GGTH Predictor — ML Forex Trading System

EA v1.19 · Predictor v9.5 · GUI v2.3 · Sentiment v2.3

A two-process algorithmic trading system for EURUSD on MetaTrader 5. Python runs a five-model deep-learning ensemble and a news-sentiment pipeline. A compiled MQL5 Expert Advisor reads the results from shared JSON files and manages live trades in real time. Neither process blocks the other — Python can be restarted, retrained, or updated without affecting open positions.

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  PYTHON SIDE                                                    │
│                                                                 │
│  unified_predictor_v9.py          main.py                      │
│  (5-model ML ensemble)            (news sentiment pipeline)    │
│                                                                 │
│  Writes every N minutes:          Writes every 10 minutes:     │
│  ea_signal_EURUSD_M5.json         forex_sentiment.json         │
└────────────────┬──────────────────────────┬────────────────────┘
                 │       MQL5\Files          │
┌────────────────▼──────────────────────────▼────────────────────┐
│  METATRADER 5                                                   │
│                                                                 │
│  GGTH_2026_v19.mq5  (Expert Advisor)                          │
│  • Reads signal + sentiment on every tick                       │
│  • Manages entries, averaging, TP/SL, journal                  │
│  • Renders live panel on chart                                  │
└─────────────────────────────────────────────────────────────────┘

Communication is one-way through files. Python writes; the EA reads. If the signal file goes stale, the built-in watchdog suspends new entries and shows [WARN] on the chart panel. The sentiment layer is optional — if forex_sentiment.json is missing or stale, all ML signals pass through unmodified.

---

Features

Machine Learning

• Five-model ensemble: LSTM · GRU · Transformer · TCN · LightGBM
• Three prediction timeframes: 1H / 4H / 1D trained and published simultaneously
• Per-timeframe Kalman filter smoothing on ensemble output
• Adaptive ensemble weights — softmax weights updated each cycle based on each model's recent directional accuracy
• Per-model health tracking — consecutive-failure counter excludes unhealthy models until manually reset
• Persisted ensemble state — weights, prediction history, and health counters survive process restarts
• Beta-distribution adaptive confidence floor — updated from live EA trade outcomes
• HMM regime detection — classifies market into trending / ranging / volatile; regime bias applied to ensemble weights
• Hyperparameter tuning via Keras-Tuner

Sentiment Analysis

• Async news fetch from 6 free RSS feeds (no API keys required)
• Optional paid feeds: NewsAPI · Finnhub · MarketAux
• Three-model scoring ensemble: FinBERT (50%) · VADER (30%) · TextBlob (20%)
• Time-decay weighting — article weight halves every 6 hours; articles older than 48 hours discarded
• Source-quality multipliers — Reuters 1.0 down to generic RSS 0.6
• Keyword-regex currency detection across 8 currencies (USD / EUR / GBP / JPY / CHF / AUD / CAD / NZD)
• Per-pair sentiment derived from base-minus-quote currency scores
• Atomic file writes — EA never reads a partially written JSON

Expert Advisor

• Sentiment veto — blocks trades directly opposing confirmed news sentiment (auto-bypassed in Strategy Tester)
• Three-level averaging down — each level independently configurable for lots and pip trigger
• Dollar-amount profit-protection close — closes entire campaign when combined floating P&L hits target
• Configurable trailing stop with activation-pip threshold
• MA trend filter (EMA-83 default) and RSI filter (14-period)
• Multi-session trading windows — up to three independent time windows per day
• Day-of-week filter — individually enable/disable each day
• Adaptive learning — EA monitors closed-trade outcomes and adjusts signal threshold and lot multiplier within configurable bounds
• FIFO-compliant position management (NFA/CFTC)
• Trade journal CSV — per-trade decision context written to MQL5\Files for post-session analysis
• Stale-prediction watchdog — heartbeat timestamp validated every tick
• Live on-chart panel — shows predictions, regime, adaptive stats, and sentiment in real time

---

Requirements

Python

Package	Version	Notes
Python	3.9 – 3.11	3.12+ not supported — TF 2.15 requires ≤ 3.11
tensorflow	==2.15.0	Pinned — saved models are not cross-version portable
keras	==2.15.0	Must match TensorFlow exactly
protobuf	>=3.19,<4.0	TF 2.15 requirement; pip will break this silently without the pin
lightgbm	>=3.3,<5.0
keras-tuner	>=1.1,<1.5
MetaTrader5	>=5.0.37	Windows only
hmmlearn	>=0.3,<0.4	Regime detection
scipy	>=1.7,<2.0	hmmlearn dependency; explicit pin prevents silent mismatch
numpy	>=1.23.5,<2.0
pandas	>=1.3,<3.0
scikit-learn	>=1.0,<2.0
aiohttp	>=3.9	Sentiment pipeline async HTTP
feedparser	>=6.0	RSS parsing
vaderSentiment	>=3.3	Sentiment model
textblob	>=0.18	Sentiment model
transformers	>=4.40	FinBERT
torch	>=2.2	FinBERT backend

tkinter is included with Python on Windows. On Linux: sudo apt-get install python3-tk

MetaTrader 5

• MetaTrader 5 terminal (any broker)
• MetaEditor (included with MT5, press F4)
• EURUSD symbol available in Market Watch

---

Quick Start

1 — Install Python

Download Python 3.10 or 3.11 from python.org.

⚠ Check "Add Python to PATH" during installation. Without this, all batch files fail.

2 — Run the Setup Wizard

setup_wizard.bat

The wizard scans %APPDATA%\MetaQuotes\Terminal\* for MT5 installations, lets you confirm the correct one, and writes config.json atomically. If your MT5 is in a non-standard location, use the manual entry option.

To find your MT5 Files path manually:
MetaTrader 5 → File → Open Data Folder → navigate to MQL5\Files → copy the address bar.

3 — Compile the EA

1. Press F4 inside MetaTrader 5 to open MetaEditor
2. Open GGTH_2026_v19.mq5
3. Press F7 to compile — zero errors and zero warnings expected
4. The EA appears in MT5 → Navigator → Expert Advisors → GGTH_2026_v19

4 — Launch the GUI

run_ggth_gui.bat

On first run this creates a .venv, installs all dependencies from requirements.txt (5–15 min), starts the sentiment pipeline in a separate console window, and opens the GGTH GUI.

5 — Train Models

In the GUI: select train-multitf, ensure all five model types are ticked (LSTM, GRU, Transformer, TCN, LightGBM), and click Start.

First training run: 20–90 minutes depending on hardware.

When complete, ea_signal_EURUSD_M5.json appears in your MQL5\Files folder.

6 — Attach the EA

1. Open an EURUSD M5 chart in MT5
2. Drag GGTH_2026_v19 from Navigator onto the chart
3. Configure inputs (see EA Input Settings)
4. Common tab → tick Allow algorithmic trading
5. Click OK — a smiley face in the chart corner confirms the EA is active

---

Repository Layout

GGTH/
│
├── setup_wizard.bat           # ★ First-time installer entry point
├── setup_wizard.py            # GUI installer (launched by wizard.bat)
├── run_ggth_gui.bat           # Daily launcher
│
├── ggth_gui.py                # Main GUI — tkinter control panel (v2.3)
├── unified_predictor_v9.py    # ML prediction engine (v9.5)
├── config_manager.py          # config.json read/write with schema validation (v2.2)
├── model_builders.py          # LSTM / GRU / Transformer / TCN builders (v2.2)
├── logger.py                  # Thread-safe rotating file + console logger (v2.2)
│
├── main.py                    # Sentiment pipeline entry point (async loop)
├── forex_sentiment.py         # News → ensemble → per-currency → per-pair aggregation
├── news_fetcher.py            # Async RSS + optional API fetcher (aiohttp)
├── sentiment_models.py        # FinBERT / VADER / TextBlob ensemble
├── sentiment_reader_py.py     # Reads forex_sentiment.json; veto/bias logic
├── sentiment_writer.py        # Atomic JSON writer (temp-file + os.replace)
├── currency_mapper.py         # Keyword regex → currency detection
│
├── config.py                  # Sentiment pipeline settings (edit directly)
├── config.json                # Written by setup wizard — do not hand-edit
├── requirements.txt           # Pinned package versions
│
├── ggth_predictor.log         # Auto-created rotating log (5 MB × 3 backups)
├── .venv/                     # Auto-created virtual environment
│
└── MQL5/ (MetaTrader 5)
    └── Experts/
        └── GGTH_2026_v19.mq5  # Expert Advisor source

Files written to MQL5\Files at runtime:

File	Written by	Read by
ea_signal_EURUSD_M5.json	unified_predictor_v9.py	EA every tick
ea_signal_EURUSD_H1.json	unified_predictor_v9.py	EA every tick
ea_signal_EURUSD_H4.json	unified_predictor_v9.py	EA every tick
forex_sentiment.json	main.py	EA every tick
EURUSD_trade_outcomes.json	EA	unified_predictor_v9.py
EURUSD_trade_journal.csv	EA	User (post-session review)
ensemble_state_EURUSD.json	unified_predictor_v9.py	unified_predictor_v9.py on restart

---

ML Engine

unified_predictor_v9.py — the core prediction engine.

Model Ensemble

Model	Type	Specialization
LSTM	Bidirectional LSTM + Attention	Long-range sequential dependencies
GRU	Bidirectional GRU (2 layers)	Shorter-range sequences; faster training
Transformer	Multi-head self-attention (TransformerBlock)	Non-local patterns; multi-scale correlations
TCN	Temporal Convolutional Network with residual skip	Trend regimes; efficient parallel training
LightGBM	Gradient boosted trees	Non-linear feature interactions; fast inference

Prediction Cycle

1. Fetch OHLCV bars from MT5 via the MetaTrader5 Python package
2. Build feature matrix (price transforms, technical indicators, regime flags)
3. Run inference on all five models independently per timeframe
4. Combine predictions using per-timeframe softmax ensemble weights
5. Apply per-timeframe 1-D Kalman filter to smooth output
6. Write flat ea_signal_*.json files atomically to MQL5\Files
7. Read back EA trade outcomes; update Beta-distribution adaptive confidence floor
8. Update ensemble weights based on directional accuracy; persist state

Ensemble Weight Learning

Weights are updated using a softmax-over-inverse-error scheme. Models with lower recent directional error receive higher weight. The update is identity-aligned — a slot per model exists in every cycle regardless of failures, so a failing model cannot accidentally inherit another model's error metric. Models exceeding 5 consecutive failures are excluded from the cycle until the health file is reset.

Regime Detection

A Gaussian HMM is fitted at the end of training and persisted alongside model weights. Each prediction cycle classifies the current market into one of three regimes (trending / ranging / volatile) and applies a bias to ensemble model weights accordingly. Falls back to a heuristic ATR/volatility classifier if the HMM file is absent.

GUI Modes

Mode	Description
train-multitf	Train all five models across 1H / 4H / 1D in one pass (recommended)
train	Train on the primary timeframe only
predict-mtf-cont	Continuous multi-timeframe prediction at configured interval
predict-mtf-once	Single prediction cycle then exit
backtest	Generate historical signals over a date range
safe-backtest	Backtest using only training-time scalers (no data leakage)
tune	Keras-Tuner hyperparameter search (GPU recommended)
report	Print metrics summary without running a cycle

---

Sentiment Pipeline

main.py + forex_sentiment.py — runs as a separate background process.

Cycle (every 10 minutes)

RSS feeds + optional APIs
        │
        ▼
  news_fetcher.py  ──────────────────────────────────► up to ~300 articles/cycle
        │
        ▼
  currency_mapper.py  ──────────────────────────────► which currencies does each article mention?
        │
        ▼
  sentiment_models.py  (Ensemble.score)
    ├── FinBERT   50%  (finance-domain BERT, ~400 MB download on first run)
    ├── VADER     30%  (rule-based lexicon, no download)
    └── TextBlob  20%  (general NLP, no download)
        │
        ▼
  forex_sentiment.py  (aggregate + derive_pairs)
    • time-decay weighting  (half-life 6 h, max-age 48 h)
    • source-quality multipliers
    • per-currency weighted average
    • per-pair score = (base_score − quote_score) / 2
        │
        ▼
  sentiment_writer.py  ──────────────────────────────► forex_sentiment.json  (atomic write)

Supported News Sources

Source	Type	Quality weight
FXStreet	Free RSS	0.85
ForexLive	Free RSS	0.85
MarketWatch	Free RSS	0.85
DailyFX	Free RSS	0.80
Investing.com	Free RSS	0.75
Finnhub	API (free tier)	0.80
NewsAPI	API (free tier)	0.70
Reuters (via NewsAPI)	API	1.00

No API keys are required for basic operation.

Sentiment Modes (EA-side)

Mode	Behaviour
veto	Block trades when |score| > InpSentimentVetoBand and signal opposes sentiment
bias	Blend sentiment into signal strength (configured in sentiment_reader_py.py)
off	Display sentiment panel only; no trades blocked

Fail-open design: missing file, stale snapshot (> InpSentimentMaxAgeSec), or low confidence (< InpSentimentMinConf) all result in all ML signals passing through unmodified.

---

Expert Advisor

GGTH_2026_v19.mq5 — compiled and attached to an EURUSD M5 chart.

Trade Lifecycle

OnTick()
  │
  ├── ReadSentimentFile()         parse forex_sentiment.json (mtime-guarded)
  ├── ReadSignalFile()            parse ea_signal_EURUSD_M5.json
  ├── Watchdog check              stale? → suspend new entries, show [WARN]
  │
  ├── IsTradingAllowed()
  │     ├── Day-of-week filter
  │     ├── Session window filter  (up to 3 independent windows)
  │     ├── Spread limit
  │     └── Margin utilization limit
  │
  ├── Signal gates (applied in order)
  │     ├── InpMinPredictionPips  (minimum predicted move)
  │     ├── MA trend filter       (EMA-83 default)
  │     ├── RSI filter            (14-period, 75/35 thresholds)
  │     └── Sentiment veto        (auto-bypassed in Strategy Tester)
  │
  ├── Open initial position       (if no campaign open and signal passes all gates)
  │
  ├── Manage open campaign
  │     ├── Averaging down        (Level 1 @ 15 pip, Level 2 @ 35 pip, Level 3 disabled)
  │     ├── Profit-protection close  (all legs when combined P&L ≥ InpProfitTargetAmount)
  │     ├── Trailing stop         (if enabled, engages after InpTrailingActivationPips)
  │     ├── Stop-loss / TP        (shared SL from blended average entry)
  │     └── Max hold time         (force-close if campaign age > InpMaxHoldHours)
  │
  ├── Adaptive learning           (every InpAdaptEveryN closed campaigns)
  │     ├── Compute rolling win rate and profit factor over last N campaigns
  │     ├── Blend-adjust InpMinPredictionPips and lot multiplier
  │     └── Write outcome to ea_signal file for Python to consume
  │
  └── DrawPanel()                 update on-chart display

FIFO Compliance

InpFIFOCompliant = true enforces NFA/CFTC first-in-first-out position closing. Required for US-regulated brokers. Disable only for non-US brokers where hedging is permitted.

Strategy Tester

Set InpStrategyTesterMode = true when backtesting. The EA reads predictions from pre-exported CSV files rather than live JSON, and the sentiment veto is automatically disabled (historical sentiment data is not available in the tester).

---

Configuration Reference

config.json

Written by the setup wizard. Contains exactly two keys. Do not add others.

{
  "mt5_files_path": "C:\\Users\\YourName\\AppData\\Roaming\\MetaQuotes\\Terminal\\<hash>\\MQL5\\Files",
  "version": "2.3"
}

Key	Description
mt5_files_path	Full path to MQL5\Files for the terminal your EA is attached to
version	Schema version gate — must be "2.3". Do not edit.

To update the path: re-run setup_wizard.bat, or edit the MT5 Files field in the GUI.

config.py (Sentiment Pipeline)

Edit directly in a text editor. Restart the sentiment pipeline window for changes to take effect.

mt5_files_path          = r"C:\Users\...\MQL5\Files"  # must match config.json
output_filename         = "forex_sentiment.json"
update_interval_seconds = 600          # how often the pipeline re-runs (10 min)
half_life_hours         = 6.0          # article weight halves every 6 hours
max_age_hours           = 48.0         # discard articles older than this
newsapi_key             = ""           # optional — newsapi.org free tier (100/day)
finnhub_key             = ""           # optional — finnhub.io free tier (60/min)
marketaux_key           = ""           # optional — marketaux.com free tier (100/day)

ensemble_weights = {
    "finbert":  0.50,
    "vader":    0.30,
    "textblob": 0.20,
}

---

EA Input Settings

Full reference for all inputs in the MT5 EA dialog. Grouped as they appear in the Inputs tab.

Testing Mode

Parameter	Default	Description
InpStrategyTesterMode	false	Read predictions from CSV files instead of live JSON. Required for Strategy Tester backtests.

Trading Setup

Parameter	Default	Description
InpSymbol	EURUSD	Trading symbol. Must match Market Watch name exactly.
InpTradingTimeframe	PERIOD_H1	Prediction timeframe the EA selects from the multi-timeframe signal file.
InpEnableTrading	true	Master switch. false = monitor only, no trades opened.
InpLotMode	LOT_MODE_FIXED	FIXED: use InpFixedLot. RISK: calculate lots from InpRiskPercent.
InpFixedLot	0.10	Base lot size (FIXED mode). Adaptive multiplier scales this up or down.
InpRiskPercent	1.0	Percent of balance risked per trade (RISK mode).
InpMagic	20260522	Magic number stamped on all EA orders.
InpFIFOCompliant	true	Enforce FIFO closing order. Required for NFA/US brokers.

Institutional Filters

Parameter	Default	Description
InpMaxMarginUsagePct	60.0	Block new entries when used margin exceeds this % of equity.

Strategy & Signal

Parameter	Default	Description
InpMinPredictionPips	17.0	Minimum predicted price move (pips) to open a trade. Adapted dynamically by the adaptive learning system.

Averaging Down

Parameter	Default	Description
InpUseAveragingDown	true	Enable multi-level averaging.
InpAvgLevel1Lots	0.20	Lot size for averaging entry Level 1.
InpAvgLevel1Pips	15	Adverse pips from initial entry before Level 1 fires.
InpAvgLevel2Lots	0.20	Lot size for Level 2.
InpAvgLevel2Pips	35	Adverse pips before Level 2 fires.
InpAvgLevel3Lots	0.30	Lot size for Level 3.
InpAvgLevel3Pips	10000	Set to 10000 to disable Level 3 (default).

⚠ All averaging levels share one stop-loss from the blended average entry. Large Level 3 lots can expose significantly more capital than the initial trade.

Profit Protection

Parameter	Default	Description
InpUseProfitProtection	true	Close all campaign legs when combined floating P&L hits target.
InpMinPositionsForProtection	2	Minimum open legs before the profit target is checked.
InpProfitTargetAmount	27.00	Combined floating profit (USD) to trigger full-campaign close.

Max Hold Time

Parameter	Default	Description
InpUseMaxHoldTime	true	Enable time-based force-close.
InpMaxHoldHours	10	Force-close any campaign open longer than this many hours.

Market Context Veto

Parameter	Default	Description
InpUseMarketContextVeto	false	Block entries during ATR spikes or abnormal candle moves.
InpVolatilitySpikeMultiplier	2.5	Block when current ATR exceeds N× the recent average.
InpMaxCandleChangePercent	0.5	Block when current candle move exceeds this % of price.
InpVolatilityLookback	20	Bars used to compute average ATR for comparison.

Take Profit & Stop Loss

Parameter	Default	Description
InpUsePredictedPrice	true	Use ML predicted target price as TP (recommended).
InpStopLossPips	60	Stop-loss distance in pips from blended average entry.
InpTakeProfitPips	200	Fixed TP in pips when InpUsePredictedPrice = false.
InpTPMultiplier	1.0	Multiplier applied to predicted TP distance.
InpMinTPPips	17	Minimum TP distance (clamped up if prediction is smaller).
InpMaxTPPips	70	Maximum TP distance (clamped down if prediction is larger).

Trend Filter

Parameter	Default	Description
InpUseTrendFilter	true	BUY only above MA; SELL only below MA.
InpTrendMAPeriod	83	MA period.
InpTrendMAMethod	MODE_EMA	MA type: EMA / SMA / SMMA / LWMA.
InpTrendMAPrice	PRICE_CLOSE	Applied price for MA calculation.

RSI Filter

Parameter	Default	Description
InpUseRSIFilter	true	Block overbought BUY and oversold SELL signals.
InpRSIPeriod	14	RSI calculation period.
InpRSIOverbought	75.0	RSI level above which BUY signals are blocked.
InpRSIOversold	35.0	RSI level below which SELL signals are blocked.

Trailing Stop

Parameter	Default	Description
InpUseTrailingStop	false	Enable trailing stop.
InpTrailingStopPips	12	Distance behind price the stop follows (pips).
InpTrailingStepPips	5	Minimum price movement before stop is moved.
InpTrailingActivationPips	25	Minimum profit in pips before trailing engages. Set 0 to engage immediately.

Trading Days

Parameter	Default
InpTradeMonday	true
InpTradeTuesday	true
InpTradeWednesday	true
InpTradeThursday	true
InpTradeFriday	true
InpTradeSaturday	false
InpTradeSunday	false

Trading Sessions (up to 3 windows)

Parameter	Default	Description
InpUseSession1	true	Enable Session 1.
InpSession1StartHour	0	Session 1 start (server time, 24h).
InpSession1EndHour	17	Session 1 end. Default covers London + NY overlap.
InpUseSession2	false	Enable Session 2 (e.g. Asian session).
InpUseSession3	false	Enable Session 3.

All times are broker server time as shown in MT5's bottom-right clock.

Display Settings

Parameter	Default	Description
InpFontSize	14	Panel font size in points.
InpLineSpacing	12	Extra vertical spacing (px) between rows.
InpPanelPadding	10	Inner padding (px) between panel border and content.
InpTextColor	clrWhite	Standard text color.
InpUpColor	clrLimeGreen	Color for UP predictions and bullish labels.
InpDownColor	clrRed	Color for DOWN predictions and bearish labels.
InpXOffset	20	Horizontal pixel offset from left edge.
InpYOffset	30	Vertical pixel offset from top.
InpShowDebug	true	Show diagnostic rows (file age, veto reasons). Disable for cleaner production display.

Adaptive Learning

Parameter	Default	Description
InpEnableAdaptiveLearning	true	Enable online adaptive parameter adjustment.
InpAdaptLookback	20	Campaigns in the rolling evaluation window.
InpAdaptEveryN	5	Trigger an adaptation step every N closed campaigns.
InpAdaptRate	0.15	Blend rate for threshold/multiplier updates (0.05=slow, 0.30=fast).
InpAdaptMinPredFloor	1.0	Hard lower bound on adapted signal threshold (pips).
InpAdaptMinPredCeil	60.0	Hard upper bound on adapted signal threshold (pips).
InpAdaptLotMultFloor	0.25	Minimum adaptive lot multiplier.
InpAdaptLotMultCeil	1.50	Maximum adaptive lot multiplier.
InpAdaptResetOnInit	false	Reset all learned params on EA init. Use after changing base lot size or SL.
InpShowAdaptiveDebug	true	Show adaptive stats row in panel.

Stale Prediction Watchdog

Parameter	Default	Description
InpEnableStaleWatchdog	true	Suspend new entries if signal file is stale.
InpStalePredictionMaxMinutes	90	Maximum signal file age before watchdog fires.
InpFailClosedOnMissingHeartbeat	false	Block trading if no heartbeat received since EA init.
InpShowWatchdogStatusOnChart	true	Show [OK] / [WARN] banner on chart.

Trade Journal

Parameter	Default	Description
InpEnableTradeJournal	true	Write per-trade decision context to CSV.
InpJournalAveragingTrades	true	Also journal averaging-down legs.
InpShowJournalDebug	false	Print journal write events to MT5 Experts log.

Sentiment Filter

Parameter	Default	Description
InpUseSentiment	true	Enable sentiment veto (auto-disabled in Strategy Tester).
InpSentimentFile	forex_sentiment.json	Sentiment JSON filename in MT5 Common Files.
InpSentimentMinConf	0.20	Minimum confidence before sentiment is acted on.
InpSentimentMaxAgeSec	1800	Maximum snapshot age (seconds) before it is ignored (30 min).
InpSentimentVetoBand	0.10	Minimum |score| required to veto. Scores in (−0.10, +0.10) never block.
InpShowSentimentPanel	true	Show sentiment section in on-chart panel.

---

GUI Reference

ggth_gui.py is a tkinter control panel that builds and runs the predictor CLI command with the selected options.

Key Controls

Control	Description
Symbol	Trading symbol to train/predict (default EURUSD)
Action	Predictor mode (see GUI Modes)
Force Retrain	Overwrite existing saved models
Interval (min)	Prediction cycle interval for continuous modes
Models	Toggle each of the five model types independently
Kalman Filter	Apply per-timeframe Kalman smoothing to predictions
Python exe	Path to Python interpreter (defaults to active venv)
Predictor	Path to unified_predictor_v9.py
MT5 Files	Path to MQL5\Files — saved to config.json on click of Save MT5 Path
Train Start / End	Optional date window for training data
Predict Start / End	Optional date window for prediction/backtest
Start	Build CLI command and launch subprocess
Stop	Graceful shutdown: CTRL_BREAK_EVENT → 5s wait → terminate

Output streams to the scrollable log area in real time. Save Log writes the current log area content to a timestamped .txt file.

---

On-Chart Panel

GGTH PREDICTOR v1.19  │  EURUSD
Price: 1.08432    Regime: trending
────────────────────────────────────────────────────────
[OK] Watchdog  age=2m / limit=90m
────────────────────────────────────────────────────────
PREDICTIONS
1H  UP    1.08500  (+0.06%)     Acc: 45/72  (62.5%)
4H  DOWN  1.08100  (-0.31%)     Acc: 12/20  (60.0%)
1D  UP    1.09200  (+0.71%)     Acc:  5/8   (62.5%)
────────────────────────────────────────────────────────
ADAPTIVE LEARNING
WR: 58.3%  PF: 1.42  Kelly: 16.2%  LotMult: 0.95x
AdaptPred: 18.4 pips  #adapt: 5  Trades: 47
────────────────────────────────────────────────────────
NEWS SENTIMENT
BULLISH  score: +0.241  conf: 68%
Base: +0.312  Quote: -0.071  Age: 4m
SELL signals may be vetoed

Row	Description
Header	EA version, symbol, current price, detected regime
Watchdog	[OK] or [WARN] with current file age vs configured maximum
Predictions	Per-timeframe: direction, predicted target price, % move, running accuracy
Adaptive Learning	Win rate, profit factor, Kelly %, lot multiplier, adapted pip threshold, trade count
News Sentiment	Direction label, score (−1 to +1), confidence %, base/quote scores, snapshot age, active veto message

---

Troubleshooting

Panel shows [STALE] or [WARN]

• Python predictor has stopped writing fresh signal files
• Restart via run_ggth_gui.bat
• Check ggth_predictor.log for import errors or MT5 connection failures
• Verify MT5 is open and logged in — the predictor requires MT5 to fetch price data

Sentiment section shows "No data"

• forex_sentiment.json has not been written yet or is in the wrong folder
• Confirm main.py is running in the separate console opened by run_ggth_gui.bat
• On first run, FinBERT downloads ~400 MB — wait for completion
• Confirm output_filename in config.py matches InpSentimentFile in EA settings

EA opens no trades despite an active signal

• Check InpEnableTrading = true
• Predicted pip move may be below InpMinPredictionPips
• Confirm current server time is within an enabled session
• Check InpMaxMarginUsagePct — margin may be over the limit
• Check the sentiment panel for an active veto message
• Verify the Common tab → Allow algorithmic trading is ticked

Models fail to load after a package upgrade

• Keras 2.15 saved models cannot be loaded with a different Keras major version
• Restore: .venv\Scripts\activate then pip install -r requirements.txt
• If you intentionally upgraded, delete saved models and retrain via train-multitf

config.json unknown-key warning in log

• An older setup wizard wrote keys that v2.3 no longer reads — harmless
• To clean up: delete config.json and re-run setup_wizard.bat

MT5 connection errors during training

• Ensure MT5 is open and logged in
• The MetaTrader5 Python package requires 64-bit Python matching your 64-bit MT5 terminal
• MT5: Tools → Options → Expert Advisors → Allow algorithmic trading

---

License

Copyright 2026 Jason Rusk. All rights reserved.

---

Author

Jason Rusk — jason.w.rusk@gmail.com