Weather tails: non-modal buckets and sub-cent lottery tickets
Born from our audit of the "weather legends" article: test both the lottery-ticket story and the one real winner's style, at executable asks. Both lanes lost — the lottery lane catastrophically.
The hypothesisTwo claims from the weather-market hype cycle, tested side by side. TAIL: deep-tail temperature buckets at 0.2–2¢ pay 100–500× often enough to be +EV (the article's story). ADJ: mid-priced non-modal buckets are mispriced when a forecast model disagrees with the ask by 8+ points (the style of the one audited trader who actually earns).
The test
Inside the hourly weather-edge cycle, every open temperature event logged its non-modal buckets at the live executable ask: lane TAIL for asks ≤2¢ gated on our model's probability being ≥3× the ask, lane ADJ for asks ≤50¢ gated on model−ask ≥8 points with a two-sided book. Settled at resolution net of fees. Because the TAIL payoff is 100–500×, its verdict gate was set at 300 settlements, not 30 — skew that steep needs a long tail to judge a tail.
The result
Both lanes failed, at scale, quickly. The lottery lane settled 8,565 entries with two wins — −98.2% ROI — the mathematically honest version of "occasionally pays 500×": occasionally is not often enough. The non-modal lane settled 4,875 entries at −58.9%: the asks on adjacent buckets already encode the uncertainty our forecast model claimed to see, which is the same finding as the main forecast-drift probe from a different angle.
Conclusion
Negative on both counts, and a clean close to the weather-legends audit: the article's mechanism loses catastrophically, and the best real trader's mechanism is not reproducible from public forecasts — if their edge exists, it is station-specific settlement knowledge, not a recipe. Weather temperature buckets are now the most thoroughly tested market class in this lab (GPT forecaster, numeric drift at seven lead times, bias-corrected blends, observed-max latency, tails and non-modals) and every automated entry thesis has lost at the ask. The data platform stays; the trading theses are closed.