Live trading

🦉 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

This guide covers operating the bot (python main.py) once it is installed and configured.

Startup paths

At startup the bot decides where its model state comes from:

Saved brain exists at PO_BRAIN_PATH (default athena_brain.pkl): it loads it — 🧠 Brain loaded (fitted=True, batch=True, models: {...}), plus ⏱ Expiry stats loaded: ... if <brain>_expiry.pkl exists — then ✅ Loaded saved brain — skipping dataset training!. Journal reload and dataset pre-training are both skipped.
No brain (or load failed): the bot first replays trade_journal.db (🔄 Reloaded N trades from journal — models retrained!), then pre-trains from PO_DATASET if set — a several-minute in-process training run ending with a walk-forward report and ✅ Pre-trained on N samples!.
PO_RETRAIN=1: logs 🔄 Force retrain requested — ignoring saved brain., deletes the brain file, and takes the journal + dataset path above. Use this after changing feature-affecting code or when you suspect a stale brain. Unset it again afterwards — it deletes the brain on every startup. See PO_RETRAIN.

Tip: prefer training offline with python train.py (Training) so bot startups are instant.

Reading the log output

Diagnostic `⏸` lines (every 30 seconds while idle)

The trade loop evaluates gates every 2 seconds but only logs the blocking reason at most every 30 seconds. Each reason and what to do about it:

Log line	Meaning	Action
`⏸ Warming up (N/60 candles)`	Not enough candles buffered yet	Wait; ~1 candle per timeframe interval
`⏸ Max trades open (1/1)`	A trade is pending resolution	Normal; resolves after its expiry
`⏸ Wait between trades (Ns left)`	Inside `min_wait_between_trades` (60 s default)	Normal pacing
`⏸ Daily loss limit reached`	Daily P&L ≤ −`PO_MAX_DAILY_LOSS`	Trading stops until the UTC day rolls over
`⏸ Cooldown (Ns left)`	Consecutive-loss cooldown (300 s after 3 losses)	Wait; protects against tilt
`⏸ Outside trading hours (HH UTC)`	Current UTC hour not in `trading_hours`	Expected outside 00–04 / 16–17 / 19–23 UTC by default
`⏸ Adaptive cooldown (Ns)`	Recent win rate < 40% triggered an extra 120 s / 300 s wait	The bot is cold; it self-throttles
`⏸ Feature compute returned None`	Too few candles for indicators	Should clear as the buffer fills
`⏸ Volatile regime — skipping`	`skip_volatile_regime` enabled and regime is volatile	Off by default
`⏸ Low confidence: X% (need 63.0%) …`	Prediction below `PO_MIN_CONF`	The most common line — the model has no strong opinion
`⏸ Overconfident: X% (cap=85.0%) — likely overfit noise`	Above `PO_MAX_CONF`	Overfit-looking batch-model signal rejected
`⏸ Low range candle (…) — skip`	Candle range < 0.00005 (doji/indecision)	Normal in flat markets
`⏸ Indicators misaligned (conf=X%)`	Fewer than 2 of 3 (RSI, MACD hist, SMA cross) agree	Requires `PO_REQUIRE_ALIGNMENT=1`
`⏸ Signal confirmation (n/N) …`	Waiting for N consecutive candles to agree	Only with `signal_confirmations` > 1
`⏸ 5-min trend DOWN vs CALL — skip` (or UP vs PUT)	Higher-timeframe trend contradicts the signal	Multi-timeframe filter
`🧠 Adaptive skip: <reason> (conf=X%)`	Adaptive strategy blocked the regime/hour/direction or raised the threshold	See Adaptive strategy

A quiet bot cycling through ⏸ Low confidence lines is healthy — see Troubleshooting if nothing trades for many hours.

`▶ TRADE` lines

▶ TRADE  CALL  $25.00  conf=68.4%  expiry=300s  regime=trending_up  [120s:55%(20) | 300s:62%(41) | 600s:48%(12)]

Direction, stake ($), model confidence, the AI-chosen expiry, detected regime, and the expiry selector’s per-expiry win-rate stats (win% and sample count).

Result lines

✅  WIN  $+21.25  expiry=300s  |  W:14 L:9 WR:60.9% PnL:$87.50  |  expiry stats: ...
❌  LOSS  $-25.00  expiry=300s  |  ...
➖  DRAW  $+0.00  expiry=300s  |  ...

Profit assumes an 85% payout (win = stake × 0.85, loss = -stake). The middle section is the running performance summary; the end repeats expiry stats after this result was recorded.

Online learning while trading

Every resolved win/loss becomes a training sample:

Label derivation: win → the predicted direction was correct; loss → the opposite direction is used as the label.
Incremental retrain: the online models (SGD/PA/NB) run partial_fit after every retrain_every = 10 new results, logging Models updated | acc EMAs: {...}.
Batch retrain: GBM/RF (and MLP if present) retrain once at least 200 samples are buffered and then every 100 results, always on a sliding window of the most recent 500 samples — 🔄 Batch models retrained on recent N samples. This stops stale batch models from poisoning the ensemble.
Brain autosave: every 10 completed trades the bot writes a journal snapshot and saves the brain (🧠 Brain saved to ...), so a crash loses at most ~10 trades of learning. A clean Ctrl+C shutdown also saves.

How results feed the adaptive systems

Each result also updates:

AdaptiveStrategy (src/trading/strategy.py) — records direction/regime/confidence/UTC-hour/result; every 25 trades it reviews itself and may block losing regimes or hours, prefer a direction, adjust the confidence threshold ±3%, and set an extra cooldown (120 s if the recent-10 win rate < 40%, 300 s if < 30%). Watch for its 🧠 Review complete. ... log block.
ExpirySelector — per-expiry win/loss counts steer future expiry choices and are persisted to <brain>_expiry.pkl.
MoneyManager — profit updates the daily P&L used by the loss limit (Risk management).
FeatureLab — tracks whether experimental features help (Feature lab).

Day-to-day checklist

Session IDs expire — if the balance call hangs or trades fail, refresh PO_SSID.
Check the startup banner matches your intended config (confidence band, hours, expiries).
Let the daily loss limit and cooldowns do their job — don’t restart the bot to bypass them; the loss limit resets automatically at UTC midnight, but the daily P&L counter lives only in memory, so a restart does reset it.
Back up athena_brain.pkl, athena_brain_expiry.pkl, and trade_journal.db occasionally (Persistence).

Risk note

The bot trades real money autonomously within your configured limits. Binary options can lose the full stake on every trade; past performance and walk-forward results do not guarantee future profits. Keep PO_MAX_STAKE and PO_MAX_DAILY_LOSS at levels you can afford to lose.