BRANDYCANDY

⚙ System

System Console — Build 11 v5

Live ring buffer of all bot output — stdout, logging, and unhandled exceptions. Updates every 5s. Filter by severity/source/search. Use "Copy All" to share with AI or paste into a handover.
Recommended filter: WARNING+ for day-to-day monitoring, ERROR+ to only see failures, DEBUG+ when actively debugging a new feature.

auto-refresh

Loading…

Severity filter

DEBUG+	Everything — use when reproducing a specific bug
INFO+	Trades, startups, MT5 connects, [P10] gate outcomes
WARNING+	Recommended default — daily monitoring
ERROR+	Only show failures — quickest "is anything broken?" check
CRITICAL	Catastrophic only — should be empty

Line colour

Red — blocked trade, crash, error, failure

Green — trade placed, training complete, bot recovered

Amber — partial qualify, cooldown, warning

Cyan — scanner/multi-TF/migration info

Muted — routine background noise

P10 gates (auto-trade pipeline)

Gate 2	SMC_AMD sequence stage < required
Gate 3	Non-SMC strategy blocked by SMC-only toggle, or Asia toggle off
Gate 4	Strategy score below min (`p10_X_score_min`)
Gate 5	Ensemble disagreement between parameter sets
Gate 6	SP-4 unvalidated (need 20+ tagged trades)
Gate 7	Upcoming high-impact news
Gate 7b	Post-news cooldown active
Gate 7c	DD velocity breach (account or per-strategy)
Gate 8	Session mismatch (SMC/Breakout need London/NY; MeanRev needs Asia/NY PM)
Gate 9	Regime mismatch (SMC/Breakout need trending/expansion; MeanRev needs ranging)
Gate 10	Spread above `spread_mult`× typical
Gate 11	Instrument cooldown active (default 60m)
Gate 12	All accounts at daily DD hard stop
Gate 13	Instrument already has an open trade
Item 5	Funded consistency block (best day ≥15% of total PnL)
Gate 14	ML Layer A win probability below threshold
Gate 15	ML Layer B continuation prob or phase filter

Tip: type [P10] in the search box to see only auto-trade events. Combine with WARNING+ to see only blocks & execs.

Bot Control & Health

Restart the bot, check system health, and see library versions — no RDP needed.

Uptime

—

DB Size

—

Trades

—

Paper Trades

—

Events

—

PyTorch

—

XGBoost

—

Waitress

—

Restart Live Bot

Restarts the live trading bot to activate code changes. The Flask process exits cleanly; the watchdog detects the process is dead and respawns it within ~30 s. Does NOT affect running backtests — those run as separate subprocesses launched by run_backtest.py and continue independently.

Watchdog Controls

Start, stop, or restart the watchdog PowerShell process. Tries the Windows Scheduled Task first (BrandycandyWatchdog), falls back to direct script launch / process kill.

Click "Check Status" to see watchdog state.

Background Jobs

ML training + backfill runs asynchronously. Pending and running jobs appear here with progress. Cancel in flight without restarting the bot.

auto-refresh

No jobs yet.

General

Trader Name

Fatigue Warning (hours)

TP1 Split % (default 50)

Default Risk % (default 1)

Revenge Warning — losses before alert (default 2)

SP-4 Strategy Activation GateENABLED

On

When ON, strategies require 20+ closed journal trades per instrument before firing at full score — prevents signals on unvalidated setups.
Turn OFF if your journal trades have no strategy tags, or on a fresh install.

Settings Debug — Raw DB Values

Click Refresh to load raw DB values

System Info

API Key

Checking…

Saves to database — persists across restarts. Get your key from console.anthropic.com

Bot Control

⬆ DEPLOY FILESBot will restart automatically via wrapper script

Deploy Password:

Set a deploy password, then use brandycandytrading.com/deploy from any device to upload updated zip files. Bot auto-restarts after deploy.

Kill Zones

Session	UTC	NZ Time
Loading…

Prop Firm Rules

Profit target: 6%
Daily loss limit: 3%
Max trailing drawdown: 6%
Payout split: 80%
Min profitable days: 3 (≥0.5%)
Consistency: best day ≤ 20% of total
Payout cycle: every 14 days

MT5 Connection

Checking MT5 status…

▸ Manual credentials (if account not in list)

MT5 Server Name

MT5 Login

MT5 Password

Scan Frequency (active / idle)

/

Active = chat used in last 60 min. MT5 candle pulls are local/free.

Terminal Manager

Loading terminals…

Check boxes = auto-launch on bot start. Force Launch opens the terminal without connecting. Secondary account imports now run every 5 min (configurable) instead of every 60s to prevent terminal reconnection storms.

MT5 Trade History Import

Account

Days back to import

MT5 must be connected first.

📊 Trading Behaviour

Initial SL Sizing (ATR)

Per-instrument-class ATR multipliers used when computing the initial stop-loss distance at entry.
Runtime trail uses Mode 3 (Structure Trail) + Adaptive Trail below; the legacy Mode-2 ATR runtime trail has been removed.

ATR Multiplier Defaults

XAUUSD

NAS100 / Indices

Forex Pairs

Oil (WTI/Brent)

Crypto

Silver (XAGUSD)

SL Management — Adaptive Trail replaces BE-at-1R when enabled

Profit-level-dependent SL ladder. Each band cuts more risk or locks more profit as price moves in your favour.
SL only tightens, never widens. Above 3R, control hands off to the structure trail (Mode 3).

Enable Adaptive TrailON by default — replaces the rigid BE-at-1R rule below

Gradient Bands (multiplier of original SL distance)

Negative = SL below entry (LONG; mirror for SHORT). Positive = SL above entry (locks profit). Example LONG entry 1.0800, SL 1.0780, dist=20pips. 1p5R_sl=0.3 → SL = 1.0800 + 0.3×20pips = 1.0806.

0.7R reached → SL @ (default −0.5)

1.0R reached → SL @ (default −0.2)

1.5R reached → SL @ (default 0.3)

2.0R reached → SL @ (default 0.8)

Above 3R: hand off to Structure Trail (Mode 3)

Presets

Conservative: wider SL — leaves more room (0.7R: −0.7 / 1R: −0.4 / 1.5R: 0.0 / 2R: 0.5).
Balanced: bot defaults (0.7R: −0.5 / 1R: −0.2 / 1.5R: 0.3 / 2R: 0.8).
Aggressive: tighter — locks profit sooner (0.7R: −0.3 / 1R: 0.0 / 1.5R: 0.5 / 2R: 1.0).

ⓘ Adaptive Trail is enabled — the legacy BE-at-1R rule (configured via early_trail_at_r) is bypassed for new positions. Disable Adaptive Trail above to fall back to BE-at-1R.

Signal Sensitivity Advanced

⚠ Changes affect signal frequency, not signal quality. Lowering thresholds = more signals, more false positives.

Strategy score minimums (SMC_AMD, Breakout, MeanRev) are now managed in the Execute tab → Scalp Panel.

Entry Filter Thresholds

FVG Wick Rejection % (default 0.40)

Breakout Body % (default 0.60)

Sweep Detection Sensitivity (ATR multiplier)

Controls how far price must exceed a swing high/low before it counts as a liquidity sweep. Lower = easier to detect. Check sweep debug logs before adjusting.

Forex (def 0.30)

Metals/XAU (def 0.20)

Indices (def 0.20)

Energy/WTI (def 0.25)

Crypto (def 0.20)

Performance Feedback Loop

⚠ Warning: Do NOT enable until ≥50 real closed trades are logged across multiple instruments AND the strategy engine (Items D/E/F/H) has been verified live for ≥2 weeks. Enabling early will cause the system to optimise against noise and actively degrade signal quality.

Enable Feedback LoopOFF by default

When enabled, the system analyses closed trade win rates per instrument×strategy and proposes parameter threshold adjustments. Requires ≥30 trades per combination to activate. Rolls back automatically if performance drops >10% within 20 trades of a change.

SL Management — Mode 3 (Structure Trail) BETA — off by default

Automatically trails SL to the most recent 1H swing low (LONG) or swing high (SHORT) when a better level exists.
Runs every 60 s inside the MT5 monitor loop. SL can only move in your favour.
⚠ Enable only after verifying MT5 connectivity and reviewing the place_order() lot validation fix.

Global Structure TrailLoading…

Per-position overrides available on each open position card in the Execute tab.
Swing detection: 5-candle lookback each side on 1H chart (configurable in code via lookback=5).

v5.17 — Execution Quality (Task 4)

Spread gate measures a rolling EMA of spread per instrument. When the live spread at fire time exceeds multiplier × EMA, placement delays and re-checks once. If still wide, the trade is skipped. Target: kill the 0.1–0.3R slippage on wide-spread entries. Defaults ON.
Overlap block hard-blocks nominated instruments during the LDN/NY overlap window (12:00–16:00 UTC). Default OFF — opt in by listing instruments.

Spread-Aware Gate (4a)

Enable spread gate

Trigger Multiplier (live > N × EMA)

Delay on Wide Spread (sec)

Min Samples Before Gate Active

EMA Alpha (0.05–0.5 typical)

Higher alpha = EMA reacts faster to recent samples. Lower = smoother but slower to adjust after a regime change.

LDN/NY Overlap Block (4b)

Enable overlap block

Instruments to block (comma-separated)

Start hour (UTC)

End hour (UTC)

Window is inclusive-start, exclusive-end. Only blocks instruments in the list; others trade normally.

v5.22 — Limit-Order-Only Mode

What this does: wires mt5_service.get_auto_order_type() into the AUTO-trade fire path. When the scanner entry is far enough from live price to count as a pullback, a LIMIT order is placed at the scanner entry instead of a market order — avoiding slippage on volatile instruments.
When to enable: if you're seeing slippage costs on NAS100 / XAUUSD or if filled prices consistently worse than scanner entry. Leave OFF for the Execute-tab manual chain — it has its own order-type selector.
Per-strategy: turn on only for strategies where pullback entries are the norm. Breakout_Continuation should stay OFF — breakouts need immediate fills.
Stop-order handling: when get_auto_order_type returns "stop" (breakout above live price), the trade is SKIPPED — market-on-breakout has the same slippage problem we're avoiding.
Failure mode: if MT5 rejects the limit or the resolver fails, the trade falls back to market. Zero silent drops.

Master: enable limit-order-only mode for AUTO trades

SMC_AMD (recommended ON) Breakout_Continuation (recommended OFF) Mean_Reversion_Trap (recommended ON)

Limit Expiration (minutes)

Limit orders auto-expire. 15m is the sweet spot — long enough for pullbacks, short enough that stale setups don't fire.

Tally (all time):— limit orders fired · — stop-route skips

Loading EMA state…

v5.18 — Partial-Close Ladder (Task 5a)

Replaces the current single 50%-at-TP1 partial with a configurable 3-rung ladder. Each rung closes a percentage of the original lot size at a different profit level. Remainder rides with BE+buffer trail.
Default OFF — enable deliberately after observing baseline. Broker TP shifts past TP2 as safety net when ladder is active.

Rung Percentages

Enable partial-close ladder

Rung 1 — % at TP1

Rung 2 — % at TP1 + 1.5R

Rung 3 — % at TP2

Rider: 10% rides with BE+buf trail

Advanced

Rung 2 R-Multiplier (1.0=TP1, 2.0=TP2)

Broker TP Safety Buffer (R past TP2)

Broker TP shifts to TP2 + buffer × SL_distance when rung 1 fires. If the bot stops monitoring, the broker still catches the position at this safety TP.

Continuous Shadow Mode BETA

Fires paper trades 24/7 alongside live trades for calibration data. Shadow trades are DB-only — never sent to MT5. Generates ground-truth data for calibrating the backtest engine (Phase 3). Recommended: run for 30+ days before Phase 3.

Enable Continuous Shadow Mode —

Off-Hours Penalty System L1 ACTIVE · L2 SHADOW

Soft-penalty applied to strategy_score at Gate 4 when a setup fires during Off-Hours and spread is elevated vs the instrument's normal-hours baseline. Coexists with Gate 8 (hard session block) and Gate 10 (hard spread cap). News-aware: penalty is reduced by the configured factor when a high-impact news event is imminent.

Layer 1 Configuration (Active)

Enable Layer 1 (spread-based dynamic penalty)

Base Penalty (default −2.0)

Max Multiplier (default 5.0)

Baseline Window (days, default 7)

News Reduction Factor (default 0.5)

Layer 1 Live Status

Active baselines

—

Missing baselines view list

—

Penalties applied (24h)

—

Blocked by penalty (24h)

—

Passed despite penalty (24h)

—

Layer 2 Configuration (Shadow Mode)

StatusSHADOW MODE

Auto-activation threshold0 / 1000 Off-Hours trades

When the threshold is met AND Layer 2 outperforms Layer 1, Layer 2 will auto-activate after a 1 hour notification window (you'll receive an email). Manual override available below.

Layer 2 Per-Class Weight Overrides (Advanced) — different factor weights for forex vs indices vs crypto

Each asset class can have its own factor weight distribution (sum should equal 1.0 for clean baselines). Defaults are tuned to each class's liquidity dynamics. Changes are logged to parameter_change_log.

Layer 3 ML Model Status NOT_READY— gradient-boosting model trained on accumulated L2 shadow data

Training data progress0 / 10000 Off-Hours samples

Manual controls (advanced)

🧠 ML & Optimisation

ML Layer A — Trade Classifier SHADOW

XGBoost model predicting trade win probability from scanner snapshots. Trains on your journal data. NOT used in live trading until you explicitly enable it.

Data Readiness

Click "Check Status" to see data readiness.

Training

▸ Training History & Gate Eligibility

Loading…

Historical Backfill Engine

Replays scanner against historical MT5 candles to generate training data. 6 months across all 72 instruments typically produces 2000–5000+ snapshots.

Months

Instruments

Status

Idle

Window (H1 bars)

Step (bars)

Min Score

Lookahead (bars)

Window: H1 bars of context per evaluation (50–500, rec 200). More = better indicator calc, slower. Step: Bars between evaluations (1–24, rec 4). Lower = more snapshots but slower. 1 = every hour, 4 = every 4h. Min Score: Minimum signal score to capture (0.0–3.0, rec 0.5). Lower = more data for ML. 0.0 captures everything with a direction. Lookahead: H1 bars to check for TP/SL outcome (5–50, rec 20). Longer = more wins hit TP but less realistic for day trading.

ML Layer B — Sequence Model (LSTM) SHADOW

Learns temporal patterns across sequences of market events (sweep → displacement → BOS → FVG → retest). Answers "what is the market doing next?" — complementary to Layer A's "should I trade this setup?". Requires PyTorch (pip install torch). NOT used in live trading until you explicitly enable. MFE Extension adds a second model head that predicts the distribution of Maximum Favourable Excursion so TP1/TP2 can eventually become data-driven.

Full ML Backfill (recommended on first deploy / after long downtime)

Runs Layer B scanner-replay (paper_trades snapshots feeding Layer A context), multi-timeframe event-stream detection (M5/M15/H1/H4/D1 across all instruments → event_stream for LSTM sequences), MFE / true-MFE / MAE backfill (24h post-close window for losers — Principle 8 self-bias correction), and finally a Layer A feature-CSV refresh — all in one orchestrated background job.

When to run: first deploy on a new VPS, or after ≥ 1 week of bot downtime. Safe to run while the bot trades — uses job_runner, polled via /api/jobs/<id>. Typical runtime 20–60 min depending on months + instrument count.
Requires: MT5 connected (scanner replay + MFE both pull historical candles via copy_rates_range). Pandas + PyTorch not required for the backfill itself — only for subsequent Layer B training.
Failure mode: each sub-step is isolated. If MFE backfill fails, the preceding stages still wrote their data; re-run later. Individual per-instrument errors are logged but never abort the whole job.
Cancel: click Cancel on the row in the Background Jobs panel. Orchestrator checks cooperative-cancel between sub-steps and inside each inner loop.

All individual backfill buttons below still work if you want to run just one sub-step.

Months of history

Last run: —

Operational · Train / Predict / Diagnostics

Accuracy expectations as data grows:

• Below 10k sequences — model is immature, accuracy noisy

• 20k sequences — expect plateau around low-to-mid 80s

• 50k sequences — expect plateau around mid-to-high 80s

• 100k+ sequences — plateau holds; further gains require new features, not more data

Data Readiness

 Click "Check Status" to see event counts. 

Check Status: shows how many events have been logged, PyTorch version, and whether a model is trained. View Recent Events: pops up the last 30 events for an instrument — useful for debugging what the scanner is detecting.

Training & Live Inference

Train Layer B Model: trains the LSTM on all accumulated event sequences using the hyperparameters below. Takes a few minutes on CPU. Safe to run anytime — replaces the existing model on success, leaves old one untouched on failure. Predict Current Sequence: runs the trained model against the latest 20 events for a chosen instrument and returns phase prediction + continuation probability. Model Diagnostics: runs the trained model on a sample of historical sequences, reports label distribution, naive-baseline accuracy, per-class precision/recall, and confusion matrix. Tells you whether the model is learning real signal or just predicting the majority class.

Training thresholds: 500+ sequences minimum, 10,000+ for a usable model, 100,000+ for stable production use.

Multi-Timeframe Event Backfill (Build 12)

Replays historical candles through the event detectors on M5, H1, H4 and D1 timeframes. Expected outcome: 10k events → 100k–500k events in one operation, dramatically expanding Layer B's training data. Scanner is paused during backfill to avoid DB lock contention. Uses the cached MT5 candle data — no extra API calls.

Months

Timeframes

M5 H1 H4 D1

M5 step (bars)

Larger = fewer events, faster

Clean Regime Flapping: pre-v5.8 the regime_change detector had no hysteresis and would log every classifier flip, producing thousands of spurious events when ADX hovered near thresholds. Click this to purge any regime_change events logged within 30 minutes of a prior one on the same instrument+timeframe. The first click runs a dry-run preview. Safe to run anytime.

Run Diagnostics: evaluates the trained Layer B model and reports label distribution, naive baseline accuracy (always predict majority class), model accuracy, per-class precision/recall/f1, and the confusion matrix. Use this to sanity-check that the model's accuracy is genuine signal, not just an artifact of class imbalance. Takes ~1-2 minutes on a 5000-sequence sample.

Live Inference Gate (Auto-Trade Integration)

What does "inert" mean? When the gate is OFF, Layer B's LSTM is never loaded and never runs during live trading. No predictions, no blocking, no influence on auto-trade decisions whatsoever — the code path is skipped entirely. Event logging and training still work independently. You can turn the gate ON only after you've trained a model and want to test it.

Enable Layer B gate

Master switch. OFF = fully inert (no influence on trading).

Shadow mode (recommended)

When ON, Layer B predicts & logs but doesn't block trades. Use for weeks before enabling the real gate.

Min Continuation Prob (%)

Range: 50–90. Default: 65. Minimum probability the market will continue in the scanner's proposed direction. Higher = stricter (fewer trades, higher quality). Lower = more permissive.

Phase Filter

Which market phases allow trading. "Any" = no phase filtering. "Not uncertain" blocks only when the model is unclear. "Trending only" is strictest — only allows trades during clear trends.

⟡ ML Layer B — MFE Model (Staged Rollout)

Stage: —

Distribution heads for data-driven TP1/TP2. Backend auto-advances stages once each criterion is met — no operator decision burden.

Loading MFE status…

Stage 1

MFE Backfill

—

Stage 2

Model Heads

—

Stage 3

Shadow Mode

—

Stage 4

TP Display

—

Stage 5

Auto-Use TPs

—

About ML Layer B staged rollout

Layer B predicts the distribution of Maximum Favourable Excursion (MFE) for each setup. Once shadow-mode validation confirms calibration is healthy, TP suggestions become data-driven — derived from how similar historical trades actually ran, not from static R:R multipliers.

Stage 1 — MFE Backfill: walks every closed trade, computes observed MFE + bias-corrected true MFE (24h post-close window for losers). Idempotent.
Stage 2 — Model Heads: retrains Layer B with mean+std MFE heads using Gaussian NLL loss. Auto-enabled when backfill coverage > 95%.
Stage 3 — Shadow Mode: auto-trades log Layer B's suggested TPs. Trades still use scanner TPs. Auto-enabled at Stage 2 accuracy threshold.
Stage 4 — TP Display: scanner/calculator panels show suggested TPs alongside scanner's. Auto-enabled after 2+ weeks shadow data.
Stage 5 — Auto-Use TPs: auto-trades use Layer B's suggested TPs. Auto-enabled only after 4+ weeks favourable calibration.

Admin — manual ops

Force-stage-advancement (overrides auto-advance)

Force enable MFE model heads (Stage 2)Shadow mode (log only)Force show suggested TPs in UI (Stage 4)

Optimiser Instrument Profiles

Save and recall named sets of instruments for the Parameter Optimiser. Profiles persist across restarts.

Active: none

Parameter Optimiser ADVANCED

Grid-searches RSI / BB / ATR / EMA parameters per instrument using 5yr historical data.
80/20 walk-forward split. Auto-locks validated edges (WFA stable or insensitive params). Degraded instruments unlock and rebuild automatically.
⚠ Requires historical data import first. Select instruments below to run a targeted rebuild.

Optimised

—

instruments

Last Run

Never

Duration

—

seconds

Instrument Selector

Loading instruments…

0 selected

WFA Status Dashboard

✅ Validated(0)

⚠ Needs Attention(0)

❓ No Data / Building(0)

Historical Data Import

Drop MT5 CSV exports into brandycandy_fixed/historical/ then click Import.
Supports H1 and H4 exports from the BrandycandyExport script (2019–present).
Data is used by the Parameter Optimiser (the Parameter Optimiser) for statistically robust backtesting.

Total Rows

—

Instruments

—

DB Size

—

CSV Files Ready

—

🔔 Integrations & Alerts

🌐 MACRO DATA APIS (J6)

External feeds for macro fields the broker doesn't subscribe (DXY/VIX/indices, US Treasury yields). Computed proxies fall back when keys missing.

FRED API key

Free with registration at fred.stlouisfed.org/docs/api/api_key.html. Used for us_10y_yield, us_2y_yield, yield_curve_spread.

Twelve Data API key

Free tier (800 calls/day) at twelvedata.com/pricing. Used for DXY, VIX, US500, DE40, JP225 fallback.

⟡ OPERATIONAL ALERTS

Operator-supervision emails: critical bot problems (always on) and configurable state summaries. Uses a dedicated Gmail account with App Password (not your main Gmail). Generate app password here ↗

Send Alerts To (your email)

Alert Gmail Address (sender)

Gmail App Password

Bot Problem Alerts

ALWAYS ON

Immediate email on critical operator-attention events: MT5 disconnect · Flask crash · watchdog restart failure · DB errors · scanner stall >10min · auto-trade unexpectedly disabled · critical exceptions. Not queued, not throttled.

Last problem alert: —

State Summary Emails

Scheduled supervision summaries: P&L, balance per account, DD status, session performance, open positions, recent auto-promotions, anomalies. Markdown-formatted.

Daily

Hour (NZST)

Weekly

Day

Hour (NZST)

Monthly

Day of month

Hour (NZST)

Next scheduled summary: —

Phase Auto-Activation Alerts

Email the operator when a phase / layer transitions from shadow to active automatically (data-volume threshold + performance criteria met). The promotion is delayed 1 hour after the email so you can override.

Master toggle (all auto-activations) Off-Hours Layer 2 ready Off-Hours Layer 3 ready Off-Hours Layer 3 rolled back Vision Stage activations

📰 Live News Feed (Auto-active) —

Live ForexFactory economic calendar. Auto-cached hourly. Blocks scanner signals and auto-trades within 30 min of high/medium impact events. Active automatically — no config needed.

Loading…

Account size	Count
10K
25K
50K
150K

Strategy	Live score	Live risk
SMC_AMD	—	—
Breakout_Continuation	—	—
Mean_Reversion_Trap	—	—

Instrument	Live SL	Live TP
forex	—	—
xauusd (metals)	—	—
index	—	—
oil	—	—
crypto	—	—

Metric	Value	Metric	Value
Sharpe ratio	–	Sortino ratio	–
Profit factor	–	Expected value / trade	–
Max consecutive losses	–	Recovery factor	–
Calmar ratio	–	Annual return	–