Skip to main content

​
Hugging Face Hub

Required. All ML model binaries (.pkl and .onnx) are stored on Hugging Face Hub. The trading bot pulls the latest models at startup.

​
Setup

  1. Create a free account at huggingface.co
  2. Create a new model repository (private)
  3. Generate an access token with write permission at Settings → Access Tokens
  4. Add to your .env:
HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxx
HF_REPO_ID=yourname/novosky-models

​
Usage

# Pull latest models (run at startup / before trading)
python ml/hf_hub.py --pull

# Push after retrain
python ml/hf_hub.py --push

# Push with a tag (e.g. after a major retrain)
python ml/hf_hub.py --push --tag phase-16

# List available revisions
python ml/hf_hub.py --list
The push command uploads all .pkl and .onnx files from models/ plus the models/model_compat.json and models/risk_metadata.json metadata files.

​
What is stored

FileDescription
rf_signal.pklRandom Forest signal model (raw)
rf_signal_calibrated.pklCalibrated RF signal model
rf_signal.onnxONNX export of RF signal model
xgb_signal.pkl, _calibrated.pkl, .onnxXGBoost equivalents
lgb_signal.pkl, _calibrated.pkl, .onnxLightGBM signal equivalents
position_*.pklPosition model files
sltp_*.pklSL/TP model files
risk_*.pklRisk multiplier model files
model_compat.jsonFeature list hash + model paths — compatibility check
risk_metadata.jsonRisk model metadata: trained_at, val_mae, features

​
Supabase

Optional. Supabase stores trade logs, open position snapshots, and bot status. The website dashboard reads from Supabase for trade history display.

​
Setup

  1. Create a free project at supabase.com
  2. Run the schema migrations from supabase/migrations/
  3. Add to your .env:
SUPABASE_URL=https://xxxxxxxxxxxx.supabase.co
SUPABASE_KEY=your-service-role-key

​
Schema

-- Core tables (from supabase/migrations/)
trades          -- closed trade records with broker deal data
positions       -- open position snapshots
account_snapshots  -- periodic balance/equity snapshots
bot_status      -- heartbeat and current state

​
Data integrity rules

For cent accounts, always store raw MT5 values — do not convert USC to USD before writing.
  • Store closing_deal.profit as-is (e.g. 230.0 USC, not 2.30 USD)
  • For price-based profit estimates: use _usd_to_raw(price_diff × lot × contract_size) which multiplies by 100 for cent accounts
  • entry_at timestamps: always strip microseconds → datetime.now(UTC).replace(microsecond=0).isoformat()

​
Trade reconciliation

After every sync_fallback or estimation path, _reconcile_trade_bg(ticket, ...) fires in the background. It retries _get_deals_for_position() every 60 seconds for up to 4 hours, then overwrites Supabase with real broker deal data. VT Markets can delay closing deal history by hours for manual (mobile) closes. The reconciliation loop handles this automatically.

​
Telegram

Optional but strongly recommended. Telegram sends trade alerts, weekly optimize reports, and error notifications.

​
Setup

  1. Create a bot with @BotFather and copy the token
  2. Get your chat ID (send a message to your bot, then check https://api.telegram.org/bot<token>/getUpdates)
  3. Add to your .env:
TELEGRAM_BOT_TOKEN=7234567890:AABxxxxxxxxxxxxxxxxxxxxxxxx
TELEGRAM_CHAT_ID=123456789

​
Notification types

EventMessage type
Trade openedBUY/SELL signal, lot, entry price, SL/TP, confidence
Trade closedP&L, duration, exit reason (TP/SL/EXIT signal)
Weekly optimize — successNew score, improvement %, key config, pushed to HF
Weekly optimize — rollbackOld score retained, reason
Hard haltEquity drawdown % that triggered halt
MT5 connection errorAPI URL, error code

​
Telegram commands

The bot responds to commands sent from your chat: 
/status   — current position, equity, today's P&L
/stats    — last 20 trades summary
/pause    — pause trading (next cycle will skip)
/resume   — resume trading
/halt     — trigger immediate hard halt
Commands are handled by trading/telegram_commands.py.

​
Service dependency summary

ServiceRequiredWithout it
MT5 REST APIYesBot cannot trade
Hugging Face HubYesCannot load models
SupabaseNoNo trade log, no website dashboard
TelegramNoNo alerts or reports