FrontierGuard DEA

One rigorous Python toolkit for efficiency, productivity, and network analysis.

FrontierGuard DEA is a new open-source package being built to unify the fragmented DEA workflow behind one typed, inspectable API. Multi-stage and network DEA are architectural primitives, not afterthoughts.

The project is an alpha research preview. The implemented foundation currently includes:

• radial input- and output-oriented envelopment models;
• CRS, VRS, NIRS, and NDRS technologies;
• an optional second phase for slacks, targets, and peers;
• weighted additive, non-oriented SBM, and explicit directional-distance models;
• separate evaluation and reference technologies;
• radial super-efficiency with explicit infeasibility handling;
• scale efficiency and IRS/CRS/DRS diagnosis;
• exact input- and output-oriented Free Disposal Hull models;
• CRS multiplier and solver-selected cross-efficiency models;
• type-I assurance-region restrictions in original data units;
• cost, revenue, technical, and allocative efficiency with common or DMU-specific prices;
• adjacent-period Malmquist productivity panels with audited efficiency and technical-change decomposition;
• a general CRS relational model for any number of serial stages; and
• an experimental graph-defined relational model for acyclic stage networks;
• immutable results with solver status, model conventions, data fingerprints, and independent post-solve audit residuals.

SciPy-HiGHS is the zero-configuration solver. Install the highs extra to use the direct native HiGHS adapter with duals, reduced costs, sparse matrices, and deterministic solver controls.

The larger roadmap includes RAM/BAM and Russell families, environmental models, aggressive/benevolent cross-efficiency, super-SBM, robust partial frontiers, window analysis, global and Malmquist–Luenberger productivity, bootstrap inference, second-stage analysis, dynamic/network SBM, richer reporting, and cross-language validation.

Installation

After version 0.1.0a1 is published to PyPI, install it with:

python -m pip install --pre "frontierguard-dea==0.1.0a1"

The optional native HiGHS adapter is available through the highs extra:

python -m pip install --pre "frontierguard-dea[highs]==0.1.0a1"

Contributors can clone the repository and run uv sync --all-groups. Until the PyPI publication is confirmed, install the reviewed source branch or tag directly from GitHub rather than assuming the index entry exists.

Quick start

import pandas as pd
import frontierguard as fg

frame = pd.DataFrame(
    {
        "dmu": ["A", "B", "C"],
        "staff": [1.0, 2.0, 3.0],
        "cases": [1.0, 1.0, 2.0],
    }
)

data = fg.DEAData.from_frame(
    frame,
    inputs=["staff"],
    outputs=["cases"],
    dmu="dmu",
)

result = fg.RadialDEA(orientation="input", rts="vrs").fit(data)
print(result.to_frame())
print(result.peers("B"))

Multi-stage DEA

network = fg.SerialNetworkData(
    inputs=[[1.0], [2.0], [2.0]],
    links=([[1.0], [1.0], [2.0]],),
    outputs=[[1.0], [1.0], [1.5]],
    dmu_names=("A", "B", "C"),
    stage_names=("operations", "delivery"),
)

result = fg.SerialNetworkDEA().fit(network)
print(result.to_frame())

The serial relational model reports overall and stage efficiencies while verifying that their product closes to the overall score.

For branched or general acyclic production systems, construct NetworkStage, NetworkLink, and NetworkData objects and fit RelationalNetworkDEA. The general engine reports stage-gap contributions and verifies that they sum to overall inefficiency.

Trust policy

FrontierGuard will not label a method “supported” until it has:

1. an explicit mathematical contract and conventions;
2. hand-calculated or published golden cases;
3. primal feasibility and identity tests;
4. scale, permutation, and unit-invariance tests where theoretically valid;
5. a cross-language oracle against an established implementation; and
6. documented failure modes.

See the feature matrix, ecosystem audit, architecture, methods, and validation strategy. The executable examples cover classical, multi-stage, and economic workflows.

Community and project status

FrontierGuard DEA is licensed under the MIT License and maintained by Vamsi Chinda. Before contributing, read the contribution guide, governance, and Code of Conduct. Report security concerns through the private process in SECURITY.md.

Research users can cite the project using CITATION.cff. Release history and compatibility-impacting changes are recorded in CHANGELOG.md.

The publication programme includes a versioned technical white paper (PDF), an explicit AI-use record, and an executable cross-language conformance benchmark. These materials describe current evidence and limitations; they do not imply peer review or complete cross-language validation.