Configuration

Every run is controlled by a dictionary that is converted into TradeConfig. Treat this config as the experiment contract: it records what data was used, what dates were tested, how orders were filled, and where outputs were written.

Minimal Config

config = {
    "data_loader": data,
    "tickers": ["AAPL"],
    "date_from": "2024-01-02",
    "date_to": "2024-03-01",
    "setup_name": "demo_aapl",
    "execution_timing": "next_open",
    "save_results": True,
    "silence": True,
}

This is enough to run a basic single-ticker backtest. Add costs and liquidity controls before using results for serious comparison.

Required Decisions

Field	Example	How to choose
`data_loader`	`FinsaberParquetDataset(path)`	Object implementing `TradingData`.
`tickers`	`["AAPL", "MSFT"]` or `"all"`	Use a small list for debugging; use `"all"` only when the loader and strategy can handle the universe.
`date_from`, `date_to`	`"2020-01-01"`	Inclusive test range. Keep early smoke tests short.
`setup_name`	`"ma_cross_2020"`	Stable name used in output directories.
`execution_timing`	`"next_open"`	Prefer `next_open` for date-level features.

Capital And Cost Fields

Field	Default	Meaning
`cash`	`100000.0`	Initial portfolio value.
`risk_free_rate`	`0.03`	Annual risk-free rate used in Sharpe and Sortino calculations.
`commission_per_share`	`0.0049`	Per-share commission before min/max constraints.
`min_commission`	`0.99`	Minimum commission per trade.
`max_commission_rate`	`0.01`	Maximum commission as a fraction of transaction value.
`slippage_perc`	`0.0`	Fixed adverse price movement per trade.
`slippage_impact`	`0.0`	Extra adverse price movement based on participation rate squared.
`liquidity_cap_pct`	`0.0`	Max order size as a fraction of average prior volume.
`llm_cost_as_trade_cost`	`True`	Adds tracked LLM costs into `total_trading_cost`.

Practical Presets

Fast debuggingResearch comparisonLLM strategy

config.update({
    "tickers": ["AAPL"],
    "date_from": "2024-01-02",
    "date_to": "2024-02-01",
    "save_results": False,
    "silence": True,
})

config.update({
    "execution_timing": "next_open",
    "commission_per_share": 0.0049,
    "min_commission": 0.99,
    "slippage_perc": 0.0005,
    "liquidity_lookback_days": 20,
    "liquidity_min_history_days": 5,
    "liquidity_cap_pct": 0.025,
    "save_results": True,
})

config.update({
    "execution_timing": "next_open",
    "llm_cost_as_trade_cost": True,
    "save_results": True,
    "log_base_dir": "runs",
})

Rolling Windows And Selection

Rolling-window runs split the full date range into repeated evaluation periods. A selector can choose the ticker universe for each window.

config.update({
    "rolling_window_size": 2,
    "rolling_window_step": 1,
    "selection_strategy": my_selector,
})

Selectors must use only data available inside the allowed training or prior window. Do not select tickers using returns from the future test period.

Output Fields

Field	Meaning
`save_results`	If `True`, writes CSV and JSON artifacts.
`log_base_dir`	Root output folder. Defaults to `backtest/output`.
`result_filename`	Optional pickle filename for legacy result compatibility.
`print_trades_table`	Enables more verbose trade printing where supported.
`silence`	Suppresses progress output and plots when `True`.

What To Save In Papers Or Reports

Always report the config fields that materially affect performance: ticker universe, date range, execution timing, adjusted price policy, commission, slippage, liquidity cap, and whether LLM cost is counted. The generated run_config.json is designed to make this reproducible.