My Polymarket BTC up/down bot picked the winning side in 6 of its 7 real-money trades. It still lost money. Its own log puts the loss at $1.36.
That is a small number, and I know it. Each trade put $3 to $4 at risk. But the way it lost is the useful part, because the same math works at any size.
This post shows every trade from the bot’s own trade log, and what those numbers say. Seven trades is a tiny sample. So I am not claiming the strategy works or fails in general. I am showing what happened when it ran.
This is not financial advice. These are my own results, and they do not predict yours.
I traded my own money in small amounts and lost some of it. Do not risk money you cannot afford to lose.
What is a Polymarket BTC up/down market?
Polymarket lists short markets on the price of Bitcoin. Each one asks whether BTC will be up or down at the end of a time window. Some windows are 5 minutes long and some are 15.
You buy an “Up” share or a “Down” share. If your side wins, the share pays $1 when the market settles. If it loses, the share pays nothing. You can also sell your share before that, at whatever price a buyer will pay.
Six of my seven trades were in 15 minute markets. One was in a 5 minute market. So when I say up/down bot in this post, I mostly mean the 15 minute version.
What did my bot do?
The bot watched the BTC price on Binance. When it saw a sharp move, it bought the matching side on Polymarket. A sharp move up meant buy “Up”, and a sharp move down meant buy “Down”.
It placed its buys as limit orders, the kind that wait on the order book for a seller. Then, in five of the seven trades, it sold the share right as the window closed instead of waiting for the market to settle.
I built the signal side of this in the Binance to Polymarket signal pipeline. That post shows how the bot reads Binance trades and spots a move.
What happened in the 7 live trades?
The bot traded on April 26 and 27, 2026. Here is every trade from its log, in order. The last column shows how each market settled.
| # | Market | Side | Bought at | Ended at | How it ended | Result | Settled |
|---|---|---|---|---|---|---|---|
| 1 | 15 min | Up | $0.92 | $0.9495 | Sold at the close | +$0.10 | Up |
| 2 | 5 min | Up | $0.92 | $0.995 | Held until it settled | +$0.25 | Up |
| 3 | 15 min | Up | $0.55 | $0.9495 | Sold at the close | +$2.16 | Up |
| 4 | 15 min | Up | $0.94 | $0.945 | Sold at the close | +$0.02 | Up |
| 5 | 15 min | Up | $0.23 | $0.00 | Settled as a loss | -$3.91 | Down |
| 6 | 15 min | Up | $0.93 | $0.9495 | Sold at the close | +$0.06 | Up |
| 7 | 15 min | Down | $0.96 | $0.9495 | Sold at the close | -$0.03 | Down |
| Total | -$1.36 |
Each result is rounded to the cent. The total comes from the unrounded log, so the rounded rows add up to -$1.35.
A few honest notes on these numbers. They are the bot’s own records in USDC, and the log does not split out fees. On four trades it also records slightly fewer shares sold than bought. If I count only what each sale brought in against what each buy cost, the loss is about $1.43. Trade 2 settled on the right side but the log records 99.5 cents a share, not $1, and I have not traced why.
So the bot was right about direction 6 times out of 7. It made money on 5 of the 7 by its log. And the account still went down.
Why did six right calls still lose money?
Look at what my bot paid. In five of the trades it bought at 92 to 96 cents. A share that costs 93 cents can pay at most $1, so the most it can ever make is 7 cents.
Then it sold early, at about 95 cents, so it kept even less. On the four early sales from those high prices, it made between minus 1 and plus 3 cents a share. Trade 7 is the odd one. It picked the right side, bought at 96 and sold at about 95, so it lost 3 cents on a correct call.
Those five high-priced trades made about 40 cents together. Then trade 5 bought “Up” at 23 cents, and the market settled Down. The bot lost the whole $3.91 it put in, which is about ten times what those five trades made.
Trade 3 is the only one that looks like my idea working. It bought at 55 cents and sold at about 95, for a $2.16 gain. It was not enough. Trade 5 took it all back and more.
What if the bot had held every share to settlement instead of selling at the close? Using the logged prices and share counts, and before any fees, the five early sales would have made about 94 cents more. The test would still have lost about 42 cents. One wrong call at full stake was bigger than six right ones.
Why is a 93 cent share dangerous?
None of my high-priced trades settled against the bot. My big loss came from the 23 cent bet. But the math on the expensive side worries me more, because that is where most of its trades were.
Say you buy at 93 cents and sell at 95. You make 2 cents a share when you are right. Two cents. If you are wrong and the market settles against you, you lose 93 cents. To break even, you need to be right about 98 times out of 100. One wrong call wipes out about 46 right ones.
Seven trades cannot tell you whether a bot is right 98 times out of 100. That is the real problem with this test.
How many orders did the bot actually fill?
The bot’s calibration log shows 20 buy attempts during this test. Seven filled. The other 13 did not go through.
I think that matters as much as the losses, so I want it in front of me next time. A test that assumes every signal gets filled is measuring a bot I never ran. And I do not know which way the 13 missed trades would have gone.
What would I check before the next real-money test?
These are the checks this record points to.
- Log the real fill price, the real exit price and the share count for every trade. The table above exists only because the bot logged them.
- Compare the average win to the average loss in dollars. The win rate alone hid this loss.
- Count how many signals actually filled, and test only on those.
- Run on live data with no real money until there are enough finished trades to say something. Seven is not enough. Not even close.
If your win rate has to beat the price you pay, you can check the gap with the Kelly criterion calculator. When the edge is zero or less, it tells you to bet nothing.
Before you trade
Polymarket’s international site does not allow trading by people in the United States and some other countries. US residents have a separate app, Polymarket US, run by a CFTC-regulated exchange. Check the terms where you live before you trade. I am not affiliated with Polymarket and earn nothing from it.
What should you take from this?
If a Polymarket bot shows a high win rate, ask what it paid for each win. Buying at 93 cents to sell at 95 means one wrong call can wipe out about 46 right ones.
My bot picked the right side 6 times out of 7 and lost money. Seven trades do not settle whether the idea works. They do show exactly where the money went.
The prompts I use to keep Claude from touching real money too early are in Claude trading bot prompts. The full bot build is in how I built the Polymarket trading bot.