“THE COMPLETE GUIDE TO TRADING WEATHER MARKETS LIKE A QUANT” — @polydao

Why this is in the vault

Founder flagged this article with a specific question: “Does this have any meat or rigor, or is it sponsored slop — or worse, a survivorship bias scam?” This doc is the answer, filed with a deliberate critical assessment rather than a straight summary.

The short verdict first, then the reasoning.

TL;DR verdict

It is not a scam in the strictest sense. It is legitimate content marketing for Kreo, a prediction-market copy-trading platform, disguised as a quant tutorial. The technical content (Python code, API references, methodology) is real and reusable. The claimed track records are cherry-picked survivor profiles. The business model is funneling readers to copy-trade the named “winners” via Kreo (which takes a cut). It’s structurally identical to 1970s-80s direct-mail “hot stock tip” newsletters, updated for Web3: same survivorship display, same “this could be you” framing, same upsell to a paid mirror service.

Is there meat? Yes — the code patterns and API lists are useful primitives for any probability-based prediction market work (including our own Elon tweet forecaster).

Is there rigor? No. The article never measures the author’s own P&L, never shows a forward-test, never addresses base rates, never compares the claimed strategies against the market’s own baseline efficiency. Every “success” claim is a backward-looking cherry-pick.

Is it actionable for us? Not in the way it frames itself. Our PM1 Brier-calibration work actually measured the thing this article hand-waves over, and found Polymarket is efficient enough that directional strategies from a retail quant stack should fail on average.

What the article claims

• “One trader turned $100 into $5,000 trading the weather”
• “Another runs a fully automated bot that prints five figures a month”
• “A third quit his job to do it full-time”
• Names 5-6 specific Polymarket “weather cartel” traders (neobrother, Hans323, Handsanitizer23, ColdMath, HondaCivic, BeefSlayer) with claimed P&Ls, some with dramatic numbers ($1k → $80k, $600 → $16k on a single market)
• Provides working Python code to: fetch Polymarket markets via Gamma API, pull forecasts from Open-Meteo, compute bucket probabilities via Monte Carlo around the point forecast, compare to market prices, log the edges
• Lists forecast APIs: NOAA, Open-Meteo, ECMWF/Copernicus
• Lists visual model viewers: Windy, Ventusky, Windy.app
• Walks through “temperature laddering” as the core strategy: buy multiple cheap buckets on the same market, accept that most expire worthless, hope one lottery-ticket hit pays 50-100x
• Ends with a CTA to Kreo (kreo.app/@trade) — a platform where you can mirror the named traders’ positions in real time

What’s actually in the content

The technical parts are legitimate:

• The Python code for fetch_market_by_slug, fetch_outcome_prices, fetch_hourly_forecast, build_probability_table, compute_edges, and the CSV logging pattern are ~identical to what we’d write ourselves. In fact the compute_edges function is structurally the same as our PM1e Elon forecaster.
• The API references (NOAA, Open-Meteo, ECMWF) are real and free for light use.
• The Gamma API pattern (fetch_market_by_slug) is the same one we use in autoinv.polymarket.
• “Temperature laddering” is a real technique — buy multiple adjacent probability buckets to smooth variance.

The quant-methodology parts have real gaps:

• The probability model is just “Monte Carlo around the point forecast with σ=1.5°C.” That σ is picked arbitrarily. Real meteorologists would use ensemble spreads or historical error distributions. Using a fixed σ means the probability estimates don’t respond to genuine forecast uncertainty — useless in exactly the regimes (weather fronts, rapid changes) where there might be edge.
• No factor model. No accounting for crowd anchoring on named cities, time-of-day patterns, recency effects, or any of the structural biases that might give informational edge.
• No out-of-sample test. The article never says “I ran this strategy on 2024 data, here’s the Sharpe.” It says “here’s the code, these traders make money.”

The survivorship bias problem

This is the core issue and the founder’s specific concern.

The article names 5-6 “top traders” and describes their P&Ls as if they prove the strategy works. Polymarket has ~500,000 total accounts and probably 10,000+ active weather traders. The article profiles the top winners. The 9,994+ losers don’t get screenshots, don’t get profile links, don’t get mentioned.

Classic selection bias. If 10,000 traders each flip a coin 100 times, roughly 5 will hit heads 70+ times by chance alone. Those 5 are then profiled as “the coin-flipping cartel.” The article does not compute a base rate against the full population, does not measure risk-adjusted returns, does not show the losing trades from the named winners, and does not discuss any structural reason to believe the named winners will continue winning prospectively.

This is the modern version of the classic direct-mail stock picking scam (Martin Gardner wrote about this in the 70s): send 10,000 people a stock tip, split them into halves predicting up or down, repeat 10 times. You end up with ~10 people who got 10 consecutive correct picks. To them, you look like a genius. You then sell them a newsletter subscription. The direct-mail version is illegal in the US now; the content-marketing-into-copy-trading-upsell version isn’t, because nobody’s explicitly promising returns.

The named traders are probably real people who really are winning. That’s not the question. The question is whether the article gives retail readers the tools and context to reliably replicate them — and it doesn’t. It gives them enough Python to feel like they could, then routes them to Kreo to copy-trade without having to actually build edge.

The Kreo upsell — the actual business model

The article’s closing CTA:

“Want to start without building the stack yourself? The traders in this article are live on Kreo — a copy-trading platform where you can mirror their weather positions in real time. Follow the top weather traders directly: kreo.app/@trade”

This is the tell. The article exists to drive signups to Kreo. The “quant cartel” profiles are the product. Kreo presumably takes a cut when users mirror positions.

Copy-trading prediction markets is structurally one of the worst forms of copy trading:

1. You’re buying AFTER the edge has been identified and announced to every Kreo user simultaneously
2. Your trade execution lags the source trader’s, so you pay worse prices on the entry
3. You have the same downside risk but a slimmer upside because you bought at a higher price
4. You’re paying Kreo’s platform fee on top of Polymarket’s trading fee
5. The source trader has no incentive to pick trades sized for YOUR risk tolerance — they’ll still take their own high-conviction bets
6. If the source trader gets lucky in 2024, they get another year of mirror subscribers in 2025, but their luck has no memory

How this compares to our own PM1 findings

The article’s thesis — “Polymarket weather markets are inefficient, retail quants can extract edge” — is directly testable, and we have adjacent data.

Our PM1 Polymarket baseline measured actual calibration on 611 resolved markets:

• Top volume tier ($10M+): Brier 0.0258, lift +0.13 over majority baseline
• $2M-$10M: Brier 0.0731, lift +0.07
• $500K-$2M: Brier 0.1487 (borderline fails gate)
• $100K-$500K: Brier 0.1202 (right at gate)
• $10K-$100K: Brier 0.1493 (fails gate)

Weather markets specifically would fall somewhere in the mid-volume tiers (maybe $100K-$500K for most cities, $500K-$2M for NYC/Chicago peaks). That’s the zone where Polymarket is still pricing within ~15 points of a majority-baseline Brier. A forecaster would need to beat the market’s own Brier of ~0.12, not just the discipline gate.

A naive Monte-Carlo-around-point-forecast with σ=1.5 is unlikely to meet that bar. The article doesn’t test whether it does.

The “quant cartel” likely really is making money. But they’re making money because they have edge the article doesn’t teach — better forecast ensembles, faster execution, better station microclimate models, genuine meteorological insight. The article teaches the scaffolding, not the edge.

What IS worth stealing from the article

Filing the usable primitives:

1. The Python patterns. fetch_market_by_slug + fetch_outcome_prices + compute_edges is almost identical to our existing Polymarket client. Validates our design.
2. The forecast API list. NOAA, Open-Meteo, ECMWF/Copernicus are all free for light use. Copernicus in particular is interesting for ensemble spreads.
3. The conceptual pattern. “Fetch market → compute model probability → compare to market → log edge → paper-trade → automate” is the correct shape for any probability-based strategy. It’s what our pm1e_elon_forecast already does.
4. The reminder that most prediction-market alpha is NOT in the highly-competitive liquidity zones. If there’s edge, it’s in the small, retail-dominated, narrow-bucket markets — the Elon tweet count style, not the NYC daily high.

What to do with the article

• Do NOT sign up for Kreo. Never copy-trade prediction markets.
• Do NOT try to replicate the temperature-laddering strategy on current Polymarket weather markets without first measuring the actual Brier calibration of that sub-category (same way we did PM1 for general markets). It would take ~30 minutes and would either confirm there’s edge (unlikely per our base rates) or kill the idea cleanly.
• DO steal the Python patterns and the forecast API references. They’re free.
• DO use this article as a teaching example of how content marketing disguises itself as quant tutorials. Every “here’s how the traders win” article should trigger the same survivorship-bias check we ran here.

One thing the article actually gets right

Near the end, buried in the “Risks” section:

• “Crowding: every article like this pushes more quants into the niche; edges compress over time.”

This is an honest admission. Any public article describing a strategy kills the strategy. If this article were about a real, repeatable edge, publishing it would destroy the edge. The fact that the article is being published suggests either (a) the author doesn’t believe the retail reader can execute, (b) the author’s business model is not trading but selling copy-trading, or (c) both.

Both are consistent with the Kreo upsell.

Tracked author

@polydao (“Mr. Buzzoni”, CEO of @polynternet). Prediction-market researcher and builder. Do not follow for trading signals. Could be worth following for API / methodology content, with the strong caveat that the framing around “easy edge for retail” is business-driven not research-driven.

Related

• ../01-projects/automated-investing/experiments/pm1-polymarket-baseline — our actual Brier measurements on Polymarket
• ../01-projects/automated-investing/experiments/pm1cd-category-and-stability — why sports/weather inherent variance limits what forecasters can achieve
• ../01-projects/automated-investing/experiments/pm1c3-other-breakdown — where narrow-bucket retail mispricing actually lives (Elon tweet counts, not weather)
• ../01-projects/automated-investing/experiments/pm1e-elon-forecast — our own “fetch market, compute probability, compare, log” pipeline
• 2026-04-10-halls-moore-algo-trading — the Ch 3 four-biases framework applied here

Copyright note

Direct quotes are short and in quotation marks. All analysis and framing is original. The article is public-facing content marketing; reproducing it verbatim would violate copyright, citing specific claims for criticism is fair use.