Troubleshooting

🦉 AthenaAI is an AI-powered ensemble machine-learning trading bot for PocketOption, built on BinaryOptionsToolsV2 → Get it on GitLab
💬 Need help? Join the Chipa Discord · ⚡ Go no-code with ChipaEditor

Problems grouped by lifecycle stage. Each row quotes the real log/error text from the code.

Startup and installation

Symptom	Cause	Fix
`ERROR: PO_SSID is not set.` (exit code 1)	No session ID in the environment or `.env`	`export PO_SSID='...'` (bash) / `$env:PO_SSID='...'` (PowerShell), or copy `.env.example` to `.env` and fill it in. Run from the repo root — only `./.env` is read. See PO_SSID.
`ERROR: BinaryOptionsToolsV2 not installed.` then immediate exit	The broker client isn’t on PyPI; importing `src/bot.py` hard-exits without it	Install the wheel for your platform from https://gitlab.chipatrade.com/chipadevorg/BinaryOptionsTools-v2/-/releases. See Installation. Note `train.py` doesn’t need it.
`WARNING: scikit-learn not found. Install with: pip install scikit-learn`	sklearn missing — the ensemble falls back to a non-learning stub, so the bot “runs” but never really learns	`pip install scikit-learn` (or `pip install -r requirements.txt`). Treat as mandatory.
`Warning: could not parse '...', using (120, 300, 600)`	Malformed `PO_EXPIRY_OPTIONS`	Use comma-separated integers in seconds, e.g. `120,300,600`.
Hangs at `Connecting to PocketOption …` or balance is wrong	Expired/invalid SSID	Log in to PocketOption again and refresh `PO_SSID`. Session IDs expire.

Brain and model state

Symptom	Cause	Fix
`Failed to load brain: ...` at startup	Corrupt/incompatible brain pickle (e.g. saved with a different sklearn version)	The bot automatically falls back to journal reload + dataset pre-training. To rebuild deliberately: retrain with `python train.py`, or start once with `PO_RETRAIN=1`.
`Feature dimension changed (N → M) — resetting.`	The feature engine’s vector size changed (code update, changed experimental features) while old state expected N features	Expected after feature changes: the ensemble resets itself and relearns. For a clean slate, retrain the brain; note journal reload also skips old trades whose feature dimension doesn’t match (Persistence).
`🔄 Force retrain requested — ignoring saved brain.` on every start	`PO_RETRAIN=1` left set — it deletes the brain each startup	Unset `PO_RETRAIN` (or set `0`) once the retrain is done.
`No trained models to save.`	Saving was attempted before anything was fitted	Harmless; train first (dataset or `train.py`).
MLP silently contributes nothing on the trading machine	PyTorch not installed there — the MLP degrades to a neutral 0.5 vote	Install torch on the trading machine, or accept sklearn-only predictions. See GPU training.

Training and datasets

Symptom	Cause	Fix
`Dataset too small (N candles).` (trainer aborts) / `Dataset too small (N candles), skipping pre-training.` (bot)	Fewer than 50 candles parsed — wrong path or unrecognized format; bad rows are skipped silently	Check the file path, then compare the file against the data format spec. The log’s `Parsed N candles from file.` versus your row count reveals silent skipping.
`Only N valid samples — need at least 200.` (trainer) / `Only N valid samples from dataset, skipping.` (bot, <20)	Data parsed but too short after the ≥26-candle feature warm-up and label horizon; flat closes are also skipped	Use a longer history — tens of thousands of candles minimum.
`Holdout too small (N) — skipping walk-forward test.`	Fewer than 100 test samples	More data or a larger `--test-split`.
Walk-forward WR below breakeven	No edge found on this data	Don’t deploy that brain. More history, different `--expiry`, or different asset. See Training.
`PyTorch not installed — skipping neural model`	Informational	`pip install torch`, or pass `--no-nn` intentionally.
`--check-gpu` reports CPU despite an NVIDIA GPU	CPU-only torch wheel installed	Reinstall from the CUDA index: `pip install torch --index-url https://download.pytorch.org/whl/cu128`.
CUDA out of memory during MLP training	Batch/model too large for VRAM	Lower `--batch-size` (e.g. `512`) or shrink `--hidden`.
`Not enough data for batch models (N)`	Fewer than 100 buffered samples when GBM/RF training was attempted	Provide more data; the online models still work meanwhile.

No trades happening

First: a bot that mostly logs ⏸ Low confidence is working correctly — it trades only when the model has an edge. Walk the ⏸ diagnostic (printed at most every 30 s) through this table; the full explanations are in Live trading.

`⏸` reason	Why	What unblocks it
`Warming up (N/60 candles)`	Candle buffer below `max(warmup_candles, 60)`	Time — one candle per timeframe interval; check the stream subscribed (`📡 Using subscribe_symbol_time_aligned …`)
`Max trades open (1/1)`	A pending trade	Its resolution (expiry + ~2 s)
`Wait between trades (Ns left)`	60 s pacing floor	Time, or lower `min_wait_between_trades` in `src/config.py`
`Daily loss limit reached`	Daily P&L hit −`PO_MAX_DAILY_LOSS`	UTC-midnight reset (or, reluctantly, a restart — the counter is in-memory)
`Cooldown (Ns left)`	3 consecutive losses → 300 s pause	Time
`Outside trading hours (HH UTC)`	UTC hour not whitelisted in `trading_hours` (default excludes 05–15 and 18 UTC)	Wait, or edit `trading_hours` in `src/config.py`
`Adaptive cooldown (Ns)`	Last-10 win rate < 40%	Time / better results
`Feature compute returned None`	Not enough candles for indicators	Fills with the buffer; persistent → check candle stream health
`Volatile regime — skipping`	`skip_volatile_regime=True` (not default)	Calmer market or disable the flag
`Low confidence: X% (need 63.0%) …`	Prediction below `PO_MIN_CONF`	Normal. Persistent for days → consider retraining, or cautiously lower `PO_MIN_CONF`
`Overconfident: X% (cap=85.0%) — likely overfit noise`	Above `PO_MAX_CONF`	Usually right to reject; frequent occurrences suggest overfit batch models — retrain
`Low range candle (…) — skip`	Doji/indecision candle (range < 0.00005)	Market movement
`Indicators misaligned (conf=X%)`	<2 of 3 indicators agree	Set `PO_REQUIRE_ALIGNMENT=0` to disable (more, lower-quality trades)
`Signal confirmation (n/N) …`	Waiting for consecutive agreeing candles	Only if `signal_confirmations` > 1
`5-min trend DOWN vs CALL — skip` / `UP vs PUT`	Higher-timeframe trend disagrees	Market alignment
`🧠 Adaptive skip: <reason>`	Adaptive layer blocked regime/hour/direction or raised the threshold	Self-clears as its stats evolve; full reset = delete learned state (Persistence)

Also check the candle stream: Candle stream error: ... stops the whole bot (_running = False) — restart it and check connectivity/SSID.

Trade results

Symptom	Cause	Fix
A trade stays pending long after expiry	`check_win` returns something unparseable; the checker first tries `win`/`loss`/`draw` (dict `result`/`status` keys or plain string), then substring-matches `win` / `loss`/`lose` / `draw` in the raw response, else retries every 3 s	Usually resolves on a later poll. Enable debug logging to see `check_win(id) raw=… parsed=…`.
`Abandoning stale trade <id> after timeout: ...`	`check_win` kept erroring until the trade was older than `expiry × 5`	The trade is dropped unresolved — not journaled as win/loss and not learned from. Occasional occurrences are tolerable; frequent ones point at connection/SSID problems or a broker API change.
Logged profit doesn’t match your account	The bot books results at a hardcoded 0.85 payout	Cosmetic/accounting only; the broker’s actual payout applies to your balance.

Risk note

If the bot stops trading because of the daily loss limit or cooldowns, that is the risk management working as designed — bypassing it by restarting or loosening limits increases the money you can lose. See Risk management.