# Using ENSOFORECAST in research

This review edition displays 18 source-derived spatial slices and documents coverage across all nine catalogued core files. It is a preservation and interpretation tool. It has not established a forecast-skill result or a complete historical issuance record. Original-author and peer review of the new edition have not been recorded.

## Objects and contributions

Keep five things distinct: the original Zebiak–Cane model; the LDEO1/LDEO2 products; IRI's custody and catalog documentation; Aaron Kaplan's archive transfer; and Vultax's preservation, decoding and interface. Those are different contributions. Do not assign an ORCID, affiliation, model authorship or endorsement by name matching alone.

The displayed maps use three-decimal rounding from the source float fields. The original binary files remain unchanged in the privately preserved transfer. Source hashes, member names and display transformations let another reader trace the maps. The 40 additional transfer files remain outside the documented core edition.

## Time, observations and interpretation

- Source `T` denotes verification/target time. Half-integer monthtime values map to months since January 1960. Keep the numeric source axis when converting; do not fabricate day-level dates or an issue timestamp.
- The source describes a composite of six consecutive monthly initializations, followed by spatial post-processing. LDEO's 1996–1997 bulletin articles count the nominal lead from the latest of the six starts (ledger E3), but the archive does not state its convention. Subtracting a nominal lead from a target month does not recover a verified issuance date or the exact initialization window.
- The 1982 model fields are retrospective by an explicit inference from the [Cane–Lee model chronology](https://ocp.ldeo.columbia.edu/res/div/ocp/pub/cane/CaneLee.2013.pdf), which places the first future predictions in early 1986. The exact production date remains unknown. The 1997 example fields have unverified issue-time evidence.
- Hindcasts are legitimate evaluation material. Their evidential meaning depends on how training, tuning and post-processing were separated from evaluation. This inspection has not established that separation and makes no finding of improper use of future information.
- Comparisons use the archive's observation file, not a newly assembled independent observation product. Its anomaly reference period, revisions and the relationship to modern reference datasets need further documentation.
- LDEO1 and LDEO2 remain the archive's product names. Zebiak and Cane's December 1997 bulletin calls the improved initialization of Chen et al. (1995) "LDEO2" (evidence ledger E4); the archive's own documentation does not confirm the correspondence. Do not infer a better model from the numbering.
- Maps are continuous sea-surface temperature anomalies in degrees Celsius, not calibrated event probabilities. The two-row equatorial profile is not an official Niño index. An apparent field maximum is not a global ranking of El Niño event strength.

## Full-core coverage now checked

`data/coverage.json` was generated from every source time coordinate and valid-cell mask in all nine files. Time axes are monthly, contiguous and duplicate-free. All nine have 316 target months in common. Individual model/lead products overlap the supplied observations for 316–338 target months. The same 1,077 cells are valid in every month of every file.

These are coverage results. They do not establish common historical availability, statistical independence, out-of-sample evaluation, or forecast skill. Different target ranges can be legitimate; do not score products on unequal samples and attribute the difference solely to the models.

Run `python3 validate.py` to check local slice dimensions, units/coordinates, timing-evidence labels, masks, coverage consistency and file hashes. With the original transfer available:

```bash
python3 validate.py --source-tgz /path/to/ENSOFORECAST.tgz
```

The second command verifies the transfer and nine core file hashes, recomputes every time label and valid-cell count, checks common forecast/observation masks, and compares all 18 preview slices with the original floats at the documented rounding. It reads the archive without extracting files or executing source programs. The validator reports source comparison as not run when no transfer is supplied.

## The worthwhile next research

1. Compare archived fields with the dated forecast figures in the 1996–1997 *Experimental Long-Lead Forecast Bulletin* issues listed in ledger E7, and look for further issue receipts and product documentation. Record evidence per product and time period; unknowns stay visible. File modification times alone are not issue receipts.
2. Resolve anomaly reference periods and the definition of the observation product before comparing with external datasets. Preserve those external versions separately.
3. Prespecify an evaluation on matched dates and masks: area weights, region, leads, season, continuous metrics and a fair persistence/climatology baseline. Baselines must use only information allowed at each initialization; they cannot borrow later observations or a whole-sample fitted climatology without being labelled retrospective.
4. Account for serial dependence and overlapping forecast horizons. Treat two famous episodes as illustrations, not the population. Show uncertainty and sensitivity rather than an unqualified model leaderboard.
5. Publish a reproducible conversion recipe with source axis and metadata retained. Validate modern-format claims instead of merely assigning a CF-compliance label.

## Maintaining a durable edition

Deposit the documented, rights-checked core with citation, versioned metadata, immutable source hashes, conversion code and validation outputs in an independent scholarly repository after review. Archive the readable documentation and static article as well as the interactive frontend. Version corrections and keep a clear successor/location record if the Vultax host changes.

The [FAIR principles](https://www.nature.com/articles/sdata201618) call for reusable metadata and provenance, including metadata that can survive the data's unavailability. They are a useful design target; these local files do not establish a completed FAIR assessment, repository deposit or indefinite preservation.