$all_about_ai

My First Winning Agentic AI Trading Strategy On Polymarket

By · · 5 min read

This is the first Polymarket strategy I have had real success with while letting the system run autonomously in the background. The important shift is simple: instead of trying to be a taker who predicts direction and pays fees/slippage, the strategy sits on the maker side and only takes fills when the price is far enough below my own fair value model.

Watch the video:

Nothing here is financial advice. This is a live experiment around agentic AI trading, and the strategy only makes sense if your fair value model is good enough. Without that, the rest of the system is just a nice-looking way to lose money slowly.

The problem with being a taker

Most people think about Polymarket as a taker venue: you see a price, you click buy, and you take whatever is available in the order book. That is fine for normal betting, but it is not great if you are trying to extract a small repeatable edge from the 5-minute Bitcoin up/down markets.

Two things eat the edge immediately:

That is why this setup does not try to chase the current market price. It tries to post resting orders at prices I actually want. If nobody hits them, fine. No trade is better than a bad trade.

The core idea: fair value minus a discount

The whole strategy depends on calculating a fair value price for the current 5-minute BTC market. If my model thinks the true probability of "up" is 51%, then 0.51 is my fair value. I am not saying Polymarket is always wrong; I am saying my bot only wants to trade when somebody is willing to sell to me at a meaningful discount to that fair value.

In the version I have been testing, the model settled on a 4-cent discount requirement. So if fair value for "up" is 0.51, the bot is only willing to bid around 0.47 for an up share. If the fair value moves down to 0.45, the bid moves down to 0.41. The resting order follows the model, not the crowd.

The down side is just the inverse. If up fair value is 0.51, then down fair value is roughly 0.49. With the same 4-cent discount, the bot only wants to buy down around 0.45 or lower.

The trade only happens when an impatient trader crosses the spread and sells into that resting order. That is the entire edge: do not predict every window, do not force trades, just sit on the book with a price that already has positive expected value baked in.

Why AI helps here

The math is not magic. The hard part is getting the fair value close enough that the discount actually means something. That is where I have been using models like Codex, Claude Code, and Fable-style reasoning to collect data, grade snapshots, inspect model behavior, and keep improving the calibration.

For the 5-minute up/down fair value model, I collected roughly:

That is the real work. Anyone can write a loop that posts a bid. The valuable part is building enough historical and live market context that the bid is anchored to something better than a guess. This is why the strategy fits the agentic trading data pipeline approach so well: collect, grade, analyze, adjust, and then let the pod run.

The current results

The strategy has been running fully autonomously and is up almost $70 in the current test period. The raw count at the time of recording was 32 wins, though that includes some early model data before the parameters were tightened.

After looking at the data, I adjusted the gap so the bot should trade less often but, hopefully, with a steadier return profile. That is the tradeoff I want: fewer fills, better fills. I do not care if this only makes $25 a week if it costs almost nothing to run and behaves well over months.

This is one pod inside the larger agentic AI trading pods idea. I do not want one giant strategy that needs constant babysitting. I want many small, independent strategies that can run quietly, produce data, and survive my own urge to over-tinker.

Maker-side trading is not the rebate game

Polymarket has market maker rewards and rebates, but that is not what this particular strategy is built around. I am not trying to optimize for reward-pool participation. I am using maker mechanics for two practical reasons: avoid taker fees where possible and avoid slippage from chasing the book.

That distinction matters. A pure market-making strategy needs to think about two-sided quoting, inventory, rewards, and staying competitive in the book. This strategy is narrower. It only wants to buy discounted shares when the fair value model says the discount is large enough.

The biggest risk: overfitting the fair value model

The danger is obvious: you can overfit the parameters to the exact week of data you just collected. I have already seen early versions where changing a parameter made the backtest look better but did not necessarily make the live behavior better.

That is why I am trying to make small changes, then keep monitoring. The latest change was a 3-cent gap adjustment after reviewing about seven days of data. I also asked whether we should run Monte Carlo simulations now, but the answer was: later. First we need more live data. Monte Carlo can help answer questions like:

Those are useful questions, but if the sample is too small, the simulation just makes fake confidence. More live data comes first.

Why this is a good agentic strategy

I like this strategy because it does not require the model to predict Bitcoin perfectly. It asks the model to do something more realistic: estimate fair value, compare it to available prices, and only place a resting order when the spread is large enough.

That is a much better fit for an autonomous AI trading loop. The agent can monitor the market, update fair value, move orders, log every skipped trade, and periodically summarize whether the calibration is still holding. The system is not trying to be a genius. It is trying to be patient.

This also connects back to the earlier Polymarket AI trading bot work, but with a more mature lesson: the first version proved the plumbing. This version is about improving the edge.

Resources