Prediction-market agents

Prediction markets make the relationship between a belief and a price unusually direct. A contract specifies an outcome, and its price gives participants a way to express what they think about it. I’m interested in what happens between forming that belief and deciding whether it justifies a position.

The first task is to understand the contract precisely. A market’s title is a shorthand for a set of conditions: what counts as the event, which source determines the result, and when the question closes. On Polymarket, for example, resolution follows the market’s predefined rules. A good forecast of the headline question can still be poorly matched to what the contract actually settles on.

The next task is to assess the difference between a probability estimate and an available price. That difference only matters if there is enough confidence in the estimate and enough room for the costs of expressing it. Spreads, liquidity, and the time capital remains committed all affect the decision. A model’s confidence deserves scrutiny of its own; a numerical probability can be precise in appearance while resting on limited evidence.

My Polymarket agent explores this as a repeated research process. It keeps notes and moves towards an explicit decision to buy, sell, or pass. Passing matters because the absence of a position is a legitimate outcome of analysis. A useful agent needs the ability to recognise when its view offers too little advantage or too much uncertainty.

Memory is important for the same reason. If a forecast changes, I want to know whether new evidence caused the change or whether the explanation simply drifted. Preserving earlier reasoning makes it possible to compare what was expected with what occurred. Over time, that creates a basis for studying calibration: whether events assigned similar probabilities actually happen at a corresponding frequency.

A related HIP-4 research project examines short-horizon binary markets near settlement. Here, the distance from a threshold, the remaining time, and the uncertainty in the underlying price become central. A seemingly small move can matter a great deal near expiry. The relevant probability has to be considered alongside the actual order book and the exact settlement mechanism.

Position sizing connects the forecast to its consequences. A stronger estimated advantage does not remove the possibility that the estimate is wrong. My research uses constrained sizing and paper evaluation to study those decisions before drawing conclusions from them. I care about how a strategy behaves across repeated uncertain outcomes, including sequences in which several reasonable-looking views fail.

These projects are a way to study judgment in a setting where a belief eventually meets a defined result. The most useful output is a record of decisions that can be revisited: what was known, what was assumed, what price was available, and how much confidence the position expressed. That record gives learning somewhere concrete to begin.