Skip to content

Backtesting integration

MFDRO is a signal component, not a backtesting engine.

Intended pipeline

point-in-time returns and memberships
    → MFDRO rho and audit
    → radius or regime policy
    → portfolio optimizer
    → target weights
    → next tradable execution
    → holdings, costs, and return ledger
    → performance and inference

The separation is intentional. Data eligibility, optimization, execution, and accounting rules vary by market and should not be hidden inside the signal estimator.

Timing rule

An estimate dated t may use observations through t. A backtest must not assume that the resulting weights were held during the same observations. Execution must occur at a price and timestamp that were actually available after the signal and optimization completed.

Do not apply a blind row shift without checking the trading calendar. Map each formation timestamp to an explicit decision timestamp and next tradable execution timestamp.

Typical integration

path = engine.estimate_path(
    returns,
    lookback_months=36,
    formation_dates=formation_dates,
    memberships=memberships,
    reference_calendar=trading_calendar,
    on_insufficient="raise",
    seed_namespace="research_universe",
)

signal = path.estimates.set_index("date")["rho"]
radius = radius_policy(signal)  # User-defined and point-in-time.
target_weights = optimizer.solve(radius=radius, data=optimizer_inputs)
ledger = simulator.run(target_weights, execution_schedule)

radius_policy, optimizer, and simulator are deliberately not supplied by MFDRO.

Complementary packages

  • CVXPY is appropriate for expressing a custom convex DRO portfolio problem and selecting an open-source or commercial solver.
  • cvxportfolio provides portfolio policies, constraints, transaction-cost models, and a simulator.
  • vectorbt is useful for fast vectorized diagnostics, order simulation, fees, slippage, and parameter sweeps.
  • skfolio provides portfolio estimators and time-aware model-selection tools such as walk-forward and purged cross-validation.
  • Zipline Reloaded provides an event-driven engine with market calendars and data bundles.

These packages should remain optional integrations, not mandatory MFDRO dependencies.

Minimum backtest ledger

A rigorous downstream ledger should record at least:

  • signal formation timestamp;
  • information cutoff;
  • asset membership and order;
  • radius-policy inputs and output;
  • optimizer status and solver;
  • target and realized weights;
  • execution timestamp and price convention;
  • turnover, fees, slippage, and holding costs;
  • delisting and corporate-action treatment;
  • gross and net returns;
  • configuration and source-data identities.

MFDRO's audit should be retained alongside this ledger rather than reduced to a single merged return series.