About ten dollars
Months of building prediction-market bots earned roughly the price of a sandwich. Here is what the money couldn't measure.
· 2 min read · markets · engineering
For a few months this year I ran automated strategies on Polymarket: some directional trading, some market making, some attempts at arbitrage. I built the data collection, the backtests, the risk gates and the bots. The total profit over all of it was about ten dollars.
That number is the honest headline, and I would rather say it plainly than dress it up. It is also the least interesting thing that came out of the project.
Activity is not edge
The bots were busy. They watched markets, placed orders, cancelled them and logged everything. On a dashboard, busy looks like progress. It isn’t. A strategy that trades a lot and earns nothing has only proven that it can trade.
The question that matters is narrower: in which situations do I know something the price doesn’t? For most markets, most of the time, the answer was “nowhere”. The market was better informed than my code.
The backtest is the first suspect
Every good-looking backtest I produced had a leak in it somewhere. Sometimes the leak was data from the future sneaking into a feature. Sometimes it was a fill price that assumed the order book would wait for me. Sometimes it was simply testing on the same period I had tuned on.
I now treat a great backtest the way I treat a too-good benchmark: as a bug report. The fix is boring. Use only the data that existed at that moment, simulate fills conservatively, and decide the evaluation before you look at the results. The same discipline runs through my forecasting project, TidalMoon.
The gate that says no
The most useful thing I built was the part that stopped trades. Every order had to pass a gate first: limits on size and spend, sanity checks, and a kill switch that stopped everything.
After enough tightening, that gate rejected almost everything. That is safer than the alternative, but a gate that always says no isn’t a strategy either. Deciding “should this go through?” quickly and well turned out to be its own research problem. It is one of the reasons I started slaya, a project on small, fast decision models.
What the ten dollars bought
- A data pipeline and backtesting harness I trust, because I broke it so many times.
- A habit of writing the evaluation before the experiment.
- Kill switches and spending caps that now guard everything else I run unattended.
- A clear sense of how efficient a well-traded market really is.
The code stays private. The lessons are here.