policy slice missing fallback

Symptoms:

FALLBACK: policy slice missing -> neutral scaling for BollAverages on GBPUSD/H1 demo=true

Diagnosis:

Check policy slices:

Get-Content "...\policy.json" | ConvertFrom-Json | Select-Object -ExpandProperty slices | 
    Where-Object {$_.strategy -eq "BollAverages" -and $_.symbol -eq "GBPUSD" -and $_.timeframe -eq 60}

No result → slice missing from policy

Solutions:

Collect more training data for this slice
Re-train model with broader coverage
Accept fallback for new slices (data collection mode)

Issue: Exploration caps reached immediately

Symptoms:

GATE: blocked ADXStrategy on EURUSD/H1 reason=explore_cap_day (day=2/2, week=2/3)

Diagnosis:

Check counters:

Import-Csv "...\explore_counts_day.csv" | Where-Object {$_.key -like "*ADXStrategy|EURUSD|60*"}

Verify caps not too restrictive:

ExploreMaxPerSlicePerDay = 2  // Very conservative

Solutions:

Increase caps (e.g., 5/10 for PaperEA, 1/2 for LiveEA)
Reset counters manually (delete CSV files)
Use NoConstraintsMode=true for initial data collection (bypasses all caps)

Issue: Exploration not triggering

Symptoms:

GATE: blocked ADXStrategy on EURUSD/H1 reason=below_winrate (WR=0.42 < 0.50)

(Slice exists but fails threshold, no exploration bypass)

Diagnosis: Exploration only triggers when slice missing entirely, not when slice exists but fails thresholds.

Solutions:

This is correct behavior (no-slice-only bypass)
To allow trading despite low performance:
- Lower insights thresholds temporarily
- Use NoConstraintsMode=true
- Delete slice from insights.json to force exploration
- Wait for performance to improve

Issue: NoConstraintsMode not working

Symptoms: Still seeing gate blocks despite NoConstraintsMode=true

Diagnosis:

Verify setting applied:

if(NoConstraintsMode) Print("NoConstraintsMode is TRUE");

Check if using unified system:

input bool UseUnifiedSystem = true;  // Required for NoConstraintsMode

Solutions:

Ensure UseUnifiedSystem=true
Restart EA after changing NoConstraintsMode
Check logs for shadow decisions: [SHADOW] gate=... result=block (but allowing)

See Also:

Configuration-Reference.md - Policy and exploration parameters
Execution-Pipeline.md - When policy/exploration evaluated
Observability-Guide.md - Policy/exploration telemetry events

16 KiB Original Enlace permanente Blame Histórico

Policy & Exploration Guide — DualEA System

Table of Contents

Policy System Overview

Purpose

Files

Policy Structure

JSON Schema

Fields

Policy Gating Logic

Configuration

Decision Flow

Code Example

Policy Fallback Modes

Fallback: No Policy

Fallback: Slice Missing

Neutral Scaling

Fallback Safety

Policy Scaling

Policy Scaling (LiveEA Only)

Scaling Examples

Exploration Mode

Purpose

When Exploration Triggers

Configuration

Exploration Caps

Cap Types

Counter Persistence

Cap Checking

Resetting Caps

Interaction with NoConstraintsMode

Troubleshooting

Issue: Policy fallback always triggering

Issue: Policy slice missing fallback

Issue: Exploration caps reached immediately

Issue: Exploration not triggering

Issue: NoConstraintsMode not working

16 KiB

Original Enlace permanente Blame Histórico