Nordic Electricity Prices Stopped Moving Together After the Energy Crisis — But the Forecasting Value of Wind and Solar Data Didn't Split the Same Way

Prices diverged across Nordic zones post-crisis, but VRE forecast value tracks national generation mix, not continental coupling. A useful diagnostic for where to invest in forecasting infrastructure.

Imagine you live in a row of houses that all used to get water from the same reservoir. One day, a drought partially disconnects some houses — their pressure drops, their flow patterns change. You'd assume the houses with weird pressure also need different plumbing advice. This paper says: not necessarily. The pipes inside each house matter more than the connection to the reservoir. The committed claim: after the 2021-2022 energy crisis, Nordic bidding zones diverged sharply in how closely their day-ahead electricity prices track Germany-Luxembourg (DE-LU) — but this co-movement divergence does NOT predict where wind and solar forecasts are most valuable for price prediction. That second finding is the one that matters for practitioners. Four southern zones (SE3, SE4, NO1, NO5) still correlate strongly with DE-LU at intraday correlations of 0.54-0.71, while four northern zones (SE1, SE2, NO3, NO4) correlate weakly at 0.25-0.30. You'd expect the tightly-coupled southern zones to benefit most from continental VRE forecasts. They don't. The forecasting architecture is deliberately conservative: a single model selected on 2024 data, applied to forecast 2025 prices at 11:00 on the previous day. The model tests which inputs — neighbouring zone prices, TSO wind/solar forecasts, reservoir levels — actually reduce mean absolute error (MAE). Previous-day neighbouring prices helped by at most 2.2% MAE reduction in any zone, equally across coupled and isolated zones. The real discriminator was national, not zonal: removing TSO wind and solar forecasts increased MAE by at most 6% in Norway, 8-16% in Sweden, and a massive 31% in DE-LU. This ordering held across 2024 validation. The integrity picture is solid for an applied econometrics paper. The model was selected on 2024 data and tested on 2025 — a genuine temporal holdout. The co-movement analysis spans 2018-2025, giving pre-crisis, crisis, and post-crisis regimes. Code and results are published on GitHub. The one notable absence: no formal pre-registration, though the temporal holdout design partially substitutes for this. The reservoir-filling analysis in Norway found no detectable dependence of VRE forecast value on hydro storage levels, which is a useful null result honestly reported. What makes this paper genuinely useful rather than merely descriptive is the negative finding at its core. The energy policy instinct is to look at price coupling as a diagnostic for everything — market integration, forecast value, infrastructure investment priority. This paper cleanly separates two things that looked like they should track together and shows they don't. The value of VRE forecasts is a property of the national generation system (hydro-dominated Norway vs. wind-heavy Sweden vs. VRE-dependent Germany), not of how tightly a zone is coupled to continental markets. The ladder position is clear but modest. This is not competing with deep learning SOTA for price forecasting accuracy. The contribution is diagnostic, not predictive: it tells you where to aim your forecasting resources, not how to build a better forecaster. The single-model design is a feature (controls for model selection effects) but limits performance claims. Against the broader electricity price forecasting literature (Weron 2014, Lago et al. 2021), this is a well-executed applied contribution, not a methods breakthrough. The obvious next experiment is cross-border transmission capacity constraints as an explicit variable. The paper measures co-movement as an outcome but doesn't model the physical bottlenecks (NordBal Cable, NorNed, SwePol) that create the divergence. The authors likely view this as a different paper — transmission physics versus price information content — but it's the natural follow-on that would make the negative result mechanistically complete.