quantlab.backtest.config.ApproximateBacktestConfig#

class quantlab.backtest.config.ApproximateBacktestConfig(*, price_dataset, start_date, end_date, rebalance_periods, cost=0.001, benchmark_dataset=None, chunk_bars=63, max_iterations=20000, tolerance=1e-05)[source]#

Bases: FrozenConfig

Config of an approximate backtest: every rebalance decided at once, the window simulated as arrays.

Built once over a price dataset and run on many (predictions, rule) pairs (quantlab.backtest.approximate.ApproximateBacktest). The rule is not here: it comes to run with its own unchanged config, so a rule chosen by a sweep goes to a backtest run as it is. Execution is reduced to one cost on traded weight; the rest of what is kept and given up is in docs/user-guide/backtesting.md.

Examples

>>> cfg = ApproximateBacktestConfig(
...     price_dataset=price_dataset,
...     start_date="2023-01-01",
...     end_date="2023-12-31",
...     rebalance_periods=5,
... )
>>> cfg.cost, cfg.chunk_bars, cfg.max_iterations, cfg.tolerance
(0.001, 63, 20000, 1e-05)

benchmark_dataset = None#: A market dataset holding exactly one symbol, bought and held on the window’s bars at the same cost, as a backtest run’s benchmark_dataset is; the account then reports its curve, its return statistics and the excess over it. None for none.

chunk_bars = 63#: Bars of the window decided together: a chunk’s decision inputs (its covariance estimates, for a mean-variance rule) are held in memory at once. The chunks are decided in time order, each from the weights decided before it, so the size changes memory, not the result.

cost = 0.001#: Cost per unit of traded weight at each fill, standing for fees and slippage together (a backtest run’s fees + slippage).

end_date#: Last date of the window, inclusive.

max_iterations = 20000#: An iterative rule’s iterations per decision bar before the bar is reported as not converged (quantlab.portfolio.approximate.Stopping).

price_dataset#: The dataset whose prices the simulation trades on.

rebalance_periods#: Rebalance every this many bars from the window’s first bar, as a backtest run does.

start_date#: First date of the window, inclusive.

tolerance = 1e-05#: An iterative rule’s bar has converged when its gradient-mapping residual is at most this times the largest entry of its starting gradient: its objective is then within about twice that of the optimum (quantlab.portfolio.approximate.Stopping).