A house, its meters, and the weather
The UCI Appliances Energy Prediction dataset was assembled for regression modelling of appliance energy use in a low-energy house. It brings together records with different formation histories. According to the source description:
| Record | Source-described formation | Role in this case |
|---|---|---|
| Appliance energy | M-bus meters logged energy every 10 minutes. | Appliances, in Wh. |
| Airport weather | Hourly weather from Chièvres Airport was interpolated and merged by date and time. | T_out, in degrees Celsius. |
| Indoor temperature and humidity | Wireless nodes transmitted about every 3.3 minutes; values were averaged over 10-minute periods. | Background only; not used in this calculation. |
These are the source's descriptions, not measurement chains independently reconstructed for this case. In particular, T_out is not identified here as a thermometer reading taken outside the house at every row's timestamp.
The small step that changes the question
The published table makes a precise operation possible: take Appliances and T_out from each same row and calculate their Pearson correlation.
But consider an additional sentence:
Because the values occur in the same row, they were measured at the same physical instant.
That sentence needs more than a pairing rule. It needs a documented connection between the timestamp label, the energy measurement window, and the weather observations supporting the interpolated value.
This is a test sentence introduced by this project. It is not an allegation that the dataset creators or the original paper made this claim.
What the calculation supports
The calculation uses all 19,735 pairs from one exact archived CSV, keeping the original same-row membership. No values are filled, filtered, resampled, standardised or shifted by this analysis. Source-side interpolation remains part of the input's history.
The supported numerical statement is:
Under the frozen selected bytes and declared same-source-row record pairing, the sample Pearson statistic between Appliances and T_out is r = 0.099154650739960.
Here, “frozen” means the exact file version identified in the methods and evidence, not a claim that the underlying physical process has been fully reconstructed. The displayed decimal places identify the calculation result; they do not express sensor accuracy or physical certainty.
What survives a changed ordering
Reverse the order of the already formed pairs, keeping each energy value with its original weather value. The computed coefficient remains exactly unchanged. The scalar therefore does not retain the source-row chronology. The original rows still retain their order in the archived input; that information is absent from the scalar, not erased from the evidence package.
A different control keeps the energy values fixed and rearranges which weather value is paired with each one. Across 256 deterministic rearrangements, the separate collections of energy and weather values are preserved while the original pair membership is broken. The detailed outputs are retained in the control record. This is a test of pairing dependence, not a population-significance test or a way to rule out confounding.
Neither control tests whether the original measurements were simultaneous.
What remains unresolved
The reviewed records do not establish the timezone, daylight-saving treatment, the start or end anchoring of the energy measurement window, or the precise weather interpolation method and supporting observations. Same-row membership is known; the physical-time relationship needed by the stronger sentence is not.
This does not demonstrate that the timestamps are wrong or that the records are necessarily misaligned. It also does not invalidate the dataset for its original modelling task, whose predictive performance is not evaluated here.
The coefficient alone does not establish causality, a physical response delay, or the complete relationship between weather and appliance use. A numerical physical-lag result was not produced. The claim-by-claim record keeps the supported statements and the project-authored stronger statements that were not supported.
The distinction to carry forward
A calculation can be correct for a published table without settling every physical interpretation of that table. For a question about a shared physical window or a response delay, the next evidence would concern timing conventions, measurement windows, and the original weather-to-row transformation—not another decimal place in the correlation. Even those records would not, on their own, establish causality.
What connects this row's timestamp to the physical interval your question actually needs?
Sources and inspection
- Methods, numerical references and reproduction
- Source manifest and saved-source identities
- Exact result and numerical crosscheck
- Dataset citation: Candanedo, L. (2017), Appliances Energy Prediction, UCI Machine Learning Repository.
UCI declares the dataset to be licensed under Creative Commons Attribution 4.0 International. The source data are unchanged; this case's analysis, controls and wording are separate work. No endorsement by the data providers or paper authors is implied.