TL;DR
The 2021-2022 energy crisis permanently split Nordic electricity zones into two groups: those that move in lockstep with continental Europe and those that remain isolated. This study reveals a surprising divergence: while price movements have split, the information needed to forecast those prices has not—a zone's coupling to the continent is a poor predictor of its forecasting needs, which are instead driven by the local generation mix (hydro vs. wind/solar).
What Problem Does This Paper Address?
Electricity markets are undergoing two simultaneous shifts: increased physical interconnection (market coupling) and a rising share of variable renewable energy (VRE) like wind and solar. Traditionally, researchers used price convergence as a proxy for market integration. However, the 2021-2022 energy crisis created a "regime break" where northern Nordic zones became insulated from continental price shocks due to grid bottlenecks, while southern zones remained tightly coupled.
The authors ask: Does this divergence in price movement change the information carried by prices? Specifically, if a zone is "coupled" to Germany, does it suddenly need German wind forecasts or neighboring price data to achieve accurate forecasts? Understanding this helps System Operators and traders decide which forecasts are worth producing and publishing before the day-ahead auction closes.
Methods
The researchers employed a rigorous out-of-sample forecasting framework across 10 bidding zones.
- Co-movement Analysis: They measured the correlation of intraday price profiles between Nordic zones and the Germany-Luxembourg (DE-LU) reference zone from 2018 to 2025.
- Forecasting Model: They utilized a per-hour LightGBM model. To handle the extreme price volatility of the crisis years, they implemented a "level-and-scale" target normalization. Instead of predicting absolute prices, the model predicts the deviation from the previous day’s mean, scaled by the previous day’s standard deviation. This allows the model to learn from "expensive" crisis data and apply those patterns to "cheaper" periods.
- Information Value: Following the predictive logic of Granger causality, they measured the "value" of an input (like VRE forecasts or neighboring prices) by observing the increase in Mean Absolute Error (MAE) when that input was removed.
- Double Machine Learning (DML): This was used to diagnose which market conditions (e.g., reservoir levels, temperature) were associated with higher forecast errors.
Datasets
The study uses a comprehensive dataset spanning 2018–2025, with 2024 used for model screening and 2025 for the final rolling test.
- Price Data: Day-ahead prices for 13 zones (Nordics + DE-LU + NL).
- TSO Forecasts: Load, wind, and solar forecasts from ENTSO-E.
- Physical Data: Weekly hydro reservoir filling (Norway/Sweden) and hourly weather reanalysis (Open-Meteo).
- Commodities: TTF gas, EUA carbon, coal, and oil futures.
Key Insights and Ideas
- The Persistence of the Split: The "Nordic Split" wasn't a temporary crisis artifact. As shown in Figure 2, the divergence in co-movement that began in 2020/2021 has persisted through 2025, creating two distinct market types: "Coupled" (NO1, NO2, SE3, SE4) and "Isolated" (NO3, NO4, SE1, SE2).
- Coupling $\neq$ Information Needs: Counter-intuitively, southern Norwegian zones (NO1, NO2) move almost exactly like Germany but gain very little from VRE forecasts. Conversely, northern Swedish zones (SE1, SE2) are isolated from Germany but gain significantly from VRE forecasts.
- Generation Mix is King: The value of wind and solar forecasts is dictated by the local generation system, not the interconnectors. In hydro-dominated Norway, reservoirs act as a buffer, absorbing local VRE fluctuations before they hit the price. In Sweden, which has more wind and nuclear, the price is far more sensitive to weather-driven supply.
- Neighboring Prices are Redundant: Adding the previous day's prices from neighboring zones improved accuracy by less than 2.2% across the board. The zone's own price history already captures the necessary continental signals.
Key Results
- VRE Forecast Impact: Removing wind and solar forecasts increased the relative MAE (rMAE) by 31% in Germany, 8–16% in Sweden, but only 0–6% in Norway.
- Forecasting Accuracy: The LightGBM model with normalization reduced the MAE of a naive D-1 forecast by 24–37% in Nordic zones and 47% in DE-LU.
- Isolated Zone Paradox: In isolated zones, prices are actually easier to forecast on high-wind days. The authors suggest that when local wind supply exceeds export limits, prices fall "low and flat," making them highly predictable.
Limitations and Future Work
- Point Forecasts vs. Tails: The study focused on Mean Absolute Error. Continental shocks might still be highly relevant for predicting "extreme" price events or probabilistic tails, which this study did not explore.
- Publication Timing: The authors assumed TSO forecasts are available at 11:00 D-1. In reality, VRE forecasts are often published later (18:00), meaning their "value" in this study represents a potential gain if publication rules were changed.
- Causality: The relationships found are associations. Future work using event studies (e.g., the exact moment a new cable like NordLink went live) could better isolate causal drivers of the split.
Summary updated