LoRA adapter for 24-hour-ahead hourly day-ahead (DA) electricity price forecasting in Poland (ENTSO-E bidding zone PL). Fine-tuned on top of amazon/chronos-2 using ENTSO-E price history. Weather covariates were not used during training but can be supplied at inference.

Intended use

Univariate DA price forecasting for the Polish bidding zone (PL)
Load with the Chronos-2 pipeline (PEFT LoRA adapter, not a standalone full model)
Optional Open-Meteo weather covariates at inference via predict_df
Research and prototyping — not financial or trading advice

Usage

Install dependencies:

pip install chronos-forecasting peft torch

Load the adapter (Chronos merges the LoRA weights automatically):

python

from chronos import Chronos2Pipeline
import numpy as np

pipeline = Chronos2Pipeline.from_pretrained("albertooooz/chronos-2-lora-pl-day-ahead")

# Univariate context: shape (n_series, n_variates, history_length)
context = np.array(your_hourly_prices, dtype=np.float64).reshape(1, 1, -1)
forecast = pipeline.predict(inputs=context, prediction_length=24)

Training data

Field	Value
Source	ENTSO-E Transparency Platform (day-ahead prices)
Zone	PL
Frequency	1h (resampled from 15min where needed)
Train rows	85,523
Date range	2015-01-04 → 2025-07-31
Holdout	Last 12 months excluded from training (benchmark eval window)
Weather at train	none

Training hyperparameters

Parameter	Value
Fine-tuning mode	lora
LoRA rank (r)	8
LoRA alpha	16
Learning rate	1e-05
Steps	1000
Batch size	64
Context length	2048
Prediction length	24
Hardware	RunPod CUDA (PyTorch 2.4.1+cu124)

Target modules: self_attention.q/k/v/o, output_patch_embedding.output_layer.

Evaluation

Walk-forward backtest on the held-out 12 months: 2 random origin days per calendar month, horizon 24h, random_seed=42, 1h frequency — same protocol as benchmark v1.

Model	Covariates	Zeroshot MASE	LoRA MASE	Δ MASE
chronos-2-small	actual_weather	0.814	0.609	+25.1%
chronos-2-small	none	0.843	0.632	+25.0%
chronos-2	actual_weather	0.779	0.578	+25.8%
chronos-2	none	0.824	0.614	+25.5%

Metrics: MASE (primary), MAE. MASE scale uses in-sample seasonal naive on context up to each origin.

See eval_results.json in this repository for structured metrics.

Limitations

Trained only on Polish (PL) day-ahead prices; do not expect strong transfer to other zones or markets without retraining
Does not model fundamental drivers (fuel, outages, cross-border flows) beyond price history (and optional weather at inference)
Outputs are probabilistic point/quantile forecasts — validate before production use
Not a substitute for professional energy market forecasting or risk management

Base model

This is a LoRA adapter — load it together with amazon/chronos-2. If you use this adapter in research, cite the Chronos-2 foundation model.

Model card contact

albertooooz