Isotonic Quantile Regression Averaging for uncertainty quantification of electricity price forecasts
Authors: Arkadiusz Lipiecki, Bartosz Uniejewski
Organizations: Doctoral School, Faculty of Management, Wrocław University of Science and Technology · Department of Operations Research and Business Intelligence, Wrocław University of Science and Technology,50-370 Wrocław, Poland
Quantifying the uncertainty of forecasting models is essential to assess and mitigate the risks associated with data-driven decisions, especially in volatile domains such as electricity markets. In this paper, we propose a computationally efficient isotonic variant of Quantile Regression Averaging (QRA), called Isotonic Quantile Regression Averaging (iQRA), for generating probabilistic forecasts from ensembles of point forecasts. Building on QRA, we introduce stochastic order constraints as a hyperparameter-free regularization mechanism that reduces computational complexity while maintaining competitive forecast accuracy. We further demonstrate that the same isotonic regularization principle can be applied to the smoothed estimator of QRA. In an extensive forecasting study of two European day-ahead electricity markets, we show that iQRA and its smoothed variant consistently rank among the top state-of-the-art postprocessing methods in terms of reliability and sharpness. It produces well-calibrated prediction intervals, with smoothed iQRA providing superior reliability even when compared against coverage-oriented conformal prediction. We further introduce a battery arbitrage evaluation, showing that the leading probabilistic methods achieve the highest profit per trade and that probabilistic risk control improves the economic value of trading decisions. Isotonic regularization offers a scalable hyperparameter-free variable selection for linear quantile regression, making iQRA attractive for practical decision-support systems in energy markets.
Figures & tables
Figure 1: Day-ahead electricity prices in the Germany-Luxembourg (top) and French (bottom) markets for the period 8.1.2015–31.12.2025. The first vertical dashed line marks the end of the 1456-day training window for the point forecasting models. The second dashed line marks the end of the 364-day calibration window for the post-processing methods and the beginning of the 2192-day out-of-sample test period. The red dot in the French market series indicates the peak price of 3000 EUR per MWh.
Figure 2: Visualization of the NARX network with five hidden neurons with hyperbolic tangent activation functions and one linear output neuron.
Germany
France
2020
2021
2022
2023
2024
2025
2020
2021
2022
2023
2024
2025
HS
1.53
4.71
10.90
5.79
5.12
5.23
1.46
5.32
12.49
5.99
5.14
5.04
CP
1.51
4.73
10.89
5.80
5.10
5.26
1.45
5.43
12.46
5.92
5.18
5.13
IDR
1.60
4.76
11.19
5.11
5.08
4.69
1.51
5.38
12.71
5.53
5.17
4.83
QRA
1.60
4.69
11.38
5.69
5.25
5.16
1.50
5.38
13.14
6.04
5.43
5.33
SQRA
1.59
4.70
11.32
5.70
5.23
5.14
1.49
5.39
13.07
6.04
5.39
5.30
Table 1: Continuous ranked probability score (CRPS) for a postprocessed ensemble of 25 NARX point forecasts and six years of out-of-sample period on German and French markets.
Figure 3: Number of years in which the model on the vertical axis outperformed the model on the horizontal axis according to the Giacomini-White test at the 1% significance level. Results obtained for an ensemble of 25 NARX forecasts.
Germany
France
10
25
50
10
25
50
HS
5.592
5.542
5.525
5.960
5.904
5.893
CP
5.594
5.544
5.527
5.982
5.926
5.915
IDR
5.428
5.403
5.404
5.880
5.853
5.856
QRA
5.455
5.626
5.973
5.922
6.134
6.578
SQRA
5.440
5.609
5.965
5.903
6.109
6.576
Table 2: Continuous ranked probability score (CRPS) for different sizes of the ensemble of NARX point forecasts, on German and French markets, aggregated for the entire out-of-sample testing period.
Germany
France
10
25
50
10
25
50
IDR
✓
✓
QRM
✓
SQRM
✓
LQRA
✓✓
✓✓
✓✓
✓✓
✓✓
✓✓
iQRA
✓✓
✓
✓
✓✓
✓✓
✓✓
Table 3: The results of the Model Confidence Set based on the daily CRPS values and the Tmax statistic ( Hansen et al., 2011 ) . The columns correspond to two markets and three different ensemble sizes. The test was run with 10,000 bootstrap replications and two significance levels: α=1% and α=5% . One checkmark indicates that the model was included in the 99% confidence set, two checkmarks denote models included in the 95% confidence set. Models omitted from the table were included in the test, but not retained in either confidence set.
Germany
France
1-99%
2-98%
5-95%
10-90%
1-99%
2-98%
5-95%
10-90%
HS
1.99
2.68
4.21
5.08
2.08
2.84
3.90
4.79
CP
1.44
2.18
3.58
4.57
1.58
2.27
3.53
4.18
IDR
7.09
7.48
7.64
7.33
6.92
7.10
7.05
6.58
QRA
7.11
6.83
7.37
7.52
6.58
6.37
7.07
7.54
SQRA
5.99
5.81
5.48
5.05
5.43
5.33
5.21
4.85
Table 4: Average Coverage Error (%) for the ensemble of 25 NARX point forecasts and four prediction interval levels, across two markets.
Germany
France
1-99%
2-98%
5-95%
10-90%
1-99%
2-98%
5-95%
10-90%
HS
0.93
1.45
2.59
3.93
0.95
1.51
2.73
4.15
CP
0.92
1.44
2.60
3.93
0.94
1.50
2.74
4.18
IDR
0.97
1.40
2.39
3.65
1.09
1.56
2.64
3.96
QRA
0.99
1.42
2.50
3.81
1.07
1.54
2.72
4.15
SQRA
0.99
1.42
2.48
3.79
1.06
1.53
2.69
4.11
Table 5: Prediction Interval Pinball Score (PIPS) for the ensemble of 25 NARX point forecasts and four prediction intervals, across two markets.
Figure 4: Gain per trade (EUR) and loss avoidance (%) from strategies based on postprocessed ensemble of 25 NARX forecasts, on the German market, aggregated for the entire sample period. Baseline profit per trade equates to 43.1 EUR and is marked by a dashed line. The bars represent strategies with different risk thresholds - 1/3 for dark blue, 2/3 for blue, and 1 for light blue color.
Quantity
QRA
iQRA
Nonnegative variables
2M+2T+2
M+2T+2
Equality constraints
T
T
Nonzero entries in A
T(2M+4)
T(M+4)
Table 6: Sizes of the QRA and iQRA linear programming formulations in the standard form minAx=b,x≥0c⊤x , where M is the number of regressors and T is the number of observations. Both formulations include an unrestricted intercept, while iQRA constrains the slope coefficients to be nonnegative. The nonzero-entry counts assume that all entries of the regressor matrix are nonzero.
QRM
3.03 ± 0.02
SQRM
3.76 ± 0.04
Ensemble size
10
25
50
QRA
4.56 ± 0.23
7.56 ± 0.15
11.73 ± 0.13
SQRA
14.78 ± 0.37
96.71 ± 2.40
473.93 ± 16.02
LQRA20
117.98 ± 2.13
220.40 ± 8.45
456.14 ± 22.96
LQRA1
5.90 ± 0.11
11.02 ± 0.42
22.81 ± 1.15
Table 7: Time (s) required for each quantile regression method to generate quantile forecasts for a single day (99 quantiles for each of the 24 hours) in Julia with a single execution thread on Apple M2 Pro, not including the compilation time. The values include average and standard deviation over 10 runs. LQRA20 corresponds to the time for estimating the model over the full regularization path consisting of 20 λ values, while LQRA1 indicates the time per single λ value.
Figure 5: The plots show how often (in percentages) given predictions are included in the final model (corresponding βi has non-zero value). We report the percentages separately for each forecasted quantile (vertical axis) and for each point prediction in the ensemble (horizontal axis). Recall that the 1-st prediction corresponds to the ensemble minimum, 13-th to the median, and 25-th to the maximum. The aggregated percentage of selected variables is 11.4% for LQRA, 13.7% for iQRA and 14.1% for iSQRA. The plots correspond to the ensemble of 25 NARX forecast in the German market.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Germany
France
2020
2021
2022
2023
2024
2025
2020
2021
2022
2023
2024
2025
NARX 10
4.07
11.89
29.24
14.85
13.48
14.24
3.89
13.74
32.86
15.22
14.32
14.25
NARX 25
4.02
11.76
29.00
14.72
13.37
14.13
3.85
13.63
32.46
15.07
14.24
14.15
NARX 50
4.00
11.72
28.89
14.64
13.33
14.10
3.83
13.58
32.46
15.03
14.22
14.13
XGB 10
4.17
14.50
32.41
13.91
12.74
12.72
4.07
15.29
35.63
14.81
14.09
13.90
XGB 25
4.15
14.47
32.31
13.81
12.71
12.70
4.02
15.24
35.51
14.71
14.05
13.83
Appendix
Table 8: Mean Absolute Error (MAE) for point forecasts obtained as an averaged prediction of NARX or XGB ensembles, for ensemble sizes of 10, 25 and 50 models. MAE is reported separately for each of the six calendar years in the out-of-sample period, on German and French markets.
Number of trades
Profit per trade
π
1/3
2/3
1
1/3
2/3
1
HS
535
433
357
51.3
60.4
67.6
CP
539
429
355
50.8
60.1
67.8
IDR
586
444
337
42.1
54.1
61.5
QRA
550
449
348
42.2
51.1
58.2
SQRA
545
449
343
43.7
51.6
59.9
Appendix
Table 9: Number of trades and profit per trade for the economic evaluation of the postprocessed ensemble of 25 NARX models in the German day-ahead market. Profit per trade corresponds to the values plotted in Fig. 4 .
Parameter
Range of values
Maximum depth
{2,...,10}
Learning rate
[10−4,1]†
Subsampling rate
[0.5,1]
Minimum child weight
{0,...,10}
Minimum loss reduction ( γ )
[0,5]
L1 regularization strength ( λ )
[10−3,10]†
Appendix
Table 10: Hyperparameters and the considered value ranges for the XGB model. Logarithmic scale was used for the ranges indicated with † .
Germany
France
2020
2021
2022
2023
2024
2025
2020
2021
2022
2023
2024
2025
HS
1.56
5.69
12.26
5.57
4.92
4.80
1.50
6.00
13.63
5.95
5.06
5.03
CP
1.56
5.87
12.24
5.56
4.88
4.84
1.52
6.16
13.67
5.92
5.13
5.12
IDR
1.64
5.54
12.68
5.20
5.01
4.61
1.51
5.98
14.05
5.65
5.15
4.97
QRA
1.65
5.42
12.67
5.48
5.06
4.95
1.54
5.87
14.27
5.89
5.31
5.28
SQRA
1.64
5.42
12.63
5.48
5.03
4.93
1.53
5.88
14.22
5.89
5.27
5.25
Appendix
Table 11: Continuous ranked probability score (CRPS) for a postprocessed ensemble of 25 XGB point forecasts and six years of out-of-sample period on German and French markets.
Figure 6: Number of years in which the model on the vertical axis outperformed the model on the horizontal axis according to the Giacomini-White test at the 1% significance level. Results obtained for an ensemble of 25 XGB forecasts.
Germany
France
10
25
50
10
25
50
HS
5.818
5.800
5.792
6.220
6.194
6.192
CP
5.840
5.822
5.814
6.277
6.250
6.249
IDR
5.775
5.777
5.784
6.219
6.215
6.226
QRA
5.644
5.867
6.313
6.098
6.356
6.833
SQRA
5.631
5.854
6.318
6.080
6.338
6.829
Appendix
Table 12: Continuous ranked probability score (CRPS) for different sizes of the ensemble of XGB point forecasts, on German and French markets, aggregated for the entire sample period.
Germany
France
10
25
50
10
25
50
LQRA
✓✓
✓✓
✓✓
✓✓
✓✓
✓✓
iQRA
✓✓
✓
✓✓
iSQRA
✓
✓✓
✓✓
✓✓
✓✓
✓✓
Appendix
Table 13: The results of the Model Confidence Set based on the daily CRPS values and the Tmax statistic ( Hansen et al., 2011 ) for the ensemble of XGB point forecasts. The columns correspond to two markets and three different ensemble sizes. The test was run with 10,000 bootstrap replications and two significance levels: α=1% and α=5% . One checkmark indicates that the model was included in the 99% confidence set, two checkmarks denote models included in the 95% confidence set. Models omitted from the table were not retained in either confidence set.
Germany
France
1-99%
2-98%
5-95%
10-90%
1-99%
2-98%
5-95%
10-90%
HS
1.89
2.69
4.03
4.63
2.15
3.02
4.03
4.41
CP
1.57
2.30
3.80
4.28
1.69
2.50
3.75
3.99
IDR
8.57
8.63
8.49
7.70
8.41
8.43
8.15
7.05
QRA
6.32
6.19
6.41
6.36
6.04
5.95
6.24
6.32
SQRA
5.32
5.13
4.62
3.78
5.26
5.15
4.53
3.84
Appendix
Table 14: Average Coverage Error (%) for the ensemble of 25 XGB point forecasts and four prediction intervals, across two markets.
Germany
France
1-99%
2-98%
5-95%
10-90%
1-99%
2-98%
5-95%
10-90%
HS
1.02
1.60
2.84
4.22
1.03
1.64
2.97
4.46
CP
1.02
1.61
2.85
4.24
1.03
1.65
3.00
4.50
IDR
1.19
1.63
2.69
3.98
1.30
1.80
2.93
4.31
QRA
1.00
1.47
2.59
3.97
1.13
1.64
2.87
4.35
SQRA
1.02
1.48
2.58
3.95
1.14
1.65
2.85
4.33
Appendix
Table 15: Prediction Interval Pinball Score (PIPS) for the ensemble of 25 XGB point forecasts and four prediction intervals, across two markets.
Figure 7: Gain per trade (EUR) and loss avoidance (%) from strategies based on postprocessed ensemble of 25 XGB forecasts, on the German market, aggregated for the entire sample period. Baseline profit per trade equates to 40.1 EUR and is marked by a dashed line. The bars represent strategies with different risk thresholds - 1/3 for dark blue, 2/3 for blue, and 1 for light blue color.
Number of trades
Profit per trade
π
1/3
2/3
1
1/3
2/3
1
HS
522
409
317
51.4
60.6
69.1
CP
519
409
306
51.5
59.6
71.0
IDR
587
472
340
40.5
49.7
58.3
QRA
597
452
316
36.4
48.6
61.9
SQRA
591
449
316
36.7
49.0
62.7
Appendix
Table 16: Number of trades and profit per trade for the economic evaluation of the postprocessed ensemble of 25 XGB models in the German day-ahead market. Profit per trade corresponds to the values plotted in Fig. 7 .
School of Electrical and Mechanical Engineering, The University of Adelaide, Ingkarni Wardli Building, Adelaide, SA 5005, Australia · Watts A/S, Brogade 19D, Køge 4600, Denmark