Forecast-Only Mode

Forecast-Only Mode

forecast_only=True runs models in operational mode: trains on all available history and predicts the next horizon days from today. No evaluators run, no exporters fire, and report.predictions() returns a clean DataFrame with no “actual” column.

---

Minimum Example

import os
os.environ["THREADS_PER_PROCESS"] = "8"
os.environ["MAX_PROCESSES"] = "2"

from epftoolbox2.pipelines import DataPipeline, ModelPipeline
from epftoolbox2.data.sources import EntsoeSource, OpenMeteoSource, CalendarSource
from epftoolbox2.data.transformers import TimezoneTransformer, ResampleTransformer, LagTransformer
from epftoolbox2.models import OLSModel

ENTSOE_API_KEY = os.environ["ENTSOE_API_KEY"]

if __name__ == "__main__":
    df = (
        DataPipeline()
        .add_source(EntsoeSource(country_code="PL", api_key=ENTSOE_API_KEY, type=["load", "price"]))
        .add_source(OpenMeteoSource(latitude=52.2297, longitude=21.0122, horizon=7, prefix="warsaw"))
        .add_source(CalendarSource(country="PL", holidays="binary", weekday="onehot", daylight=True))
        .add_transformer(TimezoneTransformer(target_tz="Europe/Warsaw"))
        .add_transformer(ResampleTransformer(freq="1h"))
        .add_transformer(LagTransformer(columns=["price"], lags=[1, 2, 7], freq="day"))
        .run(start="now_d-730", end="today_d+7", cache=True)
    )

    report = (
        ModelPipeline()
        .add_model(OLSModel(predictors=["price_d-1", "price_d-2", "price_d-7","warsaw_temperature_2m_d+{horizon}"], training_window=365))
        .run(data=df, target="price", horizon=7, forecast_only=True)
    )

    forecast = report.predictions()
    print(forecast)

Output — 168 rows per model (24 hours × 7 horizons):

run_date (today)	target_date	hour	horizon	prediction	model
2026-04-03	2026-04-04	0	1	312.4	OLS
2026-04-03	2026-04-04	1	1	287.1	OLS
…	…	…	…	…	…
2026-04-03	2026-04-10	23	7	341.8	OLS

---

Date Keywords

Both pipelines accept runtime date keywords so the script runs correctly every day without editing dates:

Keyword	Resolves to
"today" / "now"	Today’s local date
"today_d+7" / "now_d+7"	Today + 7 days
"now_d-735"	Today − 735 days
"2024-01-01"	Passed through unchanged

The offset N in _d±N accepts any integer, so it can match any model’s training_window.

Tip

Set end="today_d+7" in DataPipeline so that OpenMeteoSource and CalendarSource fetch forecast data for the full prediction window. ENTSOE columns will be NaN beyond today, but lag features (price/load from past days) remain valid and are what models use as predictors.

Caution

If test_start is not in the fetched data, BaseModel.run() raises ValueError with a message like: 'test_start' date '2026-04-03' is not in the data (data ends at 2026-04-02). Fetch data through at least '2026-04-03'.

Increase end in DataPipeline (e.g. "today_d+7") to ensure today’s rows exist after all transformers.

---

Horizon Mechanics

With test_start = test_end = "today" and horizon=7, the model runs once from today and produces predictions for:

• today + 1 day (horizon 1) — all 24 hours
• today + 2 days (horizon 2) — all 24 hours
• …
• today + 7 days (horizon 7) — all 24 hours

Total: 24 × 7 = 168 prediction rows per model.

---

YAML Workflow (Daily Automation)

Save once and schedule via cron — no date changes needed:

data_pipeline:
  start: now_d-730    # adjust to cover largest training_window + horizon
  end: today_d+7      # include forecast window for weather/calendar features
  cache: true

model_pipeline:
  target: price
  horizon: 7
  forecast_only: true   # test_start/test_end default to today

from epftoolbox2.pipelines import Workflow

report = Workflow.load("forecast_workflow.yaml").run()
forecast = report.predictions()

See Also

• Model Pipeline Overview — full run() parameter reference
• Workflow — YAML-based orchestration
• examples/05_forecast_only.py