A sparse stepwise screen matches Elastic Net at an hourly rMAE of 0.501 across six European price areas. Over 1,461 out-of-sample days, with equal weight given to each area, both methods reach the same accuracy. Yet the screen retains only 4.485 to 8.228 regressors from 562 candidates, while Elastic Net carries 88 to 134. LASSO trails at 0.514. The forecast error is equal, with a specification one tenth to one twentieth the size.
Grivas, Mandrup and Sauri keep the standard electricity price forecasting setup and replace the selector. They estimate a separate univariate regression for each delivery hour, giving 24 equations per day in every market. Each equation uses a 731-day rolling window and is recalibrated daily. The candidate set is d = 562.
Transformed prices for the previous seven days at all 24 hours enter first. The model then adds day-ahead load for day t and seven lags, together with day-ahead renewables (onshore wind, offshore wind, solar) over the same grid and six weekday dummies. Four more slots contain one t-2 close apiece for front-month TTF gas, API2 coal, Brent and EUA Dec futures. Prices are processed through the area hyperbolic sine transform, using the sample median for the shift and MAD for the scale, before forecasts return through sinh. The three multiple-testing selectors are finally refit by OLS. The in-sample window runs from 2020-01-01 to 2021-12-31. The four out-of-sample years cover 2022-01-01 to 2025-12-31 across DE-LU, FR, ES, NO1, FI and SE3.
Six selectors enter the comparison: LASSO and Elastic Net, FarmSelect (factor-adjusted LASSO), OCMT and GOCMT, plus Boosting with Multiple Testing. BMT is a forward stepwise procedure that allows at most one regressor into each stage. The paper credits BMT to Kapetanios et al. (2026) and Grivas et al. (2026), with this paper's lead author among the authors of the second study. At every stage, BMT tests each remaining candidate conditional on the variables already admitted. It accepts only the largest absolute HAC t-statistic, then stops once every remaining candidate falls below the family-wise threshold. OCMT admits all candidates that clear the threshold at each stage.
Eight variables versus 280
The admission rule produces the striking difference. OCMT and BMT use the same critical value function, yet their selected counts differ by a factor of thirty-five to fifty-seven. OCMT retains roughly 257 to 305 variables, exceeding LASSO's 107 to 132, and records the lowest value of no metric in any of the six markets. BMT averages 4.485 variables in NO1 and 8.228 in FR.
DE-LU gives the starkest count: 280.415 for OCMT against 8.107 for BMT. Elastic Net wins all four hourly metrics there, posting rMAE 0.426 with 98.942 variables. BMT follows closely at 0.428 with 8.107.
The procedures run under different settings. OCMT uses delta1 = 1, delta2 = 2 following Chudik and coauthors. BMT uses delta1 = delta2 = 1. Both set p = 0.05 and c = 1. Because the critical value increases with delta, OCMT faces the tighter later-stage threshold. BMT also calculates its threshold from the shrinking pool of remaining candidates rather than from d. Even with this looser per-stage filter, it finishes thirty-five to fifty-seven times sparser.
The paper partly credits conditional re-testing for that parsimony because it removes pseudo-signals. Its support comes from the group heat maps in Figure 4.2 rather than a test. BMT keeps only a small share of the Price lags, Load and RES groups. OCMT and GOCMT accumulate redundant clusters of highly dependent variables, which BMT avoids.
France is the break
France is the exception.
In FR, Elastic Net minimises all four hourly metrics. It records rMAE 0.470 and MAE 18.840 EUR/MWh, compared with BMT's 0.497 and 19.936. Elastic Net beats BMT in 21 of 24 hours under the univariate Diebold-Mariano test at 5%; BMT wins zero hours. The paper gives two readings of LASSO. Its FR paragraph puts LASSO ahead of BMT at the 10% level in the multivariate test, while the summary paragraph describes FR as statistically indistinguishable.
The authors disclose the FR loss. Their abstract bases its claim on statistical comparability and states the trade in the same clause, "while using less than one-tenth as many variables". Table 4.1 supports the accuracy claim on the six-area average, while FR goes the other way. Parsimony survives in every market: BMT uses 4.485 to 8.228 variables, compared with 88.135 to 133.525 for Elastic Net.
NO1 shows the opposite result. BMT wins all four metrics with 4.485 variables: rMAE 0.502, MAE 14.958, sMAPE 29.861%, RMSE 26.667. Elastic Net reaches 0.517 with 104.662 variables, and LASSO records 0.539 with 131.423. Neither shrinkage method beats BMT during a single hour. BMT is significantly more accurate than Elastic Net at 5% in FI and NO1, and at 10% in SE3. Against LASSO, it wins at 5% in DE-LU, ES and NO1, at 10% in FI, with ties in FR and SE3.
The shrinkage methods also switch order by market. Elastic Net beats LASSO in 23, 21 and 24 of 24 hours for DE-LU, ES and FR, without a result in the opposite direction. LASSO beats Elastic Net at 5% in FI and SE3. The paper supplies no ex-ante rule for choosing a selector market by market. The reversal remains clear: Elastic Net leads in FR, whereas BMT wins all four metrics in FI, NO1 and SE3.
Daily baseload is the aggregate that maps to a traded contract. BMT ranks first in 20 of the 24 area-metric combinations. Elastic Net wins the remaining 4, all in FR.
Average rMAE_t: BMT 0.408, Elastic Net 0.417, LASSO 0.424.
What a desk gets
BMT completes a full daily recalibration in 1 to 3 seconds, while the OCMT family needs 3 to 7. LASSO takes 40 to 127 seconds, Elastic Net 15 to 125, and FarmSelect 41 to 133. Every timing comes from a single thread on a Xeon Gold 6132 at 2.60 GHz. For a single production run, the authors are right to treat the difference as nearly irrelevant because the longest observed recalibration lasts 2 minutes. The saving matters when a backtest must be rerun after every proposed fundamental.
FarmSelect supplies the warning. Designed for strongly correlated covariates, it first decorrelates them through an estimated factor model and then applies the penalty. Its selected count resembles LASSO's at 102.976 to 129.063 variables. Performance does not. FarmSelect loses to LASSO on every hourly metric in every area, including 0.678 rMAE against 0.443 in DE-LU, and finishes last among all six methods for daily baseload in every area.
I would want two issues resolved before viewing this as anything beyond a development-cycle improvement. The paper motivates baseload forecasts through Base Day Futures traded in price areas that include DE-LU, FR and ES. A trader would compare the forecast with the quoted futures price for next-day delivery. Yet the analysis ends with rMAE, MAE, sMAPE and RMSE relative to a p(t-7,h) seasonal naive. Without a hit rate or settlement cost, the 0.009 rMAE_t difference between selectors remains unpriced.
The asinh shift and scale raise the other issue. They use the sample median and MAD. We did not find a statement saying whether those values are recomputed for every 731-day window or kept fixed. That choice determines whether the 2022 crisis level enters the preprocessing.
We could not test the method on our own data. Its dependent variable is the hourly day-ahead clearing price in specific European bidding zones. The inputs combine ENTSO-E day-ahead load and renewables forecasts with front-month TTF, API2 and Brent closes and EUA Dec closes. The USD-quoted API2 and Brent series are converted using the ECB EUR/USD reference rate. We hold none of these series. Replacing them with US equities or ETFs would discard the auction mechanism on which the full specification depends.
France would look more like one draw in six, rather than a real defect, if another run chose the selector separately for each market on the training window and then applied that choice out of sample. The relevant question is whether this procedure recovers Elastic Net in FR and BMT in the Nordics.