OPB: Week-ahead forecasting of hourly electricity prices in the Brazilian market — data, code and stored forecasts

This repository contains the database, source code, stored forecasts and reproduction scripts supporting the manuscript "Week-ahead forecasting of hourly electricity prices in the Brazilian market combining attentional recurrent neural networks and regularized autoregression" (submitted to the International Journal of Forecasting). The underlying research was developed in the first author's doctoral thesis (University of São Paulo, 2023). Contents forecasts/ Stored week-ahead forecasts for 2021 (the core input of the reproducibility check) df_forecasts.csv 8,568 hourly rows x 11 models + realized PLD. The OPB ensemble columns are DERIVED: OPB = hourly mean of "ANN Média" and "LEAR"; OPB-best/worst use the best/worst single network instead of the network mean. df_errors.csv Corresponding hourly errors. df_metrics*.csv Canonical accuracy tables of the original study (full year, in-crisis, out-of-crisis). df_risk_returns_price20_load5.csv Hourly acquisition-cost series, best protection scenario (Section 7). reproduction/ Scripts that regenerate every table of the paper in minutes (Python 3, numpy/pandas) regenerate_tables.py Tables 1-5 from the stored forecasts. dm_test.py Diebold-Mariano tests with HAC variance (Bartlett kernel). database/ Full relational database (SQLite, ~1.6 GB) built by module M1 from public sources code/ Source code: database module, forecaster notebooks (autoencoder + seq2seq v0.3), benchmark notebooks, and analysis scripts (metrics, multiple comparison, risk) raw_data/ Source CSV files (load and climate) as downloaded from the public agencies (~1.7 GB) trained_models/ Available trained artifacts (v0.2 weights and scalers)