---
id: VOLT-HOME-WP-011
title: "How long do negative electricity-price episodes last?"
slug: how-long-do-negative-electricity-price-episodes-last
description: "A duration-based study of contiguous negative day-ahead price episodes, preserving native market intervals and clearly separating observed episode statistics from household savings claims."
published: 2026-08-30
cluster: "Negative-price event science"
status: measured
evidence_url: "/research-data/home-papers/how-long-do-negative-electricity-price-episodes-last.json"
figure_url: "/research-media/home-papers/how-long-do-negative-electricity-price-episodes-last.webp"
figure_alt: "Chart for How long do negative electricity-price episodes last?: median contiguous negative-price episode duration, shown as P10, Median, P90."
source_ids:
  - iea-electricity-2026
  - acer-retail-2025
  - ec-sdac-15m
  - entsoe-sdac
  - volt-research-content
peer_reviewed: false
---

# How long do negative electricity-price episodes last?

## Abstract

This working paper asks a narrow descriptive question: once a European day-ahead electricity price becomes negative, how long does the contiguous episode remain negative? The public evidence snapshot identifies 1,191 observed episodes and reports a median duration of 1.5 hours. Duration is calculated from the energy represented by each native market interval rather than by treating every database row as an equal unit of time. That choice matters when hourly and quarter-hourly market time units coexist and when daylight-saving transitions make local civil days irregular.

The result describes wholesale day-ahead price curves in ten representative zones over the detailed interval window from 2025-10-01 through 2026-08-29. It is not a household bill estimate, a forecast of the next event, or proof that any appliance can use an entire episode. It also is not trading advice. The corrected evidence intentionally reports no bootstrap interval for this median estimand. The defensible public answer is therefore the recorded point summary, its sample size, its interval-aware construction, the machine-readable distribution values, and the stated coverage limitation.

## Plain-language answer

In this sample, the typical contiguous run of negative day-ahead prices lasted 1.5 hours. “Typical” here means the median across 1,191 episodes, not the average and not a promise about any particular day. The figure now supplies machine-readable P10, Median, and P90 values of 0.25, 1.5, and 11.0 hours, respectively. These describe the plotted episode-duration distribution; they are not a confidence interval around the median.

For a household, duration is only one part of usefulness. A negative wholesale interval may fall outside the time when an electric vehicle is connected, a heat pump has thermal headroom, or a battery can accept more charge. Retail mark-ups, taxes, network charges, export terms, and automation constraints can also separate a wholesale observation from a household outcome. The measurement therefore answers how the market episodes looked, not how much money a household saved.

## Research question

The primary question is: among contiguous sequences of native day-ahead market intervals with prices below zero, what is the observed median episode duration? A sequence ends when the next native interval is not negative or when continuity cannot be established. The estimand is elapsed energy-delivery time, expressed in hours, rather than a count of rows.

This framing deliberately excludes several tempting extensions. It does not estimate why episodes start or end, whether solar or wind is the dominant correlate, whether prices rebound afterward, or whether neighboring zones share the event. Those questions belong to other papers in the same cluster. It also does not convert episode length into bill savings, because no household tariff, device, load profile, charging efficiency, or counterfactual schedule appears in the evidence. Keeping those concepts separate makes this paper’s result both modest and reproducible.

## Data and provenance

The evidence declares three windows. Daily price coverage runs from 2021-01-01 through 2026-08-29. Detailed interval coverage runs from 2025-10-01 through 2026-08-29. A longer history field runs from 2015-01-01 through 2026-08-29. The episode limitation specifically says that detailed metrics use ten representative zones after 2025-10-01, so the 1.5-hour result should be read against that narrower contract rather than presented as a balanced all-Europe estimate since 2015.

The registered source tables are `day_ahead_prices`, `generation_mix`, `border_flows`, `risk_accuracy`, and `zone_holidays`. The duration calculation itself is a price-interval result. Listing the other tables records the shared family data contract; it does not imply that generation, flows, forecast scores, or holidays caused the observed duration. The publication cutoff is 2026-08-30T00:00:00Z, and the latest dates in the declared windows are 2026-08-29. Later revisions or newly ingested zones are outside this frozen public snapshot.

The current public-evidence JSON SHA-256 is `fd14524d8b137c8092bbae79ada2f8761f74b060a2b8f1b52cfcfe2066bbc688`. The evidence records a read-only transaction, a 180-second statement timeout, analysis-code SHA-256 `57c57de79cdab2b5b6d6c54c485cb5162598c5ba0b0bfe995da40d75e6c52ba9`, protocol SHA-256 `adb36bf6b447af9f96339249b8becaefc20422499cca1977242866347a97bd4b`, registry SHA-256 `7bcb91d7476d0a69fe9fa75a5c7782f8117e0153f82f9112b7e1d307d3943717`, snapshot SHA-256 `7e97489fc8528c8cc8c38830e05b48d949ce1f67b98575dff26f5d7c321e4c67`, and source-registry SHA-256 `07949550ac443ff673fda5c0209b99f137544f3ffecf6775f109bb9d09663bd6`. Those identifiers bind this narrative to a specific evidence object, analysis, source registry, and snapshot; they do not by themselves validate the substantive interpretation.

## Method

The method groups adjacent negative-price intervals into episodes. The threshold is a price below zero in the day-ahead series. Adjacency must follow the delivery timeline, and duration is accumulated from each interval’s native length. This avoids the row-counting error in which an hourly row and a quarter-hourly row are each counted as one equal period. It also makes the calculation compatible with daylight-saving changes: elapsed delivery time, not an assumed fixed number of local-clock rows, determines the duration.

After episodes are formed, the analysis takes the median duration across the recorded episode sample. A median is appropriate for a duration distribution that may contain a smaller number of long runs, because it states where the middle observed episode lies without allowing the longest events to dominate the summary. The evidence describes the wider family as episode survival, matched calendar controls, reliability scoring, and network lead-lag analysis. For this paper, however, the reported result is descriptive. No causal coefficient, adjusted treatment effect, or unadjusted significance claim is made.

The evidence states that within-family Holm control applies to inferential claims. That multiplicity rule is important for the cluster as a whole, but it does not transform this median into a hypothesis-test result. No assumptions are registered for this paper. In particular, there is no synthetic household, battery, tariff, or appliance model. Any such scenario discussed as an example would be an external assumption and is intentionally omitted from the empirical answer.

## Results

The machine-readable primary result is a median contiguous negative-price episode duration of 1.5 hours, based on a sample size of 1,191 episodes. The secondary result repeats the episode count as 1,191. The interpretation supplied with the evidence says the duration is energy-weighted from native interval lengths so that daylight-saving time and mixed market time units remain valid.

The corrected evidence stores `bootstrap_95_interval` as null and identifies the interval method as “not reported for this estimand.” That is intentional: no unsupported uncertainty interval is attached to the median. The machine-readable figure instead reports distribution points—P10 0.25 hours, Median 1.5 hours, and P90 11.0 hours—which must not be interpreted as confidence limits.

The result says nothing about the price depth within episodes. A long run could consist of prices just below zero or much more negative values; the duration statistic alone cannot distinguish them. It also says nothing about timing within the day, geographic concentration, or how often an episode occurs. Those are separate dimensions, and importing them from intuition would turn one measured result into several unsupported claims.

## Robustness and placebo checks

The most important construction-level robustness check is native-interval weighting. If the same delivery duration can be represented by different numbers of rows across market regimes, a row-count duration would create a mechanical discontinuity. Accumulating elapsed interval length removes that particular artifact. The explicit daylight-saving interpretation further establishes that irregular local days are not silently normalized to a fixed row count.

The evidence family registers matched calendar controls and within-family Holm control, but this paper’s JSON does not publish a placebo estimate, a matched-control contrast, or a corrected p-value for episode duration. We therefore do not claim those checks produced a particular numerical outcome. A useful replication should compare the native-duration result with a deliberately naive row-count result and stratify by market time unit, but those would be new analyses rather than evidence already measured here.

No bootstrap interval is reported for the median estimand. This is an explicit, intentional uncertainty-handling choice. The figure’s P10 and P90 describe variation among observed episode durations and do not quantify sampling uncertainty in the median.

## Limitations

The central limitation is coverage. Detailed episode metrics use ten representative zones after 2025-10-01. “Representative” does not mean statistically representative of every European bidding zone, and the evidence does not publish a sampling-weight scheme. The result should not be generalized to all zones, all years since 2015, or all regulatory regimes without a wider balanced analysis.

The paper observes day-ahead prices, not retail contracts. Negative wholesale prices may not pass through one-for-one to a household. The analysis does not include taxes, network tariffs, supplier margins, fixed charges, export compensation, or device control. It is therefore not a household bill study. It also does not test whether a customer could foresee an episode, receive a timely signal, or shift consumption safely.

The median compresses a full distribution into one value. Although the evidence now provides machine-readable P10 and P90 duration values of 0.25 and 11.0 hours, it does not show the longest event, tail frequency, seasonal pattern, depth, or zone heterogeneity. Finally, this is a model-assisted working paper and has not been peer reviewed. The hashes improve traceability, but independent methodological review and rerunning the analysis remain necessary.

## Practical implication

A household controller should not assume that a negative-price alert opens an all-day opportunity. The measured middle episode is 1.5 hours, so automation that depends on long uninterrupted windows could miss or only partly use a typical event in this sample. This is a design implication, not a savings estimate. A controller still needs the household’s actual retail price, device availability, minimum run time, state of charge, comfort limits, and a safe fallback when data are stale.

The result also argues for preserving native market resolution end to end. Collapsing a curve to hourly averages can hide short negative intervals or merge them with positive intervals. Conversely, treating each row as an equal duration can exaggerate the importance of quarter-hourly segments. The appropriate operational output is a timestamped interval schedule with explicit start, end, timezone, and freshness—not a generic statement that “tomorrow is negative.”

## Reproducibility

The canonical public evidence is linked in the frontmatter and identifies schema `volt-home-paper-evidence-v1`, series `home-energy-working-papers-v1`, status `measured`, and the frozen publication cutoff. A reproducer should verify the six SHA-256 values, query the declared price snapshot in a read-only transaction, retain each interval’s native start and end, sort on the delivery timeline, group only truly adjacent negative intervals, sum elapsed duration, and compute the sample median.

The reproduction report should separately confirm the 1,191 episode count, the 1.5-hour median, and the figure values 0.25, 1.5, and 11.0 for P10, Median, and P90. It should record how duplicate publications, missing intervals, revisions, timezone conversion, and daylight-saving folds are handled. It should not substitute the broader daily or long-history windows for the detailed episode window. It should preserve the intentional null interval and must not turn the plotted distribution percentiles into uncertainty bounds for the median.

The shared methodology is governed by the [Voltcast Research Content Plan](https://github.com/ossedk/voltcast/blob/main/docs/voltcast/RESEARCH-CONTENT-PLAN.md) from Voltcast. Market context is linked, rather than paraphrased into unsupported findings, through [Electricity 2026](https://www.iea.org/reports/electricity-2026) from the International Energy Agency, [Rewarding flexibility: How retail contract choice can help unlock consumer flexibility](https://www.ceer.eu/wp-content/uploads/2025/11/ACER-CEER-2025-Retail-monitoring.pdf) from ACER and CEER, [EU electricity trading in the day-ahead markets becomes more dynamic](https://energy.ec.europa.eu/news/eu-electricity-trading-day-ahead-markets-becomes-more-dynamic-2025-10-01_en) from the European Commission, and [Single Day-ahead Coupling (SDAC)](https://www.entsoe.eu/network_codes/cacm/implementation/sdac/) from ENTSO-E.

## Disclosure

Analysis and drafting were model-assisted; sources, code, assumptions, and evidence hashes are disclosed. The paper is a public working paper, not peer reviewed research. It reports a wholesale-market duration statistic and is not a household bill study. It does not recommend a tariff, appliance purchase, investment, market position, or trade, and it is not trading advice. No authors, credentials, dates, digital object identifiers, or source findings beyond the supplied registries and evidence have been invented.

## References

- International Energy Agency — [Electricity 2026](https://www.iea.org/reports/electricity-2026)
- ACER and CEER — [Rewarding flexibility: How retail contract choice can help unlock consumer flexibility](https://www.ceer.eu/wp-content/uploads/2025/11/ACER-CEER-2025-Retail-monitoring.pdf)
- European Commission — [EU electricity trading in the day-ahead markets becomes more dynamic](https://energy.ec.europa.eu/news/eu-electricity-trading-day-ahead-markets-becomes-more-dynamic-2025-10-01_en)
- ENTSO-E — [Single Day-ahead Coupling (SDAC)](https://www.entsoe.eu/network_codes/cacm/implementation/sdac/)
- Voltcast — [Voltcast Research Content Plan](https://github.com/ossedk/voltcast/blob/main/docs/voltcast/RESEARCH-CONTENT-PLAN.md)