What the project is, why it exists, where it stands, why it pauses here, and what comes next.

2 · Why does the project exist?

A forecasting system can look impressive while answering a weak question. Electricity has strong seasonality; prices change regime; renewable output can ramp sharply; random time splitting can leak future information; and an over-large experiment can become too expensive to finish.

The repository therefore follows one rule: build research controls before opening final evidence. Data contracts, leakage-resistant windows, baselines, uncertainty tools, tests, a compute-safe smoke path and publication structure are created first. The final holdout comes later.

3 · What goes into the project?

The software targets official EPİAŞ Transparency Platform REST services. Credentials stay in environment variables and never enter Git history. Raw third-party records are not redistributed by default; reproducibility is preserved through retrieval instructions, source identifiers, manifests and hashes.

4 · How does it work step by step?

5 · What model exists today?

The v0.1 model is intentionally compact. Hourly multivariate inputs are projected into a shared representation, receive positional encoding, pass through a Transformer encoder and feed multi-horizon heads. Each output supplies a central forecast and positive uncertainty scale. Split-conformal utilities can then calibrate empirical prediction intervals on a separate calibration period.

6 · What has been completed, exactly?

7 · Why has the project paused here?

It has not paused because the code failed. Software construction and empirical inference are different phases. If years of real data are inspected first, models are changed after future-period scores are seen, and the hypothesis is written only afterward, a later “confirmatory” claim becomes much weaker.

Research scaffold
      ↓
Real-data audit
      ↓
Exploratory model + runtime study
      ↓
Protocol freeze / optional OSF preregistration
      ↓
Untouched confirmatory holdout
      ↓
Evidence archive + paper

Before freezing the experiment we need to know whether data coverage is stable, whether timestamp or schema changes exist, which baselines are credible and how much compute the complete model × seed × horizon plan costs. The pause protects the scientific value of the next stage.

8 · What does the project deliver today?

It does not currently deliver an operational market forecast, trading signal, dispatch instruction or validated claim that the shared Transformer is superior.

9 · What can fool this kind of study?

10 · What would count as scientific success?

A null or negative finding is not a failed project if the experiment is controlled. The research object is the hypothesis, not the model's reputation.

11 · What can be added from here?

12 · The strongest eventual paper

When Does Cross-Task Learning Help Electricity Forecasting? A Reproducible Multi-Task Study of Load, Renewable Generation and Market Price in Türkiye.

The strongest contribution is not a marketing claim that a “foundation model” wins. It is a controlled answer about transfer, horizon dependence, market regimes and calibrated uncertainty, backed by preserved evidence that another researcher can reproduce or challenge.

13 · Research boundary

This repository is research software. It is not an EPİAŞ product, grid-operator tool, production forecasting service, trading recommendation or dispatch controller. Until a frozen real-data study is completed, it should be read as a research platform and experimental design.