Market making in prediction markets (Kalshi/Polymarket): what changes in the CLOB stochastic control problem?

External question — Quantitative Finance Stack Exchange Author: Tri4cBet Original post: https://quant.stackexchange.com/questions/85852 License: CC BY-SA 4.0 — https://creativecommons.org/licenses/by-sa/4.0/ Adaptation: HTML converted to plain text; contact email addresses removed. I've been working on market making in binary prediction markets like Kalshi and Polymarket, specifically through the Central Limit Order Book (CLOB), rather than an Automated Market Maker (AMM), and I'm trying to get a better understanding of how people think about the market making problem in this setting. I'm reasonably familiar with the standard market making literature, things like Avellaneda-Stoikov, inventory risk, adverse selection, order arrival models, etc. I've also spent quite a bit of time actually building and testing market makers for these markets. What I'm trying to figure out is how much of the usual framework really carries over to prediction markets and what needs to be treated differently. I think of a binary event contract as a claim paying 1 or 0 at resolution, so the traded price $(p_t)$ can be interpreted as the market-implied probability of the event, under the usual risk-neutral interpretation. For modeling the dynamics, I have been thinking about the corresponding log-odds $(x_t = \log(\frac{p_t}{(1-p_t)}))$ , since that puts the state on the real line and seems like a natural way to model both continuous moves and large information-driven jumps. That seems to lead to a few things that are quite different from conventional market making. For example: The event probability is bounded between 0 and 1 and eventually resolves to 0 or 1. Time to resolution seems to matter much more than it does for something like an equity. Inventory risk doesn't seem to be fully captured by just marking the position to the current mid. The terminal payoff and the remaining time seem to matter quite a bit. There often isn't a clean underlying asset that can be used to hedge the position. New information can cause a large discontinuous move in the probability. On the CLOB side, there are still all the usual issues around queue position, discrete ticks, fill probability, adverse selection, cancellations, and repricing. There are also relationships between different event contracts that seem like they could potentially be useful for managing inventory and risk. The information side is probably the part I'm most interested in. Some events seem to spend a long time in a relatively low-information state and then move very quickly when a piece of information arrives. That makes the usual continuous diffusion assumptions seem like a pretty incomplete description of what a market maker is actually exposed to. There is also an interesting asymmetric information problem. In some markets there may be participants who have a much better information set than the market maker, especially around events where information is concentrated among a relatively small group of people. At some point it seems possible that providing liquidity is just getting systematically picked off. I'm curious how people think about that problem mathematically. Is the right answer in those situations basically to recognize that the information risk is too high and stop providing liquidity, or is there a way to model the adverse selection explicitly and still earn enough spread to make providing liquidity worthwhile? The other problem I'm interested in is what I think of as the cold start problem. When a new event is listed, there may be very little trading history, little information about the order flow, and not much data from which to estimate the dynamics of the probability. That seems different from the usual problem of estimating volatility or order arrival rates for an established asset. The hedging issue also seems interesting even when the related market itself is very liquid. For example, with something like a BTC 15-minute "above" contract, you have very liquid BTC spot, perpetuals, and options that are obviously related to the event, but they still don't give you the same payoff. Even something that looks like a fairly close prediction-market hedge can have a different settlement time or settlement definition. A Kalshi 15-minute BTC contract and a Polymarket 5-minute BTC contract may be strongly related, for example, but that doesn't make the hedge equivalent. I'm curious how people think about these kinds of incomplete hedges and whether there is a useful way to quantify the residual risk rather than just treating the related market as a rough proxy. So a few things I'd particularly like to hear about: Stochastic control: Are there papers that formulate prediction market market making directly as an HJB or HJBQVI problem, rather than adapting the usual equity or FX market making models? Probability dynamics and jumps: Are there standard models for the evolution of an event probability that explicitly include discrete information arrivals or jumps? I'm particularly interested in models that distinguish ordinary short-term movement from information-driven jumps. Inventory and terminal risk: How should inventory risk be formulated when the asset ultimately resolves to a binary payoff? Is there a natural analogue of the inventory penalty in Avellaneda-Stoikov that incorporates the terminal distribution and time to resolution? Hedging: What do people normally do when there isn't a clean underlying to hedge against? Is inventory control basically the main risk-management tool, or are there useful ways to hedge using related event contracts or other markets? More generally, how do you think about hedges when the related instrument has a different payoff or settlement window? Asymmetric information: Are there market making models that are particularly relevant when some participants have a meaningful information advantage? How do you distinguish ordinary adverse selection from situations where the information risk is so large that it no longer makes sense to quote? Cold start: Are there good approaches for initializing a market maker on a newly listed event when there is very little market history? In particular, how do people think about combining an external probability estimate with what little information is available from the CLOB? CLOB microstructure: Are there prediction-market-specific models for queue position, execution probability, adverse selection, and order book dynamics that are worth studying? Literature: If you were approaching this problem from the mathematical side, what papers or books would you consider essential beyond the usual Avellaneda-Stoikov, Guéant, and Cartea and Jaimungal literature? I'm particularly interested in things that help with actually formulating the problem mathematically, rather than papers that are mainly about prediction market efficiency or forecasting accuracy. I'd also be very interested in hearing from anyone who has actually made markets in prediction market CLOBs. I'm not looking for a specific trading strategy. I'm more interested in what is fundamentally different about the problem compared with conventional market making, especially around information arrival, asymmetric information, hedging, and the lack of historical data for newly listed contracts. Any papers, models, or areas of the literature that you think are particularly worth digging into would be greatly appreciated.

Market making in prediction markets (Kalshi/Polymarket): what changes in the CLOB stochastic control problem?

Market making in prediction markets (Kalshi/Polymarket): what changes in the CLOB stochastic control problem?

Post Reply