I wanted to build a Polymarket trading bot.
I could find plenty of people talking about building bots.
Nobody was actually saying HOW they built them.
So I built one from scratch. 798 commits in 6 weeks.
Here's everything I learned. A thread 🧵
Any exchange that continues to use WUnderground History tab for resolution is not serious about weather markets.
@IBKR@Polymarket@RobinhoodApp
There are two explanations for missing METARs on their history tab. Fraud or incompetence. Neither is acceptable for resolution.
Interesting concept but the repo doesn't match the description. It's a CS-542 class project on a Kalshi demo account. The XGBoost doesn't predict NWP model error, it predicts raw temperature from Visual Crossing historical CSVs. No ECMWF/GFS forecasts, no METAR data, no model spread features anywhere in the code.
The idea of ML bias correction on NWP is legit (meteorology has done this for decades as MOS, Model Output Statistics), but building it properly needs paired forecast-vs-observation training data. This repo just has historical weather and date features.
MSE of 27-31 for Chicago/NYC is 5°F+ error, that's not tradeable on 1°C bins.
Liquidity is absolutely the bottleneck at that window. From what I've seen:
- Spreads at 18-30hr are typically 2-5c on the tighter city bins (Milan, Atlanta, Karachi) but can blow out to 10-30c on thinner ones
- Best ask is usually 0.5-3c above fair value — close enough to see the edge, too wide to capture it passively
- FAK fill rates at your intended price are rough, maybe 1 in 10 opportunities actually execute
The edge exists at 18-30hr but you're paying for it in fill rate. The real question is whether you'd rather have a high fill rate at 6-12hr with no edge, or a low fill rate at 18-30hr with real edge. I'd take fewer fills at better prices every time.
The cities matter too. Some books are consistently tighter than others at that window.
10 lessons from building a weather market research system for Polymarket.
Most of these aren't written down anywhere. I learned them from building Weather Lab, studying community traders, and digging into the data.
Save this thread. 🧵
If you found this useful:
- Follow for ongoing updates as the bot evolves
- I'll keep posting strategy updates, live results, and new features
- The repo will be public when it's ready
- Ask questions — I'm not gatekeeping any of this
The whole point was: the next person who searches "how to build a Polymarket bot" deserves real answers.
Thanks for reading. Let's keep building.
12 days ago I started posting how I built a Polymarket trading bot from scratch.
Scanner → profitability → paper trading → live trading → weather markets.
798 commits. 6 weeks. A Raspberry Pi.
Here's what's next. 🧵
The prediction market bot ecosystem needs more people sharing the HOW.
Not "I made money with my secret bot."
Not "Just use the API."
Not screenshot P/L with no context.
Actual engineering. Actual architecture. Actual numbers — including the losses.
That's what I've tried to do with this thread series. I hope it helped.
The meta hard truth:
Building a prediction market bot teaches you more about markets, execution, and probability than any course or book.
The bot itself might or might not make money. The knowledge you gain building it is worth far more than the P/L.
If you're building one: budget real money for learning, keep your infrastructure boring, and be honest about whether your edge is real.
Tomorrow: what's next, and the open-source roadmap.
Hard truth #10: The best strategy is boring.
The flashiest strategy I imagined (copy trading with big size on hot markets) lost money.
The boring approaches work:
- Start with flat 5-share trades to generate real data cheaply
- Buy near-certain bins at 95-97¢ for 3-5% return
- Diversify across uncorrelated cities
- Only trade in the sweet window when your model has edge
- Skip anything uncertain
Nobody makes a viral tweet about "I made 3% on near-certain bins" or "I traded 5 shares to measure latency." But that's what profitable looks like.