train.py CLI reference

🦉 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

train.py is the standalone trainer. It reads a historical candle CSV, trains the full ensemble (online sklearn models, batch GBM/RF, and an optional PyTorch neural model), runs a walk-forward evaluation, and saves a brain file that main.py loads at startup via PO_BRAIN_PATH. No PocketOption connection is needed.

For a task-oriented walkthrough see the training guide. For accepted CSV layouts see data formats.

Synopsis

python train.py [--data PATH] [--out PATH]
                [--timeframe SECONDS] [--expiry SECONDS]
                [--window CANDLES] [--lookback CANDLES]
                [--max-candles N] [--test-split FRACTION]
                [--no-nn] [--cpu]
                [--epochs N] [--batch-size N] [--lr FLOAT] [--hidden SIZES]
                [--check-gpu]

All flags are optional. Each maps to a field of TrainerConfig (src/training/trainer.py), which Trainer.run() consumes.

Flags

Flag	Default	Purpose
`--data`	`EURUSD_M1.csv`	Input candle CSV (HistData or headered format)
`--out`	`athena_brain.pkl`	Output brain file the bot loads
`--timeframe`	`60`	Candle period of the dataset, in seconds
`--expiry`	`300`	Trade expiry to label for, in seconds
`--window`	`20`	Rolling feature window, in candles
`--lookback`	`200`	Max candle history per feature vector
`--max-candles`	`500000`	Use at most the N most recent candles
`--test-split`	`0.30`	Fraction (0–1) held out for walk-forward testing
`--no-nn`	off	Skip the neural model; sklearn ensemble only
`--cpu`	off	Force CPU even if a GPU is available
`--epochs`	`30`	Neural model training epochs
`--batch-size`	`2048`	Neural model mini-batch size
`--lr`	`0.001`	Neural model learning rate (AdamW)
`--hidden`	`128,64`	Neural hidden layer sizes, comma-separated
`--check-gpu`	off	Print the detected training device and exit

--help prints all flags with their defaults (the parser uses argparse.ArgumentDefaultsHelpFormatter).

Exit behavior

Successful training (and --check-gpu) exits with code 0 (main() returns 0, raised via SystemExit).
The trainer aborts via SystemExit with an error message when:
- fewer than 50 candles are parsed from --data: Dataset too small (N candles).
- fewer than 200 valid samples remain after feature generation: Only N valid samples — need at least 200.
If the holdout set has fewer than 100 samples, the walk-forward test is skipped with a warning, but training still completes and the brain is saved.
If PyTorch is not installed, the neural model is skipped with an informational log line; the sklearn ensemble still trains and the run succeeds.

Examples

Train from a HistData M1 export for 2-minute trades and save an asset-specific brain:

python train.py --data GBPUSD_M1.csv --out brains/gbpusd.pkl --expiry 120

Fast CPU-only iteration on a capped dataset, skipping the neural model:

python train.py --data EURUSD_M1.csv --max-candles 100000 --no-nn

Longer neural training with a wider network:

python train.py --epochs 60 --hidden 256,128,64 --lr 5e-4 --batch-size 1024

Check what device the neural model would train on:

python train.py --check-gpu
# Device: cuda  (GPU: NVIDIA GeForce RTX 4070 (12.0 GB VRAM))

Level Up Your AthenaAI Setup

🦉 Get AthenaAI on GitLab → — clone the source, open issues, and contribute improvements.
💬 Join the Chipa Discord → — get real-time support, share your results, and swap strategy ideas with other AthenaAI users.
⚡ Try ChipaEditor → — build, backtest, and deploy trading strategies without writing a line of code, with our AI-powered strategy editor.