---
id: VOLT-HOME-WP-058
title: "Are weekends and holidays intrinsically harder to forecast?"
slug: are-weekends-and-holidays-intrinsically-harder-to-forecast
description: "A serving-lineage-matched weekend-versus-weekday MAE comparison that finds lower weekend error while leaving national-holiday effects unresolved."
published: 2026-08-30
cluster: "Forecast verification and uncertainty"
status: measured
evidence: /research-data/home-papers/are-weekends-and-holidays-intrinsically-harder-to-forecast.json
figure: /research-media/home-papers/are-weekends-and-holidays-intrinsically-harder-to-forecast.webp
figure_alt: "Chart for Are weekends and holidays intrinsically harder to forecast?: weekend minus weekday lineage-matched MAE, shown as weekday, weekend."
source_ids:
  - volt-promotion-policy
  - volt-v61-card
  - volt-research-content
  - google-article
  - entsoe-sdac
peer_reviewed: false
---

## Abstract

Weekends are not harder under the frozen fleet-level MAE comparison. Weekend lineage-matched MAE minus weekday lineage-matched MAE is -15.313262239478547 EUR/MWh, based on 4546 weekend error observations. The negative sign means average weekend error is lower than average weekday error in the recorded sample.

The evidence does not isolate holidays. Dates are divided only by weekday number into weekend and weekday groups; no national, regional, bridge-day, or substitute-holiday calendar enters the implemented outcome. The word “intrinsically” is also stronger than the design permits. Calendar groups differ in demand, generation, volatility, season, zone composition, horizons, and serving versions. The result is a descriptive operational contrast, not an inherent property of weekends and not a holiday effect. Only rows exactly matched to immutable serving lineage count as live-service evidence.

## Plain-language answer

For weekends, the measured answer is no. Weekend forecast MAE is 15.313262239478547 EUR/MWh lower than weekday MAE in the direction of the reported weekend-minus-weekday contrast.

For holidays, the evidence gives no separate answer. A holiday that falls on a weekday remains in the weekday group, and a holiday on a weekend remains in the weekend group. Country calendars also differ, so a pan-European holiday label must be joined by zone and date. The safest summary is: weekends were easier on average in this serving record, while holiday difficulty was not measured.

## Research question

The preregistered question asks whether weekends and holidays are intrinsically harder to forecast. A strong answer would distinguish calendar effects from the market conditions correlated with those calendars. It would compare the same zone, horizon, model objective, and season, and it would identify national holidays using the correct local calendar.

The implemented estimand is simpler: mean lineage-matched MAE on dates whose weekday is Saturday or Sunday minus mean MAE on Monday through Friday. It measures the operated forecasting service, not a structural law. ENTSO-E's [Single Day-ahead Coupling (SDAC)](https://www.entsoe.eu/network_codes/cacm/implementation/sdac/) describes a coupled European day-ahead market, but market coupling does not remove country-specific work patterns or holiday calendars. Those are empirical strata that this paper does not have.

## Data and provenance

The frozen evidence JSON is linked in the frontmatter and has publication cutoff `2026-08-30T00:00:00Z`. The family declares `forecast_accuracy`, `forecast_serving_lineage`, `risk_accuracy`, `generation_mix`, and `zone_temp_weighted`. The calendar outcome uses the first two. Eligible accuracy rows have non-null MAE and an exact serving-lineage match on zone, delivery date, horizon bucket, and model version.

The aggregate was extracted from production through a SELECT-only, read-only transaction with a 180-second statement timeout. Snapshot SHA-256 is `7e97489fc8528c8cc8c38830e05b48d949ce1f67b98575dff26f5d7c321e4c67`. Analysis-code, protocol, registry, and source-registry hashes are `57c57de79cdab2b5b6d6c54c485cb5162598c5ba0b0bfe995da40d75e6c52ba9`, `adb36bf6b447af9f96339249b8becaefc20422499cca1977242866347a97bd4b`, `7bcb91d7476d0a69fe9fa75a5c7782f8117e0153f82f9112b7e1d307d3943717`, and `07949550ac443ff673fda5c0209b99f137544f3ffecf6775f109bb9d09663bd6`. The evidence and figure hashes are `d87ca35af3533a70a79948c534efc6101786ca0c89f937dbf0315bccc2fa2bf8` and `b85abd877b837f46259a0e87a6eae51f273a68d8acd1b6134c41b3d32e2afaf2`.

## Method

The analysis parses the date of every eligible accuracy row. Saturday and Sunday rows enter the weekend list; all other rows enter the weekday list. It calculates the arithmetic mean MAE in each list and subtracts the weekday mean from the weekend mean. The sample field reports the 4546 weekend MAE observations used as the primary outcome sample.

The method retains whatever version was genuinely served. It does not select a special weekend model after outcomes are observed. The [Objective-specific seven-day forecast promotion](https://github.com/ossedk/voltcast/blob/main/docs/voltcast/OBJECTIVE-PROMOTION-POLICY.md) keeps point, quantile, external-challenger, and production-main identities separate, so an objective-specific version cannot silently substitute for another. The [Volt 6.1 cross-zone price model card](https://github.com/ossedk/voltcast/blob/main/docs/voltcast/V6.1-MODEL-CARD.md) documents a failed development model that never served; its offline weekend behavior, if any, is irrelevant to this live-service contrast.

## Results

Weekend minus weekday lineage-matched MAE is -15.313262239478547 EUR/MWh. The figure reports weekday MAE of 91.0632991290518 EUR/MWh and weekend MAE of 75.75003688957325 EUR/MWh, whose difference reproduces the headline. The negative result means weekend rows have lower mean absolute error than weekday rows. The weekend sample contains 4546 observations.

The regenerated evidence records `bootstrap_95_interval` as null and the interval method as “not reported for this estimand.” No uncertainty interval accompanies the -15.313262239478547 contrast, so it cannot establish significance for the difference.

There is no holiday coefficient, holiday sample size, or country-calendar result. The paper does not infer one. Likewise, the evidence does not report separate zone, horizon, season, or model-version calendar effects.

## Robustness and placebo checks

Serving lineage is the strongest implemented control. Each error is assigned to the exact model recorded for that zone, date, and horizon bucket, preventing retrospective calendar-specific model choice. The weekend rule is deterministic and familiar, so no cutoff was selected after looking at MAE.

Important placebos are absent. There is no comparison with adjacent Fridays and Mondays, no shuffled weekday labels within month, no same-zone and same-horizon matching, and no bridge-day analysis. National holidays are not joined at all. A useful placebo would assign false holiday dates within the same country-season and verify that any measured holiday effect exceeds those random contrasts.

Another useful check would preserve each zone's number of weekend rows while permuting weekend labels within narrow seasonal blocks. That would test whether the pooled difference exceeds ordinary calendar composition without treating dependent horizons as independent. A second check would compare only zone-dates for which the same serving version has both nearby weekday and weekend observations. Neither check is part of the frozen evidence, so the negative headline difference remains an unadjusted descriptive contrast.

The registered family calls for blocked uncertainty intervals for inferential claims, but this descriptive contrast reports none. Weekly dependence and uncertainty for the contrast are therefore not quantified. No Holm-adjusted inferential claim is made.

## Limitations

Weekend and weekday samples can have different zone, horizon, and version composition. If some zones have more matched weekend rows or if longer horizons are unevenly distributed, the pooled difference can reflect composition. MAE is also scale-dependent across markets.

Calendar effects are entangled with demand and generation regimes. Weekends may have lower demand, different industrial load, different maintenance patterns, and different negative-price incidence. Those conditions can affect both realized prices and forecastability. A mean difference does not identify which channel matters.

The use of the row's date assumes it is the correct local delivery-day calendar identity for all zones. The public aggregate does not audit local-versus-UTC seams in this outcome. Most importantly, the title includes holidays, but the implemented contract explicitly states that holidays are not isolated. Any claim that “holidays are easier” or “holidays are harder” would be invented.

Weekend composition also changes around public holidays. A Monday holiday can affect Sunday demand and market expectations even though the outcome labels Sunday only as a weekend and Monday only as a weekday. Bridge days and school holidays can alter load without appearing in the binary split. These interactions mean the weekend coefficient cannot be interpreted as the effect of ordinary Saturday and Sunday behavior alone.

The figure now publishes both group means, but the JSON still does not expose the weekday sample size or the underlying error distribution. Readers can verify the displayed 91.0632991290518 versus 75.75003688957325 EUR/MWh difference, but cannot inspect whether a few high-error weekdays drive it. The absence of an uncertainty interval does not close that gap.

Finally, lower average MAE does not imply better probabilistic calibration. A weekend median forecast could be closer while its uncertainty bands remain too narrow. The customer-serving objective requires point and distributional evidence to remain distinct, and this paper reports only the MAE contrast.

## Practical implication

Users should not automatically distrust weekend forecasts. In this frozen serving record, they have lower average MAE. But a household controller should still evaluate its own zone, horizon, issue freshness, and uncertainty; a fleet average cannot guarantee one weekend curve.

Model operators should add a country-aware holiday stratum before making holiday claims. Weekend monitoring can be useful, but promotion should remain objective- and horizon-specific. A model that performs well on weekends cannot replace a full customer contract if it fails weekdays or probabilistic coverage. The [Voltcast Research Content Plan](https://github.com/ossedk/voltcast/blob/main/docs/voltcast/RESEARCH-CONTENT-PLAN.md) requires the favorable weekend result and the unresolved holiday question to appear together.

For public communication, the calendar label should be carried beside the score rather than used to filter out difficult days. Operational scorecards are most credible when weekends, weekdays, missing rows, and version changes remain visible under the same lineage rules. A holiday-specific product claim should wait for the missing country-calendar contract and prospective evidence.

## Reproducibility

Verify the evidence file, manifest identity, and snapshot, code, protocol, and registry hashes. Join forecast accuracy to serving lineage on zone, delivery date, horizon bucket, and exact model version. Retain non-null MAE. Parse dates, classify Saturday and Sunday as weekend, calculate group means, and subtract weekday mean from weekend mean.

A stronger extension should use zone-specific local dates, join an official country holiday calendar, identify substitute and bridge days, balance zone-horizon-version cells, and block-resample complete weeks while recalculating the contrast. It should be a newly registered result. Google Search Central's assigned [Article structured data](https://developers.google.com/search/docs/appearance/structured-data/article) guidance supports stable public metadata and evidence linkage; it does not fill the missing holiday contract.

## Disclosure

Analysis and drafting were model-assisted. The language model organized the frozen calendar comparison and preserved the fact that national holidays were not measured. It supplied no empirical value. Sources, code, assumptions, limitations, and evidence hashes are disclosed. This working paper is not peer reviewed.

Volt has zero live traders and zero live capital. C0R is the only paper strategy. Production forecasting is a non-trading service. This calendar analysis is not trading advice, financial advice, a retail-bill forecast, or authorization to promote a model.

## References

1. Voltcast. “Objective-specific seven-day forecast promotion.” https://github.com/ossedk/voltcast/blob/main/docs/voltcast/OBJECTIVE-PROMOTION-POLICY.md
2. Voltcast. “Volt 6.1 cross-zone price model card.” https://github.com/ossedk/voltcast/blob/main/docs/voltcast/V6.1-MODEL-CARD.md
3. Voltcast. “Voltcast Research Content Plan.” https://github.com/ossedk/voltcast/blob/main/docs/voltcast/RESEARCH-CONTENT-PLAN.md
4. Google Search Central. “Article structured data.” https://developers.google.com/search/docs/appearance/structured-data/article
5. ENTSO-E. “Single Day-ahead Coupling (SDAC).” https://www.entsoe.eu/network_codes/cacm/implementation/sdac/