The exact input
The numeric input is energydata_complete.csv from the UCI distribution of Appliances Energy Prediction. It was archived without altering its bytes. The alternative UCI API file named data.csv is not used: it is a different representation, not an interchangeable copy for this computation.
| Item | Value |
|---|---|
| Source rows / paired observations | 19,735 |
| Source columns | 29 |
| First source label | 2016-01-11 17:00:00 |
| Last source label | 2016-05-27 18:00:00 |
| Spacing of successive source labels | Nominal 10 minutes |
| Blank cells in the archived table | 0 |
| Appliance field | Appliances, energy in Wh |
| Weather field | T_out, airport temperature in degrees Celsius, with source-side interpolation |
The row identifier combines its position in the source with its exact date string. Populated cells do not prove uninterrupted acquisition, sensor health, or direct weather measurement at every published label. Source formation is attributed to the saved UCI description, not independently verified at sensor level. Raw weather observations, interpolation code, calibration and physical measurement-window timing are not reconstructed here.
Exact CSV SHA-256:
2820BF712AD0275CB18B85A05250926100D8E65EBB9F4D2D016CA91EA152A25DExact ordered-pair SHA-256:
5AB3AAF96AB76F6324ADCE9930D0C22BEDEB2CA6DBFEB00A0313B1EC5D579EEAThe fingerprint concatenates the source-row number, exact date string, exact Appliances field and exact T_out field, separated by tabs, with a line feed after each row and no header. A hash identifies bytes; it does not establish physical simultaneity.
Computation
Pearson's coefficient is the sum of centred cross-products divided by the square root of the product of the two centred sums of squares. The formula record states the definition and its nonzero-variance requirement.
| Calculation setting or output | Value | Evidence identity |
|---|---|---|
| Primary working arithmetic | 80-digit Decimal | Declared method parameter |
| Issued representation | 15 decimal places, round half to even | Declared representation parameter |
| Issued sample coefficient | 0.099154650739960 | Computed result |
| Crosscheck arithmetic | Centred products with binary64 math.fsum | Different numerical implementation, not independent field evidence |
| Absolute crosscheck tolerance | 1e-12 | Declared numerical tolerance, not a measurement uncertainty |
| Pairing-control trials | 256 | Declared control parameter |
No input normalisation, imputation, filtering, resampling, clipping or lagging is performed. Pearson's internal centring and scaling belong to the formula, not to a newly normalised source file. Numeric signs are preserved; this symmetric statistic does not introduce a directional energy-flow convention.
The invocation record specifies parser rules, precision, rounding and the fixed permutation generator and seed. The summary retains the working coefficient, crosscheck, difference and component sums.
Controls and limits
Common reversal changes the order of complete pairs, not their membership; the Decimal coefficient is exactly preserved. Independent weather permutations preserve marginal values while breaking the original pairing. The control outputs retain every trial. Exchangeability, a population sampling model and causal adjustment have not been established, so these trials are not reported as p-values or significance evidence. No physical-lag coefficient is computed.
The verdict record separates table-level results from unsupported physical-time, causal, full-relationship and significance interpretations. Its stronger test sentences were written for this analysis, not extracted as claims by the original authors.
Reproduce from the evidence package
Extract the accompanying evidence ZIP into a directory and open a terminal at its root. Python with its standard library is sufficient; no account, API key, network connection, other project, Git checkout or third-party package is needed by these commands. Inspect scripts before executing a downloaded package.
python -X utf8 scripts/check_pearson_publication.py --verify
python -X utf8 cases/pearson_appliances_tout/scripts/recompute.py --no-writeThe first command checks the packaged file hashes and registered numerical-text bindings. The second uses the original calculator to recompute the statistic, controls and claim verdicts, compares deterministic outputs byte for byte, and checks the saved receipt's output hashes and declared boundaries. Neither command overwrites the frozen outputs. A mismatch ends the check with a nonzero exit code; do not regenerate old outputs just to obtain a pass.
The package is a bounded calculation companion, not a full export of project history. Its historical contracts retain their original internal and pre-computation status fields. The later computation receipt records the completed analysis. Neither historical status is a website release decision.
Attribution and corrections
Dataset: Candanedo, L. (2017). Appliances Energy Prediction. UCI Machine Learning Repository. Dataset DOI. The saved licence and citation evidence records UCI's CC BY 4.0 declaration and the source's statement about permission to distribute the embedded weather data. The CSV is unchanged; derived statistics and controls are identified as this project's work. No endorsement is implied.
Related paper: Candanedo, Feldheim and Deramaix, Data driven prediction models of energy use of appliances in a low-energy house. Paper DOI. Only bibliographic metadata are included; the paper itself is not redistributed and its model performance is not reproduced or evaluated in this case.
For a correction, identify the English edition, section or sentence, the relevant source file/field, and the evidence supporting a different reading. A useful challenge may show either an overclaim or that a narrower statement was withheld unnecessarily. New timing records would require a separately documented interpretation; they should not silently replace this edition's source basis.