Predictor Presets

epftoolbox2.presets returns the inputs of two standard benchmark models as predictor lists. The lists are plain strings with a {horizon} template, so they can be saved in a model pipeline YAML.

from epftoolbox2.presets import expert_predictors, lear_predictors, profile_columns, WEEKDAYS

lear = lear_predictors(exogenous=["load_forecast", "residual_load"])  # 247 inputs
expert = expert_predictors(exogenous=["load_forecast"])

Function	Parameters (defaults)	Returns
lear_predictors	exogenous (required), price="price", freq="1h", calendar_prefix=""	LEAR inputs (Lago et al. 2021)
expert_predictors	exogenous=("load_forecast",), price="price", weekdays=WEEKDAYS, calendar_prefix=""	Expert-model inputs (Ziel & Weron 2018)
profile_columns	column, freq="1h"	The names DailyProfileTransformer creates, e.g. price_h00 … price_h23

From “Day d” to the Issue Day

Both are day-ahead models. The papers index inputs by the day being forecast, d. This library indexes a row by the day D the forecast is issued on, and horizon h targets D+h. For h = 1 (d = D+1):

Paper	Column on the row of D	Built by
p(d-1, ·)	price, price_h00 … price_h23	DailyProfileTransformer
p(d-2, ·)	price_d-1, price_h05_d-1	LagTransformer(lags=[1], freq="day")
p(d-3, ·)	price_h05_d-2	LagTransformer(lags=[2], freq="day")
p(d-7, ·)	price_d-6, price_h05_d-6	LagTransformer(lags=[6], freq="day")
min, max, last price of d-1	price_day_min, price_day_max, price_day_last	DailyStatsTransformer
x(d, ·)	x_d+{horizon}, x_h05_d+{horizon}	LagTransformer(lags=[-1, ..., -H], freq="day")
x(d-1, ·)	x, x_h05	DailyProfileTransformer
x(d-7, ·)	x_h05_d-6	LagTransformer(lags=[6], freq="day")
weekday of d	is_monday_d+{horizon}, …	CalendarSource(weekday="onehot") + LagTransformer(lags=[-1, ..., -H])

Caution

What the presets assume is known when the forecast is issued on D:

• D’s day-ahead prices (price, price_h05), published on D-1 around 13:00. The model’s training targets still end at D-1, as always.
• The exogenous series for D+1. Use day-ahead forecasts, never realised values. ENTSO-E’s load forecast for D+1 is published around 10:00 on D and the wind and solar forecast possibly as late as 18:00, so it depends on your issue time.
• For h > 1 the price inputs stay anchored at the issue day, which gives the latest prices available. x_d+{horizon} must then really be known h days ahead, and ENTSO-E’s day-ahead forecasts are not. Run the presets with horizon=1, or give them exogenous series that are published far enough ahead.

LEAR

lear_predictors lists all slot prices of d-1, d-2, d-3 and d-7, all slot values of each exogenous series for d, d-1 and d-7, and seven weekday dummies of d. With hourly data and two exogenous series that is 96 + 144 + 7 = 247 inputs, as in the paper. Every hour’s model gets the same inputs.

from epftoolbox2.pipelines import DataPipeline, ModelPipeline
from epftoolbox2.data.sources import EntsoeSource, CalendarSource
from epftoolbox2.data.transformers import (
    TimezoneTransformer, ResampleTransformer, DSTNormalizeTransformer, LambdaTransformer,
    DailyProfileTransformer, LagTransformer,
)
from epftoolbox2.models import LassoCVModel
from epftoolbox2.presets import WEEKDAYS, lear_predictors, profile_columns

EXOGENOUS = ["load_forecast", "residual_load"]
H = 1  # day-ahead

data = (
    DataPipeline(timezone="Europe/Berlin")
    .add_source(EntsoeSource("DE_LU", api_key="YOUR_KEY", type=["load", "generation", "price"]))
    .add_source(CalendarSource("DE", weekday="onehot", holidays=False))
    .add_transformer(TimezoneTransformer())
    .add_transformer(ResampleTransformer(freq="1h"))
    .add_transformer(DSTNormalizeTransformer(freq="1h"))  # 24 slots a day, so every profile is complete
    .add_transformer(LambdaTransformer(
        columns={"residual_load": "load_forecast - generation_wind_onshore_forecast - generation_solar_forecast"}
    ))
    .add_transformer(DailyProfileTransformer(columns=["price", *EXOGENOUS]))
    .add_transformer(LagTransformer(columns=profile_columns("price"), lags=[1, 2, 6], freq="day"))
    .add_transformer(
        LagTransformer(
            columns=[c for x in EXOGENOUS for c in profile_columns(x)],
            lags=[*range(-1, -H - 1, -1), 6],
            freq="day",
        )
    )
    .add_transformer(LagTransformer(columns=[f"is_{d}" for d in WEEKDAYS], lags=list(range(-1, -H - 1, -1)), freq="day"))
)
df = data.run("2022-01-01", "2024-12-31", cache=True)

report = (
    ModelPipeline()
    .add_model(LassoCVModel(lear_predictors(exogenous=EXOGENOUS), training_window=728, name="LEAR"))
    .run(df, test_start="2024-01-01", test_end="2024-12-30", target="price", horizon=H)
)

Expert Model

expert_predictors lists, for each hour’s model: the same hour’s price on d-1, d-2 and d-7, the minimum, maximum and last price of d-1, each exogenous series for d, and weekday dummies of d.

Ziel & Weron use an intercept plus Monday, Saturday and Sunday dummies. OLSModel fits without an intercept and takes the level from a full set of one-hot dummies, so the default is all seven weekdays. This spans the intercept and the three dummies. Pass weekdays=("monday", "saturday", "sunday") for the original set with a model that fits an intercept, such as LassoCVModel.

from epftoolbox2.data.transformers import DailyStatsTransformer
from epftoolbox2.models import OLSModel
from epftoolbox2.presets import expert_predictors

# after TimezoneTransformer, ResampleTransformer and DSTNormalizeTransformer as above:
steps = [
    LagTransformer(columns="price", lags=[1, 6], freq="day"),
    DailyStatsTransformer(columns="price", stats=["min", "max", "last"]),
    LagTransformer(columns=["load_forecast"], lags=[-1], freq="day"),
    LagTransformer(columns=[f"is_{d}" for d in WEEKDAYS], lags=[-1], freq="day"),
]
model = OLSModel(expert_predictors(exogenous=["load_forecast"]), training_window=365, name="Expert")

Checking the Data

Every name a preset returns, after replacing {horizon}, must be a column of the data frame:

predictors = lear_predictors(exogenous=EXOGENOUS)
missing = [p.replace("{horizon}", str(h)) for h in range(1, H + 1) for p in predictors]
missing = [c for c in missing if c not in df.columns]
assert not missing, missing[:10]