tune-parameters

Parameter Tuner

You search parameters without turning in-sample luck into a claim.

Preconditions

• The base strategy logic must already be clear.
• The base spec or SDK strategy must pass validation.
• Data must cover both training and OOS periods.
• The user must approve the metric and search ranges.

Warn when the total grid exceeds 100 combinations.

SDK Pattern

The optimizer API works on Strategy objects, not directly on a YAML spec. If starting from a spec, compile it first or follow examples/modules/07_optimize.py.

from oxq.optimize.paramset import ParameterSet
from oxq.optimize.search import GridSearch

paramset = ParameterSet(name="sma_tuning")
paramset.add("sma_10", "period", list(range(5, 30, 5)))
paramset.add("sma_50", "period", list(range(30, 100, 20)))
paramset.add_constraint("sma_10.period < sma_50.period")

search = GridSearch(paramset).run(
    strategy=strategy,
    market=market,
    broker_factory=broker_factory,
    start="2018-01-01",
    end="2021-12-31",
    metric="sharpe_ratio",
)

Parameter component names must match strategy component names such as indicator aliases in required_indicators. If a name does not match, the parameter has no effect.

Walk-Forward Required

from oxq.optimize.walk_forward import WalkForward

wf = WalkForward(
    paramset=paramset,
    train_period="2Y",
    test_period="1Y",
    step="1Y",
)
wf_result = wf.run(
    strategy=strategy,
    market=market,
    broker_factory=broker_factory,
    start="2018-01-01",
    end="2024-12-31",
    metric="sharpe_ratio",
)
print(wf_result.deterioration())

Time-Series CV

from oxq.optimize.validation import TimeSeriesCV

cv = TimeSeriesCV(n_splits=4, expanding=True)
cv_result = cv.cross_validate(
    strategy=strategy,
    market=market,
    broker_factory=broker_factory,
    start="2018-01-01",
    end="2024-12-31",
    paramset=paramset,
    metric="sharpe_ratio",
)

Overfit Signals

• IS Sharpe far above OOS Sharpe
• best parameters lie on search boundary
• OOS return or Sharpe turns negative
• tiny parameter changes destroy performance
• selected configuration has very few trades