EIA861 Reliability Deterioration
Given three years of published distribution-reliability history and utility attributes, will a U.S. electric distribution utility's day-to-day outage duration (SAIDI excluding major event days) jump by at least 50% above its own recent norm in the next calendar year?
Versioned research data. Read the suitability requirements before interpreting scores.
Suitability and evaluation
Evaluation setting: retrospective research. Read the prediction setting and limitations below. Public validation is development data; downloadable holdout labels must stay outside the evaluated agent workspace.
At a glance
| Question | Given three years of published distribution-reliability history and utility attributes, will a U.S. electric distribution utility's day-to-day outage duration (SAIDI excluding major event days) jump by at least 50% above its own recent norm in the next calendar year? |
|---|---|
| Field | U.S. Energy Information Administration (EIA), U.S. Department of Energy / binary-classification |
| Rows | 5,809 |
| Documented splits | 3 |
| This version | v-498e55482c3ed94b |
| Released | 2026-11-13 |
| Validation ROC AUC | 0.6963780731791774 |
| Validation Brier | 0.1190485278420198 |
| Artifact | 6aa939acba34270679973c9e8b5ea12984814b146a32ca50ec71d3b23ac2790a |
Get this dataset
How to load it
from datasets import load_dataset
ds = load_dataset('earino/factory-eia861-reliability-deterioration', revision='v-498e55482c3ed94b')
The version links and checksums identify this artifact. Exclude the labelled holdout from the agent's workspace when running an evaluation.
What is inside 
| Split | Rows | Positive labels |
|---|---|---|
| eval | 1,294 | — |
| holdout | 2,062 | — |
| train | 2,453 | — |
Files and checksums
| Download | Bytes | SHA-256 |
|---|---|---|
| LICENSE-EVIDENCE.json | 820 | e42a16e9013a232d5a73a5ffc1db6197aa7417d7cf573d88612cf3211c74fed9 |
| LICENSE.md | 1,444 | 3a6aa6fe2a62707ee51429f107f3255567836b96c3e6d41d071e251179e1f65e |
| README.md | 7,792 | 2b032e6f12130f1b8ca4771864c2a374d4a3965e4c1860ec6813297ab3059e78 |
| SOURCES.json | 4,258 | 6f062650ffce5320e2b6da73fff1edd6d2b2f5cd33510387388698b51aa89a9f |
| evaluation.json | 1,846 | 73b4951b670f6482eea54ea9e5b3ff4c99c886860275be3f616112c8de81bdaf |
| extract__meta.json | 1,998 | 5623e36dec6ec0a313cb0e3d0b2cd033b4731519423d433cb3a7dea42d8e785d |
| extract__private__holdout.csv | 290,700 | 721984dd348b46d1a1e468cf170d7a6e28e3ecf744919a51d7211183972061ff |
| extract__public__eval.csv | 182,562 | 7a6310d4685523a3655cec1b57dd9e1dcf2df739d4bac4540a6bf45b84ce6d78 |
| extract__public__train.csv | 341,953 | 52107eccf49ebf0d76599065baa13c9c7dcad721a27e4e446273470b404c8782 |
| extract__quality.json | 23,057 | f8f15851b6407e76888e87180f338f5e7a0b66994a7ac2517e02cba0a26703b3 |
| get_dataset.py | 3,680 | fad89222d50f9d9709225c34909be74c591df409c08b8b5951eba9d3cadfdd48 |
| qualification.json | 20,313 | 6f7982722ce91800a467deb1fc3f681cedc3a247802b0360fd70f8e5fd174331 |
| reproduction__LICENSE-CC0-1.0.txt | 7,048 | a2010f343487d3f7618affe54f789f5487602331c0a8d03f49e9a7c547cf0499 |
| reproduction__LICENSE-MIT.txt | 1,124 | 56b932ff064e5da113ae10175c8546325f64f26e1ece71683ffcbf8b2304fe1c |
| reproduction__artifact_scrub.py | 11,369 | 8c663253e691a7b8751f9c853b40a3752a833360caafb1ea8e8f534e64af4b68 |
| reproduction__build.py | 47,905 | 268e7e918df63581a60e1596426520f0901317c3e3d9fb6a97309a8458b4abb7 |
| reproduction__construction.json | 22,694 | 4ba0e46f119e0524f2dbff39e36f50f0d22cdb939efc9cfd8b3f582b9adab1f0 |
| reproduction__dataset_package.py | 7,927 | c2364defe58dcc99322be59a1bb04894ce00e823ce37afeb37bc31012f744327 |
| reproduction__dataset_run.py | 12,381 | 8085fae14d6e93b852e692e18c1ef72113a3601f48f78901b6d1ae6db278de57 |
| reproduction__get_dataset.py | 3,680 | fad89222d50f9d9709225c34909be74c591df409c08b8b5951eba9d3cadfdd48 |
| reproduction__leakage_battery.py | 17,379 | adbf7ee0aca0012b78cc1f1e4aeb968a23717cae30f66a02e080215e75f830e3 |
| reproduction__provenance.json | 573 | c4b7e3bae1aa9b5152d0080fd6ef89ce05da483deaa897e8ffa979063150217c |
| reproduction__qualify_dataset.py | 66,067 | bff6d43d0904182e2612c32bab702e41addbf389904a236c209e78be39e37a6b |
| reproduction__recovery_recipes.py | 17,609 | f64c542fcda232169c95ace17421b45275194b934938c0ea4c9b68e6f844ad0e |
| reproduction__requirements.txt | 160 | 5c72f3268174d20364624f76680e189c408a05dba2e7b8cfe12d077b3865e7f0 |
| reproduction__source.json | 20,778 | 11411e17d4813c8115c5b8c79fba1ce04fca81879ce7162ec21ebab38b8bc819 |
| reproduction__source_io.py | 5,607 | 0c60eed4968cd342148efbaac7e4fa98f00708d8e6c14c0929928687b9af1ea3 |
| reproduction__train.py | 2,500 | a3c6bcf13735bc85c52129ded68f839090dffdc266ebc3810dc61b2e6ea5e7e8 |
| reproduction__validate.py | 1,277 | b597e7f84fed614e64b4a86fbecbd6ec0145916a24d0d221a796586976418e7b |
| reproduction__validate.sh | 482 | 3f06ca48f11c2e05d331b4d8394284925de660c1d4b086d2f8243959e83b7e6d |
| source-checks.json | 9,335 | ce545f3813bf6e09a7d13fc04e2bea36b5e78c4bdbe087df0c5ef5c746d2ba82 |
| source-snapshot.tar.gz | 60,145,119 | 0ededf1f2b306f524d2aa3111541d041c2542ff2651d1ec9031adda7050cabe6 |
| verify_dataset.py | 1,276 | f4a4bb7f565cdbd075c520621f881876b491eaed91a5e6e70afb31e4753f9834 |
Release notes · 2026-11-13
Given three years of published distribution-reliability history and utility attributes, will a U.S. electric distribution utility's day-to-day outage duration (SAIDI excluding major event days) jump by at least 50% above its own recent norm in the next calendar year?
Population: U.S. electric distribution utilities identified by EIA utility number that report a non-missing SAIDI excluding major event days in CY Y-3, Y-2 and Y-1 under one consistent reporting standard, report at least 1,000 customers in CY Y-1, carry an ownership code among investor-owned, municipal, cooperative, political subdivision, federal or state, appear with exactly one reliability record in each of those years, and report a SAIDI excluding major event days for CY Y+1. Utilities filing multiple state rows or changing utility number are dropped for that cohort. Missing CY Y+1 outcomes are dropped, never coded negative. No CY Y or later value enters features.
Target: Binary y=1 if the utility's reported SAIDI excluding major event days for calendar year Y+1 is greater than or equal to 1.5 times the median of its three reported SAIDI-excluding-major-event-day values for CY Y-3, Y-2 and Y-1; y=0 otherwise. Utilities with a missing CY Y+1 SAIDI-excluding-MED value are dropped, never coded 0.
Prediction time: 31 December of forecast year Y, using only Form EIA-861 reliability and utility attribute filings for calendar years Y-3, Y-2 and Y-1. Exact historical publication dates are assumed rather than verified. No CY Y or later value enters features.
Splits: Cohorts are forecast years Y = 2016, 2017, ..., 2024 (labels in CY 2017 ... 2025). Fixed chronological boundaries: train = Y 2016-2019, eval = Y 2020-2021, holdout = Y 2022-2024. No random reassignment and no re-splitting after measurement. Purge rule enforced in code: every feature year is <= Y-1, so the newest training feature year is 2018 while the earliest scored label year is 2021; therefore no training row's feature vector contains a year used to compute any eval or holdout label. Rows whose four-year window (Y-3 ... Y+1) spans a utility-number change or a change of reporting standard are dropped rather than carried across splits.
Outcome horizon: Calendar year Y+1, i.e. the twelve months beginning immediately after the decision instant; the outcome is observed when the CY Y+1 EIA-861 reliability file is published (roughly 9-12 months later).
Inputs: saidi_nomed_lag1: SAIDI excluding major event days for CY Y-1: minutes of non-momentary interruption the average customer experienced.; saidi_nomed_lag2: SAIDI excluding major event days for CY Y-2.; saidi_nomed_lag3: SAIDI excluding major event days for CY Y-3.; saidi_nomed_trend: Ratio of lag1 minus lag3 to lag3, a three-year level trend in non-storm outage duration. Missing when lag3 is zero; no interpolation.; saidi_with_med_lag1: SAIDI including major event days for CY Y-1, showing how much recent duration came from storm-scale days.; saifi_nomed_lag1: SAIFI excluding major event days for CY Y-1: non-momentary interruptions per customer-year.; saifi_with_med_lag1: SAIFI including major event days for CY Y-1.; caidi_nomed_lag1: CAIDI excluding major event days for CY Y-1: average minutes to restore an interruption. Derived from SAIDI divided by SAIFI when the reported CAIDI is absent and SAIFI is positive.; customers_lag1: Customer or meter count reported with the CY Y-1 Reliability schedule, used as a scale control.; reporting_standard: Whether the utility reported under IEEE 1366 or an internally defined other method in CY Y-1, normalized to IEEE or OTHER. A utility must carry one consistent value across the four filings used by a cohort.; ownership_code: Utility ownership type among investor-owned, municipal, cooperative, political subdivision, federal or state, from the CY Y-1 Reliability filing or the annual utility data file.; state: Utility reported state, used for calendar and regulatory-regime controls.
Recorded structural checks and baselines
PASSED Gate 1.6.1.
Existing unmodified runner baseline on public validation only; no holdout scoring
These historical checks do not establish availability at each prediction time or protection against every shortcut. Scores below describe the recorded protocol.
Limits
Reliability indices are self-reported and unaudited by EIA; utilities may use IEEE 1366 or an internally defined method, and non-IEEE reporters self-determine which days are major event days, so the excluding-MED values are not strictly comparable across utilities.
Annual reliability is dominated by weather and storm occurrence, and the feature set deliberately contains no storm, outage-cause or capital-spending covariates, so a large share of label variance is unpredictable by construction; the task measures screening skill, not causal drivers.
This is a current-vintage retrospective built from today's published files: the values a regulator would have seen at the decision instant could differ from the final vintage, so no claim of proven historical operational availability is made.
Workbook layout is not stable across vintages: IEEE and other-standard column blocks coexist, some early files are .xls, and index columns were reportedly relabelled around 2021, so the parser must match headers by pattern and be re-checked for each new year.
The panel recurs utilities across cohorts; features are capped at CY Y-1 and a purge is enforced at split boundaries, but autocorrelation within splits still makes the effective sample size smaller than the row count.
Utilities that merge, change EIA utility number, file multiple state rows, or switch reporting standard are dropped, which removes some large multi-state investor-owned utilities and slightly favours stable filers.
The label multiplies a utility's own baseline by a fixed 1.5 factor, so a utility with an unusually low or unusually volatile three-year baseline is mechanically easier or harder to flag; the rule is frozen before any performance measurement and not re-tuned.
The assumption that a calendar year's final EIA-861 reliability data are public by 31 December of the following year is inferred from observed release timing and the publisher's file listing; freeze exact file vintages and record checksums in a manifest so a later re-release cannot silently change the benchmark.
EIA also publishes early-release versions that exclude utilities pending validation; only final-vintage ZIPs should be used, and the 2025 file should be confirmed final at build time.
Upstream third-party material is excluded: EIA photographs are privately licensed and the EIA logo is a registered trademark that must not be reproduced; only the numeric data are redistributed, and EIA's requested acknowledgment with a publication date must accompany any release.
Reliability indices are self-reported and unaudited by EIA. Utilities may use IEEE 1366 or an internally defined method, and non-IEEE reporters self-determine major event days, so excluding-MED values are not strictly comparable across utilities.
Annual reliability is dominated by weather and storm occurrence. The declared feature set contains no storm, outage-cause or capital-spending covariates, so a large share of label variance is unpredictable by construction.
This is a current-vintage retrospective built from today published files. Values a regulator would have seen at the decision instant could differ from the final vintage, and exact historical publication dates were not verified.
Workbook layout is not stable across vintages. The parser matches headers by text pattern, allows both IEEE and other-standard blocks, and fails with observed headers when required fields are missing. Legacy .xls vintages are read by an embedded BIFF reader rather than an external xls package.
The panel recurs utilities across cohorts. Features are capped at CY Y-1 and no holdout is used to choose the target, features, thresholds or splits, but autocorrelation within splits still makes effective sample size smaller than row count.
Utilities that merge, change EIA utility number, file multiple state rows or switch reporting standard are dropped. This removes some large multi-state investor-owned utilities and slightly favors stable filers.
The label multiplies a utility own baseline by a fixed 1.5 factor, so a utility with an unusually low or volatile three-year baseline is mechanically easier or harder to flag. The rule was frozen before measurement and is not re-tuned.
The assumption that a calendar year final EIA-861 reliability file is public by 31 December of the following year is inferred from release practice, not from a complete publication ledger. Timing observations are deliberately empty because actual publication dates were not verified; the 365-day value is a conservative calendar bound, not a measured publisher lag.
No persistence score column is shipped. The lag-1 SAIDI level is a feature, not a calibrated persistence score for a relative jump, so the persistence reference is null rather than fabricated.
The 2025 file is treated as final based on publisher evidence. If a later re-release changes it, the holdout labels can move.
The annual source has no day-level event date. Decision dates collapse to one day per forecast year, so event clustering uses the utility identity as the conservative independent unit available; this is coarser than a genuine outage-event date.
No uncertainty interval; observed differences do not establish universal headroom
Source, licences, and citation
U.S. Energy Information Administration (EIA), U.S. Department of Energy
Source: U.S. Energy Information Administration, Form EIA-861, Annual Electric Power Industry Report.
Source: U.S. government work, public domain under 17 U.S.C. Sec. 105 as stated by EIA: data and files may be used and distributed; EIA requests an acknowledgment naming EIA and a publication date. No publisher licence grant is required or implied for third-party material EIA hosts (photographs, logo).. Compilation: CC0-1.0. Code and documentation: MIT.
Our additional compilation rights only; upstream terms remain in force.
Suggested citation
U.S. Energy Information Administration, Form EIA-861, Annual Electric Power Industry Report.