perfattr

perfattr is a small, auditable portfolio performance-attribution calculation library built with pandas and NumPy.

The package provides a reusable Brinson attribution calculation core and a portable preparation layer for source-period weights and returns. Portfolio accounting, vendor schemas, and presentation remain outside the package boundary.

Main features

• Accept weights and returns, or authoritative contributions when accounting results are available.
• Validate, select, and align portfolio and benchmark histories.
• Consolidate smaller source periods into complete monthly, quarterly, or yearly reporting periods using calendar and holiday rules.
• Apply static or effective-dated classification mappings before consolidation.
• Calculate Brinson-Fachler or Brinson-Hood-Beebower with compact two-effect selection or explicit three-effect selection and interaction.
• Link contributions logarithmically and attribution effects using Carino, Frongello, or Menchero optimized linking, with Carino retained as the default.
• Roll already calculated leaf attribution into a static hierarchy without recalculating Brinson effects at parent levels.
• Preserve zero-weight fee and financing contributions without inventing returns.
• Return deterministic pandas result frames with explicit financial reconciliation.

The completed initial calculation roadmap is recorded in _extras/perfattr_roadmap_1.md. The completed portable preparation work is recorded in _extras/perfattr_roadmap_2.md, while later candidates are kept in the noncommitted _extras/perfattr_roadmap_3.md. Effective-dated classification was the first promoted candidate and its completed work is recorded in _extras/perfattr_roadmap_4_effective_dated_classification.md, with its accepted contract in docs/effective_dated_classification_specification.md. The released opt-in Brinson-Fachler three-effect work is recorded in _extras/perfattr_roadmap_5_brinson_fachler_three_effect.md and docs/brinson_fachler_three_effect_specification.md. The released Brinson-Hood-Beebower three-effect work is recorded in _extras/perfattr_roadmap_6_brinson_hood_beebower_three_effect.md and docs/brinson_hood_beebower_three_effect_specification.md. The released compact Brinson-Hood-Beebower work is recorded in _extras/perfattr_roadmap_7_brinson_hood_beebower_two_effect.md and docs/brinson_hood_beebower_two_effect_specification.md. The released opt-in Frongello effect-linking work is recorded in _extras/perfattr_roadmap_8_frongello_recursive_linking.md and docs/frongello_recursive_linking_specification.md. The released opt-in Menchero optimized-linking work is recorded in _extras/perfattr_roadmap_9_menchero_optimized_linking.md and docs/menchero_optimized_linking_specification.md. The release-candidate hierarchical result-roll-up work is recorded in _extras/perfattr_roadmap_10_hierarchical_result_rollup.md and docs/hierarchical_result_rollup_specification.md. The complete portable calculation contract is defined in docs/specification.md, and the accepted roadmap 2 preparation contract is in docs/preparation_specification.md.

import pandas as pd

from perfattr import calculate_attribution, prepare_attribution

portfolio = pd.DataFrame(
    [
        {
            "from_date": "2024-01-01",
            "thru_date": "2024-01-31",
            "identifier": "Equity",
            "weight": 0.60,
            "return": 0.04,
        },
        {
            "from_date": "2024-01-01",
            "thru_date": "2024-01-31",
            "identifier": "Bonds",
            "weight": 0.40,
            "return": 0.01,
        },
    ]
)
benchmark = pd.DataFrame(
    [
        {
            "from_date": "2024-01-01",
            "thru_date": "2024-01-31",
            "identifier": "Equity",
            "weight": 0.50,
            "return": 0.03,
        },
        {
            "from_date": "2024-01-01",
            "thru_date": "2024-01-31",
            "identifier": "Bonds",
            "weight": 0.50,
            "return": 0.015,
        },
    ]
)

prepared = prepare_attribution(portfolio, benchmark)
result = calculate_attribution(prepared.portfolio, prepared.benchmark)
print(result.period_detail)

Hierarchical result roll-up

Use roll_up_attribution after calculation when leaf results also need additive parent totals:

from perfattr import roll_up_attribution

hierarchy = pd.DataFrame(
    {
        "identifier": ["Equity", "Bonds"],
        "parent_identifier": ["Total", "Total"],
    }
)
parent_result = roll_up_attribution(result, hierarchy)
print(parent_result.period_rollup)

The hierarchy is a static child-to-parent forest with complete leaf coverage. Parent effects are sums of the already calculated descendant effects; the function does not rerun Brinson formulas or linking at a higher level. period_rollup includes parent effective returns derived from summed weight and authoritative contribution. overall_rollup deliberately omits returns and cumulative values because the source result does not retain enough information to reconstruct them faithfully in every case. Leaf rows remain in the original AttributionResult, while the hierarchy result contains parent rows only and includes explicit parent and root reconciliation.

Attribution methods

The default remains the released two-effect Brinson-Fachler convention: allocation is reported separately, while portfolio-weighted selection absorbs interaction. The public method enum also provides an explicit-interaction Brinson-Fachler calculation and compact or explicit-interaction Brinson-Hood-Beebower (BHB) calculations:

from perfattr import AttributionMethod, calculate_attribution

three_effect = calculate_attribution(
    prepared.portfolio,
    prepared.benchmark,
    method=AttributionMethod.BRINSON_FACHLER_THREE_EFFECT,
)
bhb_three_effect = calculate_attribution(
    prepared.portfolio,
    prepared.benchmark,
    method=AttributionMethod.BRINSON_HOOD_BEEBOWER_THREE_EFFECT,
)
bhb_two_effect = calculate_attribution(
    prepared.portfolio,
    prepared.benchmark,
    method=AttributionMethod.BRINSON_HOOD_BEEBOWER_TWO_EFFECT,
)
print(
    three_effect.period_detail[
        ["allocation_effect", "selection_effect", "interaction_effect"]
    ]
)

For portfolio and benchmark weights wP and wB, effective returns rP and rB, and total benchmark return B, the explicit effects are:

Brinson-Fachler allocation = (wP - wB) * (rB - B)
BHB allocation             = (wP - wB) * rB
selection                  = wB * (rP - rB)
interaction                = (wP - wB) * (rP - rB)
compact selection          = total - allocation

BHB uses the group's absolute benchmark return, so overweighting a positive-return group produces positive allocation even when that return trails the total benchmark. Brinson-Fachler instead measures the group return relative to the total benchmark; the two methods can therefore assign opposite allocation signs. perfattr reports the signed formulas without labeling an effect favorable or unfavorable.

At identifier level, BHB total_effect is unadjusted active contribution, while the Brinson-Fachler total includes its benchmark-relative adjustment. With exactly normalized portfolio and benchmark weights, their period totals agree. Supplied contribution remains authoritative. If either effective return is undefined, interaction is zero and selection retains the reconciled residual rather than inventing a return.

Compact BHB is a derived perfattr reporting convention, not a claim that the original BHB methodology defined a historical two-effect model. It retains BHB allocation and total, omits the interaction column, and reports selection directly as total_effect - allocation_effect. When returns are defined, compact BHB and compact Brinson-Fachler therefore share portfolio-weighted selection but can assign different identifier-level allocation and total values.

The opt-in result inserts interaction_effect immediately after selection_effect and linked_interaction_effect immediately after linked_selection_effect wherever those channels apply. Cumulative output also places cumulative_interaction_effect immediately after cumulative_selection_effect. AttributionResult.method records the selected convention. See the Brinson-Fachler specification, BHB three-effect specification, and compact BHB specification for the complete schemas, linking rules, null policies, and independently calculated examples.

Effect linking

Carino remains the default effect linker. Frongello is an explicit opt-in for clients that need path-dependent recursive linking. Menchero optimized linking is an opt-in for clients that require order-independent, minimum-correction allocation of the multi-period compounding residual:

from perfattr import EffectLinkingMethod, calculate_attribution

frongello_result = calculate_attribution(
    prepared.portfolio,
    prepared.benchmark,
    effect_linking_method=EffectLinkingMethod.FRONGELLO,
)
menchero_result = calculate_attribution(
    prepared.portfolio,
    prepared.benchmark,
    effect_linking_method=EffectLinkingMethod.MENCHERO,
)
print(frongello_result.effect_linking_method)
print(menchero_result.effect_linking_method)

The option changes only linked allocation, selection, optional interaction, and total effects. Portfolio and benchmark contributions remain logarithmically linked, and unlinked effects and result-frame schemas do not change. Each period-detail linked effect is the originating source-period effect allocated to the complete requested horizon; intermediate cumulative rows are partial sums of those allocations, not independent as-of calculations.

Frongello's coefficient for a source period depends on earlier portfolio growth and later benchmark growth, so changing the economic chronology can change the allocation among effect channels. Menchero uses one common horizon scale plus the smallest least-squares period corrections needed to reconcile the compounded active return. Moving complete economic periods therefore moves their coefficients with them without changing complete-horizon effect totals. A one-period calculation is the identity under both policies.

Neither policy changes portfolio or benchmark contribution linking. Cash, fees, financing, authoritative contribution, signed weights, and missing identifiers all receive the same ordinary coefficient as other effects in their source period. See the Frongello specification and Menchero specification for formulas, worked examples, ordering behavior, and compatibility contracts.

GRAP is intentionally not exposed as another linker name because its unrolled result at this library's source-period boundary is the same numerical allocation already provided by Frongello. A duplicate public identity would imply a calculation choice where none exists.

Canonical CSV inputs can be loaded with read_performance_csv; optional mapping and classification readers are also available at the package root.

Mapping CSV files are headerless and use one uniform form per file. Existing static files remain two columns (identifier,classification_identifier). Effective-dated files use four columns in the order from_date,thru_date,identifier, classification_identifier. For example:

2024-01-01,2024-01-31,ASSET,Equity
2024-02-01,2024-12-31,ASSET,Fixed Income

The dates are closed and inclusive. A source period for a mapped identifier must be contained in exactly one assignment; perfattr does not split a source period at a classification boundary.

Cash, fees, and financing

Cash receives no special treatment: supply it as an ordinary identifier, or map it to a Cash classification, with the weight and return chosen by the host accounting system. perfattr never invents cash or hides a residual in it.

A fee or financing charge without exposure can be supplied as a zero-weight row with an authoritative nonzero contribution and a null return. The contribution is preserved and included in the ordinary attribution and linking calculations; perfattr does not infer the row from its name or calculate the charge. Financing with an explicit exposure and return can instead be represented as an ordinary identifier.

Development

Create and activate a virtual environment:

python3 -m venv .venv
source .venv/bin/activate

Install the package and development dependencies:

python -m pip install --editable ".[dev]"

Run the initial checks:

python -m pytest
python -m pylint src/perfattr tests scripts
python -m pyright

Run the four roadmap performance workloads:

python scripts/benchmark_core.py --samples 5
python scripts/benchmark_core.py --samples 5 --input-form authoritative
python scripts/benchmark_preparation.py --samples 5
python scripts/benchmark_hierarchy.py --samples 5

Pass --method three-effect, --method bhb-three-effect, or --method bhb-two-effect to benchmark_core.py to measure an opt-in calculation; the default remains --method two-effect. Pass --effect-linking-method frongello or --effect-linking-method menchero to measure an opt-in linker; the benchmark default remains Carino. The hierarchy benchmark runs all four attribution methods and all three linkers by default across the established selected-result sizes and hierarchy depths. Repeat --method or --effect-linking-method to select a smaller policy subset.

Add --workload monthly_121260 --profile to inspect one workload's cumulative calculation-core call profile. The preparation benchmark compares static and effective-dated mappings through quarterly consolidation. The benchmark methodology and observations are recorded in docs/performance.md.

License

perfattr is distributed under the MIT License.