I ran into an old college friend a few weeks ago
At some point he mentioned he’s been trading Polymarket weather markets and casually said he pulled around $8,000 over a few months
At first it sounded like one of those “I got lucky” stories, but the way he described it didn’t feel random at all
He was talking about probabilities, forecast models, and how markets tend to misprice weather outcomes when different data sources don’t agree
That conversation stuck with me, so I went home and started digging into it properly
I spent the next couple of days researching how these markets actually work, pulling data from different forecast models like ECMWF, HRRR, and METAR observations, and trying to understand where pricing inefficiencies come from
The key idea I kept coming back to was simple: you’re not predicting weather, you’re trading probability vs price
So I framed everything as an expected value problem
P_true = weighted probability across models
P_market = implied probability from market price
EV = (P_true x payout) − (1 − P_true) x risk
If EV is negative, the trade doesn’t exist, no matter how “likely” it looks
On top of that I added position sizing using Kelly
f* = (p x b − q) / b
Where p is your edge, b is odds, and q is the downside probability
After putting all of this together, I gave the full framework to Claude and asked it to turn it into a working system
What it built wasn’t just a script, it started with a simulation layer
It replayed historical markets, tested entries with real delays, filtered out trades where the edge disappeared too quickly, and adjusted parameters based on outcomes
Then it started adapting on its own
It recalibrates probabilities after every resolution, adjusts weights between models, and filters out setups where the disagreement between data sources is no longer meaningful
Over time it stopped taking most trades and focused only on the clean ones
After I connected it to execution, it started trading weather markets automatically
No manual input, no second guessing, just a loop of scanning, validating, sizing, and executing
Over the last month it generated around $27,000 in profit
What’s interesting is not just the number, but how it behaves
It trades less than you’d expect, but almost always enters before the move becomes obvious, when pricing is still inefficient
And the longer it runs, the more it adapts
It learns which setups actually work and slowly removes everything else
This isn’t about predicting rain or temperature anymore
It’s about systematically exploiting how those probabilities are priced
And it’s only getting better with time
I wrote a full breakdown of the logic and formulas behind this in my weather markets article below👇
https://x.com/0xChainMind/status/2049239165959917780