fix(backtest): guard zero cost_ratio in _get_buy_amount_by_cash_limit

[mateosandoval10]

Description

Exchange._get_buy_amount_by_cash_limit raises ZeroDivisionError whenever the backtest is configured without a proportional trading fee.

The helper works out the cash level at which the proportional fee overtakes the flat min_cost floor:

critical_price = self.min_cost / cost_ratio + self.min_cost

cost_ratio is open_cost + adj_cost_ratio, and adj_cost_ratio derives from impact_cost, which defaults to 0.0. So Exchange(open_cost=0, close_cost=0) makes cost_ratio exactly zero and the division raises. Setting min_cost=0 as well does not help — 0.0 / 0.0 raises too.

When there is no proportional fee the service fee is always min_cost, so no critical price exists and the min_cost branch is already the correct answer. This PR returns it directly, which leaves every non-zero cost_ratio path byte-identical:

if cost_ratio <= 0:
    # Without a proportional fee the service fee is always `min_cost`,
    # so there is no critical price to compare against.
    max_trade_amount = (cash - self.min_cost) / trade_price
    return max_trade_amount

Motivation and Context

No related issue — found while reading the backtest cost model.

The helper is reached from the buy branch of _calc_trade_info_by_order when an order is larger than the available cash:

elif cash < trade_val + max(trade_val * cost_ratio, self.min_cost):
    max_buy_amount = self._get_buy_amount_by_cash_limit(trade_price, cash, cost_ratio)

With cost_ratio == 0 that condition reduces to cash < trade_val + min_cost, the ordinary "order doesn't quite fit in the account" case — so the crash fires on the first partially-funded buy rather than in some corner of the parameter space.

A zero-cost run is a routine baseline: it is how you separate how much of a strategy's result comes from the signal and how much is being consumed by frictions. That configuration currently cannot complete a backtest.

Minimal reproduction:

from qlib.backtest.exchange import Exchange

ex = object.__new__(Exchange)   # the method is pure arithmetic over min_cost
ex.min_cost = 5.0

ex._get_buy_amount_by_cash_limit(trade_price=10.0, cash=1000.0, cost_ratio=0.0015)  # 99.5
ex._get_buy_amount_by_cash_limit(trade_price=10.0, cash=1000.0, cost_ratio=0.0)     # ZeroDivisionError

How Has This Been Tested?

• Pass the test by running: pytest qlib/tests/test_all_pipeline.py under upper directory of qlib.
• If you are adding a new feature, test on your own test scripts.

I was not able to run test_all_pipeline.py, since it needs the official dataset that the README notes is currently disabled. Happy to run it if there is a data source you would like me to use.

Added tests/backtest/test_exchange_cost.py — five unit tests, no data dependency:

• the default fee schedule below the critical price (min_cost branch),
• a cash level above the critical price (proportional branch),
• zero cost_ratio with a non-zero min_cost,
• zero cost_ratio with min_cost=0, where the full balance is investable,
• cash below min_cost, which buys nothing.

Screenshots of Test Results (if appropriate):

1. Pipeline test: not run — see above.

2. Your own tests:

$ python -m pytest tests/backtest/test_exchange_cost.py -q
.....                                                                    [100%]
5 passed in 1.73s

Against unpatched main the two zero-cost_ratio tests fail with ZeroDivisionError and the other three pass, which confirms the existing behaviour is unchanged:

$ python -m pytest tests/backtest/test_exchange_cost.py -q
FAILED tests/backtest/test_exchange_cost.py::TestBuyAmountByCashLimit::test_frictionless
FAILED tests/backtest/test_exchange_cost.py::TestBuyAmountByCashLimit::test_zero_cost_ratio_with_min_cost
2 failed, 3 passed in 1.51s

black . -l 120 --check --exclude qlib/_version.py is clean across the repository (334 files unchanged).

Types of changes

• Fix bugs
• Add new feature
• Update documentation

[mateosandoval10]

@microsoft-github-policy-service agree