src.training.trainer

🦉 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

Source: src/training/trainer.py. The standalone trainer: trains the full ensemble (online SGD/PA/NB + batch GBM/RF + optional GPU MLP) on a historical CSV and saves a brain file the live bot loads at startup. Run via train.py (see the CLI reference); the package re-exports Trainer, TrainerConfig, and detect_device from src/training/__init__.py.

@dataclass TrainerConfig

class Trainer

__init__(self, cfg: TrainerConfig)

Creates a fresh EnsemblePredictor and FeatureEngine, and resolves the training device via detect_device(force_cpu=cfg.force_cpu).

run(self) -> dict

The full pipeline. Returns the walk-forward metrics dict (see _walk_forward). Steps:

1. Banner — logs dataset/brain paths and the device line (⚡ Training device: … or 🖥 …).
2. Load candles — load_candles(cfg.dataset_path). Fewer than 50 candles → raise SystemExit("Dataset too small (N candles)."). More than max_candles → keep the most recent.
3. Build samples — label horizon = max(1, expiry // timeframe) candles; build_samples(...) produces (X, y) where y=1 means the close rose over the horizon (CALL wins). Fewer than 200 samples → SystemExit.
4. Chronological split — no shuffle (walk-forward): first 1 − test_split for training, the rest held out.
5. Online models — feed training samples to the ensemble in online_batch-sized chunks, calling ensemble.partial_fit() after each chunk (this also fits the scaler).
6. Batch models — ensemble.train_batch_models() (GBM + RF; RF already uses all CPU cores).
7. Neural model — _train_nn(...) unless use_nn is False.
8. Walk-forward evaluation — _walk_forward(test_X, test_y).
9. Save — creates the output directory (os.makedirs(..., exist_ok=True)), then ensemble.save_brain(cfg.brain_path) and logs ✅ Training complete — brain saved to <path>.

Note the trainer does not write the _expiry.pkl companion file — expiry stats are learned live only.

_train_nn(self, train_X, train_y, test_X, test_y) (internal)

Imports TorchMLPClassifier and torch; if PyTorch is missing, logs PyTorch not installed — skipping neural model (pip install torch for GPU training). and returns — training continues without the MLP.

Otherwise: builds the MLP with cfg.nn_hidden / nn_epochs / nn_batch_size / nn_lr on the detected device, fits it on scaler-transformed training data, and measures holdout accuracy (0.5 if the holdout is empty). The holdout accuracy becomes the model’s vote weight (accuracy_ema) in the ensemble. It then removes any existing "MLP" entry from ensemble._batch_models, appends {"name": "MLP", "clf": mlp, "accuracy_ema": acc}, and sets _batch_fitted = True. Because EnsemblePredictor.load_brain() rebuilds _batch_models from the saved list, the MLP survives the round-trip into the live bot, where it votes with weight acc × 0.75 like the other batch models.

_walk_forward(self, test_X, test_y) -> dict (internal)

If the holdout has fewer than 100 samples, warns and returns {"samples": 0}. Otherwise predicts each holdout sample (direction only — confidence ignored; unlike the bot’s _load_dataset, no online learning during the test), counts wins/losses, and logs:

🧪 WALK-FORWARD TEST (30% holdout): 41230 trades
   Win: 21876 | Loss: 19354 | HONEST WR: 53.1%
   ✅ EDGE: +1.0% above breakeven (52.1%)

Breakeven = 1 / (1 + payout) × 100 with payout = 0.92 → ≈ 52.1%. Returns {"samples", "wins", "losses", "win_rate", "breakeven"} (win_rate and breakeven in percent).

Usage example

from src.training import Trainer, TrainerConfig

cfg = TrainerConfig(dataset_path="EURUSD_M1.csv", brain_path="athena_brain.pkl",
                    expiry=300, use_nn=True)
metrics = Trainer(cfg).run()
print(metrics["win_rate"], metrics["breakeven"])

Or from the shell: python train.py --data EURUSD_M1.csv --expiry 300 — see the training guide.

Risk note

A walk-forward win rate above breakeven on historical data does not guarantee live profitability. Binary options trading can lose money.

See also

• CLI flag reference, dataset.md, device.md, torch-model.md
• GPU guide

💬 Unsure how this interacts with the rest of the configuration? Ask in the Chipa Discord — or prototype strategies no-code with ChipaEditor.