Zeno — forecast combination research preview (eco-u0-20260915)
Current highlighted Zeno artifact. This repository publishes one neural checkpoint that combines several forecasters into one forecast, together with the simple recent-performance baseline it has to beat.
The earlier scoped v0.15.0 release remains unchanged as programme history on
the main repository; pin revision="v0.15.0" there when reproducing it.
Main repository and release history: mbarbosa1/zeno-divergent-v1 · Research site: https://zenodivergent.dev
Research question: when does the neural model justify its additional complexity over a simple recent-performance rule?
What it does
Given several forecasts for the same case, plus each forecaster's own recent error history, the model produces learned weights and one combined forecast. It does not produce reliability warnings, abstentions or trading signals.
Three-way comparison — held-back confirmation cases, native units
Mean squared error; lower is better.
| task | equal average | recent-performance rule | Zeno preview | vs recent-performance | support |
|---|---|---|---|---|---|
| air surface pressure (hPa²) | 1.287 | 0.6142 | 0.5995 | 2.4% lower — beyond twice the day-clustered error | 2,785 cases / 70 days |
| air temperature (°C²) | 2.112 | 1.830 | 1.762 | 3.7% lower — beyond twice the day-clustered error | 4,717 / 39 |
| air precipitation (mm²) | 0.7148 | 0.7439 | 0.6897 | 7.3% lower — beyond twice the day-clustered error | 2,866 / 73 |
| air wind speed (m²/s²) | 8.275 | 7.131 | 7.123 | 0.14% lower — not distinguishable | 3,963 / 34 |
| geomagnetic Kp | 1.278 | 1.258 | 1.242 | 1.2% lower — not distinguishable | 3,364 / 35 |
The often-quoted reductions against equal averaging (53.4% on surface pressure, 13.9% on wind) remain valid, but they are against the weaker reference. The recent-performance rule reproduces most of that gain. That says nothing about what the network learned internally; it says the simple rule is a strong comparator.
Losses are kept: on sea surface pressure and sea temperature the preview is worse than both references, and those cells are thin (603 / 4 days and 16 / 4).
Limits
- Post-hoc audit of an already-inspected confirmation partition; comparator settings were chosen on development data before any confirmation loss was computed. Not a fresh sealed test.
- Histories are replayed under assumed historical availability — retrospective, not prospective evidence.
- One checkpoint, one seed. Seed robustness of these margins is unresolved.
- Broader pretraining (U) was worse than ecosystem-only pretraining (M) on the supported air tasks; this preview is the U checkpoint, and that negative result stands.
- Fitted static weights were checked against equal weighting on the training objective and never worse; they nonetheless generalise worse and are not part of this comparison.
Run both on the same inputs
pip install -r requirements.txt
python3 predict_example.py
The package is self-contained: checkpoint, model code, the baseline
(recent_performance.py), a confirmation-split sample of cases, the full
results, and MANIFEST.json with SHA-256 hashes of every file. Loading needs no
network access. Inputs carry explicit masks; outputs are converted to each
task's own units by the example. Inference reproduces the archived run's
predictions on the same cases.
- Parameters: 12,758,404
- Architecture: 4 temporal layers and 3 model-set comparison layers
- Task mapping:
task_vocab_v0, the exact mapping the checkpoint was trained with, persisted in the checkpoint and restored on load. - Licence: Apache-2.0.