Probabilistic Prediction Markets with Intermittent Contributions
Authors: Michael Vitali, Pierre Pinson
Organizations: Dyson School of Design Engineering, Imperial College London, London, United Kingdom · Halfspace, Denmark · Department of Technology, Management and Economics, Technical University of Denmark, Denmark · CoRE, Aarhus University, Denmark
Although both data availability and the demand for accurate forecasts are increasing, collaboration between stakeholders is often constrained by data ownership and competitive interests. In contrast to recent proposals within cooperative game-theoretical frameworks, we place ourselves in a more general framework, based on prediction markets. There, independent agents trade forecasts of uncertain future events in exchange for rewards. We introduce and analyse a prediction market operating with a fixed pool of agents that (i) accounts for the historical performance, (ii) adapts to time-varying conditions, while (iii) accommodates intermittent forecast submissions. The proposed design employs robust regression models to learn the optimal forecasts' combination whilst handling missing submissions. Moreover, we introduce a pay-off allocation mechanism that considers both in-sample and out-of-sample performance while satisfying several desirable economic properties. Case studies using simulated and real-world data demonstrate the effectiveness and adaptability of the proposed market design.
Figures & tables
Fig. 1 : Market design overview
Fig. 2 : Convergence of estimated weights for QR (top) and RQR (bottom) with k=1 and τ=0.5 .
Fig. 3 : Average reward per session in excess of the equal share for QR and RQR (5% missingness), split into in-sample (dark) and out-of-sample (light) components. Error bars show ±1 standard deviation across runs.
Fig. 4 : Convergence of estimated weights for QR (top) and RQR (bottom) with k=1 and τ=0.5 in a time-varying scenario.
w1
w2
w3
Missingness
Model
bias
var
bias
var
bias
var
5%
RQR
−2.8±3.2
1.91±0.16
11.7±5.6
1.87±0.23
−8.9±4.2
2.08±0.21
Last-value Impute
16.7±3.7
2.06±0.19
−20.1±6.6
2.04±0.21
3.4±4.6
2.13±0.22
Mean-impute
8.6±3.7
2.03±0.19
−7.9±6.3
2.02±0.22
−0.7±4.4
2.12±0.22
10%
RQR
−3.8±3.9
1.85±0.17
13.5±7.2
1.75±0.18
−9.7±5.0
2.08±0.21
Last-value Impute
26.3±3.8
2.10±0.19
−32.0±7.2
2.08±0.20
5.7±5.0
2.15±0.23
TABLE I : Bias and variance of the estimated weights for the proposed RQR and the two benchmark models under varying missingness rates (all values are in units of 10−3 , reported as mean ± standard deviation across runs).
Fig. 5 : Bias with varying missingness rate ranging from 5% to 90%.
Model
Quantile
0.1
0.5
0.9
ECMWF
QRF
37.76
86.68
45.78
MLP
37.18
85.45
41.05
XGB
36.19
82.38
43.92
NOAA
QRF
41.11
100.42
52.24
MLP
40.13
96.98
49.14
TABLE II : Quantile loss (in MW) for different quantile levels and methods. For each quantile level, underlined values denote the best individual model and bold values the best method overall.
Fig. 6 : Moving average and instantaneous loss of the best-performing seller as well as the online combination methods.
Fig. 7 : Seven-day moving average of the combination weights for each quantile level over the test period for RQR with 5% missingness.
Fig. 8 : Distribution of rewards per session for all market participants considering the RQR method with 5% missingness. The dashed line indicates the equal share of the total reward.
Model
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
ECMWF
QRF
334.10
333.38
344.32
334.17
343.89
346.44
337.09
348.14
334.23
332.86
MLP
337.11
334.49
346.72
334.72
345.01
344.64
332.96
345.43
336.06
332.14
XGB
337.23
336.48
345.91
334.02
345.23
345.67
337.59
345.14
329.52
332.51
NOAA
QRF
331.16
332.09
339.89
331.30
341.44
343.75
328.76
341.82
332.32
332.62
MLP
333.04
332.29
344.89
334.98
344.49
343.43
331.03
340.45
332.08
333.31
XGB
332.40
333.79
350.74
335.27
345.17
345.12
333.56
344.96
333.36
328.81
TABLE III : Monthly rewards (in £) of each seller from March to December 2025, using RQR method with 5% missingness. Bold values denote the highest reward in each month.
QR
RQR
Operation
Mean
P95
Mean
P95
Forecast aggregation
6.7
6.8
8.6
8.4
Parameter update
13.3
11.9
31.2
28.6
Pay-off calculation
6.7
6.6
15.4
14.6
TABLE IV : Wall-clock time per market session of the online operations, summed over the three quantile levels (mean and 95th percentile, in μ s).