Improvement of Daily Temperature Forecasts from a Prediction Market
Abstract
Prediction markets aggregate and distill the collective views of participants who have a financial stake in being correct. They have been shown to produce more accurate forecasts in realms like election results, which are not easily amenable to mechanistic modeling because they entail forecasting human behavior. It is far less clear that prediction markets should improve the accuracy of predictions of physical quantities where credible mechanistic models already exist, and their output is further routinely augmented with statistical post-processing and expert human oversight. Here I evaluate ForecastEx prediction markets on the daily maximum temperature at 24 US locations from 11 February 2026 to 1 October 2026 (5,405 location-days) against 20 operational forecast systems at every hourly lead from 36 hours before the end of the target day. The alternatives include state-of-the-art numerical weather prediction models with biases removed, model output statistics applied to those models, optimal model blends, an artificial intelligence ensemble, and human-supervised forecasts with access to all the aforementioned information. I converted prediction market probabilities to a central value for comparison with alternative deterministic systems that output only a single value, and I compared them using mean absolute error. Where alternative systems provided probabilistic information, I applied basic ensemble model output statistics corrections and compared them to the prediction market probabilities using the continuous ranked probability score (CRPS). I made substantial efforts to verify alternative systems against the observations that align with their forecasts when they differed from the observations that the prediction market targeted. Whether lead time was counted to the end of the target day or to the time the daily maximum was observed, the prediction market central forecast value and CRPS had the lowest error at all meaningful lead times. At 6:00 PM local time on the evening before the target day, when none of the day’s temperatures had been observed, the prediction market’s mean absolute error was 1.49°F, compared with 1.97°F for the best deterministic alternative (24% [20%, 29%] lower), and its CRPS was 1.06°F compared with 1.19°F for the best EMOS-corrected probabilistic alternative (11% [8%, 15%] lower). The margins generally held or improved through the target day, when the market’s responsiveness to live information becomes more of an advantage. The advantage was a genuine gain in accuracy rather than an artifact of forecast latency, since the market was more accurate than each alternative even when sampled only at the times when the alternative issued a new value. The prediction market’s performance improvement over model output statistics deliberately designed to remove systematic biases, and over human-expert systems with access to the same information, indicates that the market combines information and potentially incorporates additional information more flexibly and dynamically than conventional alternatives.
Related
- Rigorous uncertainty quantification of probabilistic AI weather forecasts with conformal prediction
- Enhancing Tropical Cyclone Prediction Through AI Weather Model Ensemble Forecasts and Multi‐Source Data Integration
- Evaluating E-OBS, AgERA5, MARS-STAT, ERA5, and ERA5-Land for Daily Minimum and Maximum Temperature Across Four Mediterranean European Countries
- Paraglider flight logs as hypsometers: daily temperature profiles of an Alpine valley, its diurnal heating, and the nearly level ceiling of thermals
- A Graph Neural Network for Global Daily Fire Radiative Power Prediction at Medium-Range Lead Times
BibTeX
@misc{brown2026improvement,
title = {{Improvement of Daily Temperature Forecasts from a Prediction Market}},
author = {Patrick T Brown},
howpublished = {EarthArXiv},
year = {2026},
month = oct,
doi = {10.31223/x5522d},
url = {https://eartharxiv.org/repository/view/15351/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Improvement of Daily Temperature Forecasts from a Prediction Market
AU - Patrick T Brown
PY - 2026
DA - 2026/10/03
PB - EarthArXiv
DO - 10.31223/x5522d
UR - https://eartharxiv.org/repository/view/15351/
AB - Prediction markets aggregate and distill the collective views of participants who have a financial stake in being correct. They have been shown to produce more accurate forecasts in realms like election results, which are not easily amenable to mechanistic modeling because they entail forecasting human behavior. It is far less clear that prediction markets should improve the accuracy of predictions of physical quantities where credible mechanistic models already exist, and their output is further routinely augmented with statistical post-processing and expert human oversight. Here I evaluate ForecastEx prediction markets on the daily maximum temperature at 24 US locations from 11 February 2026 to 1 October 2026 (5,405 location-days) against 20 operational forecast systems at every hourly lead from 36 hours before the end of the target day. The alternatives include state-of-the-art numerical weather prediction models with biases removed, model output statistics applied to those models, optimal model blends, an artificial intelligence ensemble, and human-supervised forecasts with access to all the aforementioned information. I converted prediction market probabilities to a central value for comparison with alternative deterministic systems that output only a single value, and I compared them using mean absolute error. Where alternative systems provided probabilistic information, I applied basic ensemble model output statistics corrections and compared them to the prediction market probabilities using the continuous ranked probability score (CRPS). I made substantial efforts to verify alternative systems against the observations that align with their forecasts when they differed from the observations that the prediction market targeted. Whether lead time was counted to the end of the target day or to the time the daily maximum was observed, the prediction market central forecast value and CRPS had the lowest error at all meaningful lead times. At 6:00 PM local time on the evening before the target day, when none of the day’s temperatures had been observed, the prediction market’s mean absolute error was 1.49°F, compared with 1.97°F for the best deterministic alternative (24% [20%, 29%] lower), and its CRPS was 1.06°F compared with 1.19°F for the best EMOS-corrected probabilistic alternative (11% [8%, 15%] lower). The margins generally held or improved through the target day, when the market’s responsiveness to live information becomes more of an advantage. The advantage was a genuine gain in accuracy rather than an artifact of forecast latency, since the market was more accurate than each alternative even when sampled only at the times when the alternative issued a new value. The prediction market’s performance improvement over model output statistics deliberately designed to remove systematic biases, and over human-expert systems with access to the same information, indicates that the market combines information and potentially incorporates additional information more flexibly and dynamically than conventional alternatives.
ER -