Backtesting

QuantRL-Lab uses a job-based backtesting architecture. An ExperimentJob describes a single training + evaluation run. BacktestRunner executes jobs and returns ExperimentResult objects.

For a complete walkthrough of a single job, see Quickstart — Use BacktestRunner. For the full API including BacktestEnvironmentBuilder, grid search, feature importance, and Optuna tuning, see Experiments.

Running a Batch

Run multiple algorithm/config combinations and compare:

from stable_baselines3 import PPO, SAC, A2C

jobs = [
    ExperimentJob(algorithm_class=PPO,  env_config=env_config, total_timesteps=100000),
    ExperimentJob(algorithm_class=SAC,  env_config=env_config, total_timesteps=100000),
    ExperimentJob(algorithm_class=A2C,  env_config=env_config, total_timesteps=100000),
]

results = runner.run_batch(jobs)
BacktestRunner.inspect_batch(results)

Result Structure

ExperimentResult contains:

Field	Description
result.status	"completed" or "failed"
result.metrics	Dict with train_avg_return_pct, test_avg_return_pct, test_avg_sharpe_ratio, test_avg_max_drawdown, etc.
result.model	Trained stable-baselines3 model
result.train_episodes	List of per-episode dicts for training set
result.test_episodes	List of per-episode dicts for test set
result.top_features	Top correlated features from AgentExplainer
result.execution_time	Wall-clock seconds

Train/Test Split

Use ratio-based or date-based partitioning from quantrl_lab.data.partitioning:

# Ratio-based (simple)
split_idx = int(len(df) * 0.8)
train_df = df.iloc[:split_idx]
test_df  = df.iloc[split_idx:]

# Date-based
from quantrl_lab.data.partitioning.date_range import DateRangeSplitter
splits = DateRangeSplitter({
    "train": ("2020-01-01", "2022-12-31"),
    "test": ("2023-01-01", "2023-12-31"),
}).split(df)
train_df = splits["train"]
test_df = splits["test"]

Saving and Loading Models

# Save
result.model.save("ppo_aapl")

# Load and evaluate manually
from stable_baselines3 import PPO
model = PPO.load("ppo_aapl")

env = env_config.test_env_factory()
obs, _ = env.reset()
done = False
while not done:
    action, _ = model.predict(obs, deterministic=True)
    obs, reward, terminated, truncated, _ = env.step(action)
    done = terminated or truncated

See Also

• Configuration — environment and strategy parameters
• Custom Strategies — build your own strategies
• notebooks/backtesting_example.ipynb — interactive end-to-end example