Closing Reaper

Picking-up the last tiny percentages from Polymarket,
a few dollars and a few minutes at a time.

It started with a very pragmatic account,
betting on Jesus not coming back, aliens not being revealed,
or Trump not invading Greenland.

Very serious investments, basically.

@pragmaticapproach on Polymarket

A tiny percentage, but in a tiny time

While placing those almost-certain bets, I noticed something pretty simple:
even very close to their end, some markets were still offering 0.5%, 1%, sometimes a bit more.

It doesn't look exciting.
Humans like big numbers, and a 1% profit sounds ridiculous next to a nice yearly +20%.
But 1% made in a few minutes isn't small at all.

The problem is that doing it manually makes no sense.
Finding the market, checking it, placing $10, then earning a few cents is definitely not a job.
So this boring idea had to be automated.

How the little reaper works

Every 15 minutes, the bot looks for markets expected to finish during the next 15 to 20 minutes.
It ranks them using their potential return and remaining time, then keeps opportunities priced between 98.5% and 99.5%.

Of course, 99% doesn't mean safe.
The bot also checks volumes and the order book, making sure enough people are actually ready to buy and sell around that price.
An empty order book can turn a nice 99% into a much riskier 96% while the order is being filled.

Each position is currently limited to $10.
I started with $100, quickly realized the whole amount was constantly locked into unresolved markets, then added $200 to let the machine breathe.
This isn't optimized money management yet, just a small enough scale to learn with real money without making every mistake expensive.

Real trades, live results

I didn't want a clean backtest explaining how rich I could theoretically be.
Here are the actual trades made automatically by @closingreaper, including wins, losses and rejected orders.


At the time of writing, the bot made around 600 to 700 trades over three weeks, including one full week paused.
The first version lost money, mostly because I was learning with the machine already running.
Recent versions recovered a good part of it, but two weeks of activity obviously doesn't prove a profitable strategy.
It may work, or it may just be a nice lucky rally. The dashboard will answer this better than this text.

Fast mistakes are useful mistakes

The first mistake was slippage.
Asking to buy at 99% doesn't magically create sellers at 99%. With a thin order book, the order keeps eating lower prices, and the supposed tiny safe profit becomes a much bigger risk.
I then added checks to make sure the book is deep enough before trading.

The second mistake was trusting Polymarket's end date.
For some football markets, the given end date was actually the beginning of the match.
Quite problematic when the whole strategy is about betting close to the end.
The bot now detects football matches and adds about 85 minutes, trying to bet closer to their actual finish.
Postponed games and extra time can still break this approximation.

I will probably also exclude markets with very little inertia.
A football result is unlikely to completely change during its last five minutes, but Bitcoin definitely doesn't care about my probability model.

One day of prompting

From the first idea to real automated trades, this took me around 7 to 9 hours.
I didn't write the code myself. My work was to structure the problem, split it into small steps, prompt, test, observe, then correct.

First listing markets.
Then connecting a Polymarket wallet.
Then using dry-runs before allowing real orders.
Then securely running everything on my server, with requests relayed through another server in Spain.

Broad prompts often produce broad wastes of tokens.
Keeping control of the structure, and validating every small block before going further, made the implementation surprisingly fast.
The code came from Ai, but the decisions, limits and paranoia had to come from me.

Why all that ?


Mostly to trust ideas that feel true, even when they also feel too stupid, too easy, or too obvious to work.

This one was stupid enough to make me doubt it,
but cheap and fast enough to test in one day.
At worst, I lose some time and $50, and buy myself a real answer.
At best, I build a small passive income with a very specific taste.

Someone probably thought about it before.
Maybe they implemented it badly.
Maybe I'm implementing it badly too.
There is only one nice way to know: making it run, with real data and real money.

A good idea and a good implementation multiply each other.
So even a silly idea deserves a clever implementation,
especially when Ai made the cost of trying it almost ridiculous.