Prediction Markets · research

Research Paper: Forecast Efficiency in CFTC-Regulated Weather Prediction Markets

Last Updated: April 1, 2026

Research Paper: Forecast Efficiency in CFTC-Regulated Weather Prediction Markets

Published: March 2026 | OddsReference Research

This paper presents a comprehensive analysis of market efficiency in CFTC-regulated weather prediction markets on Kalshi. Using 40,032 settled daily high temperature contracts and an NWS MOS ensemble model, we test whether publicly available weather forecast data can generate tradeable edge against market prices.

Download

Download Full Paper (PDF)

Abstract

We analyze 40,032 settled daily high temperature contracts on the Kalshi prediction market exchange. Using an NWS MOS ensemble model combining GFS and NAM forecasts weighted by inverse RMSE, we achieve 33% bracket accuracy (2x random) across 1,506 NYC events. Three trading strategies are tested: tail-selling (96.6% win rate, 0.62% ROI at taker fees), conditional filtering (45.2% on 93 events, likely overfit), and top-2 straddle (unprofitable at 76.5c market price vs 61.6c breakeven). Price convergence analysis on 804,248 hypothetical trades confirms smooth convergence: winners drift from 43.8c to 77c, losers from 15.2c to 4c. Markets incorporate NWS forecast updates within 10-30 minutes, but the 1-3c typical movement is smaller than round-trip trading costs of 3-3.5c. A mild favorite-longshot bias exists (0.4 pp on sub-10% contracts) but is too small to exploit. We conclude that daily weather prediction markets on Kalshi are approximately efficient in the semi-strong sense.

Key Findings

FindingDetail
Model accuracy33.0% top-1, 61.6% top-2 bracket hit rate
Best strategy ROI+0.62% (tail-selling at taker fees)
Worst strategy ROI-19.5% (top-2 straddle at taker fees)
Price convergenceWinners: 43.8c to 77c; Losers: 15.2c to 4c
NWS update lag10-30 minutes, 1-3c movement
Favorite-longshot bias0.4 pp on sub-10% contracts
Efficiency conclusionSemi-strong form efficient

Paper Structure

  1. Introduction — Research questions and motivation
  2. Market Structure — Contract design, fees, participant mix
  3. Data — 48,978 weather observations, 46,248 CLI reports, 40,032 contracts
  4. Model Development — NWS ensemble, uncertainty estimation, calibration
  5. Trading Strategy Tests — Tail-selling, conditional filtering, top-2 straddle
  6. Price Convergence Analysis — 804,248 trade scenarios, bracket rank effects, seasonal patterns
  7. Structural Biases — Favorite-longshot bias, seasonal variation, volume patterns
  8. Discussion — Efficiency interpretation, cross-domain comparison, market design implications
  9. Conclusion — Summary and future directions

This paper is part of a two-paper series on prediction market efficiency:

  • This paper: Weather market efficiency (NWS data, 40,032 contracts)
  • Companion paper: Crypto market efficiency (877,606 contracts, three pricing models)

Source Articles

The underlying analysis is documented in four detailed articles:

For real-time prediction market data across all platforms, visit our live dashboard.

Frequently Asked Questions

What is this weather prediction market research paper about?

This paper analyzes 40,032 settled Kalshi weather contracts using an NWS MOS ensemble model combining GFS and NAM forecasts weighted by inverse RMSE. We test three trading strategies across 1,506 NYC events and find that none produce positive after-fee returns, concluding that weather prediction markets are approximately efficient in the semi-strong sense.

Can you trade weather prediction markets profitably?

Our research found no profitable strategy using publicly available NWS data. Tail-selling achieved a 96.6% win rate but only 0.62% ROI at taker fees. Conditional filtering hit 45% on 93 events but every filter exceeding 40% had fewer than 120 qualifying events — the signature of overfitting. The market aggregates public forecast data within 10-30 minutes of release.

How does weather market efficiency compare to crypto markets?

Both domains show approximately efficient pricing. Weather markets exhibit a 0.4 pp favorite-longshot bias versus 2-4 pp in crypto markets. The smaller bias reflects lower emotional stakes — nobody has an identity-based attachment to a temperature bracket. Both markets incorporate public information within minutes and leave no exploitable edge after transaction costs.

Key Takeaways

  • An NWS ensemble model achieves genuine predictive skill (33% bracket accuracy, 2x random) but cannot generate after-fee trading profits on Kalshi weather contracts
  • Three strategies tested: tail-selling (marginally positive), conditional filtering (overfit), top-2 straddle (unprofitable at market prices)
  • 804,248 trade scenarios confirm smooth price convergence with 10-30 minute NWS update incorporation
  • Weather markets are approximately semi-strong efficient — the crowd aggregates public expert forecasts faster than any individual model
  • Cross-domain comparison with crypto markets shows the same efficiency conclusion despite fundamentally different underlying assets

Frequently Asked Questions

What is this weather prediction market research paper about?
This paper analyzes 40,032 settled Kalshi weather contracts using an NWS MOS ensemble model. We test three trading strategies across 1,506 NYC events and find that none produce positive after-fee returns, concluding that weather prediction markets are approximately efficient in the semi-strong sense.
Can you trade weather prediction markets profitably?
Our research found no profitable strategy. Tail-selling achieved 96.6% win rate but only 0.62% ROI at taker fees. Conditional filtering hit 45% on 93 events but is likely overfit. The top-2 straddle was unprofitable at market prices. The market aggregates NWS data faster than any model can exploit.
How accurate is the NWS ensemble model for weather trading?
The model achieves 33% top-1 bracket accuracy (2x random) and 61.6% top-2 accuracy using GFS and NAM forecasts weighted by inverse RMSE. While the model demonstrates genuine predictive skill, Kalshi markets already incorporate this same public data within 10-30 minutes of forecast release.