End-to-end workflow

MFDRO begins after source-data engineering and ends before portfolio policy. Keeping those boundaries visible is essential to interpreting the result.

Research pipeline

prices and corporate actions
    ↓ user-owned transformation
point-in-time daily simple returns + membership + calendar
    ↓ MFDRO validation and aggregation
frequency-specific empirical distributions
    ↓ scaling, barycenter, transport dispersion
rho + sqrt_rho + audit artifacts
    ↓ user-owned calibration
radius, regime, or model-combination policy
    ↓ optimizer and simulator
orders, holdings, costs, and performance

Before MFDRO

The caller decides and documents:

• price field, adjustment convention, currency, and timestamp convention;
• how simple returns are calculated;
• delistings, corporate actions, stale prices, and non-trading observations;
• the point-in-time investable universe for every formation month;
• the authoritative observation calendar;
• source snapshots, licences, and checksums.

MFDRO cannot reconstruct these choices from a return matrix. A full matrix can still contain survivorship bias or future information.

Inside MFDRO

For every formation date, the path API:

1. selects the configured calendar-month lookback;
2. selects the formation month's asset membership, if supplied;
3. rejects missing or invalid selected values;
4. checks the authoritative calendar, when supplied;
5. compounds base returns into each configured frequency;
6. places the empirical measures on the configured scale;
7. constructs the configured center and estimates dispersion;
8. records result identity, sample sizes, seed, and window evidence.

Call validate_path_inputs before step 6 to inspect every window without paying for transport geometry.

After MFDRO

rho must be transformed by an explicit downstream research policy. Examples include a time-varying ambiguity radius, a regime label, or a model weight. The mapping must be estimated using only information available at its decision time.

A portfolio workflow must additionally specify optimization constraints, solver behavior, execution delay, prices, costs, and accounting. MFDRO can feed general tools such as CVXPY, cvxportfolio, vectorbt, skfolio, or an internal engine; it does not require a particular downstream package.

Minimum reproducible hand-off

Pass the next researcher:

• the source transformation code and frozen source identifiers;
• the daily return panel, membership ledger, and calendar used;
• config.json or equivalent configuration JSON;
• the complete saved SignalPath directory;
• exact MFDRO, Python, and dependency versions;
• the downstream policy and backtest configuration.

The package can reproduce its own calculation from those inputs. It cannot make replaceable market-data downloads historically immutable.