Calibrated agents as market makers: automated liquidity provision for long-tail prediction markets

Shah, Nihar
Digital Finance · Springer Nature · 2026-10-27 · DOI: 10.1007/s42521-026-00233-z

Аннотация

Under what conditions can a forecasting agent profitably provide liquidity on a prediction market, and when does that liquidity improve the market? We derive a decomposition of the optimal market-making spread into a microstructure term and an adverse selection term proportional to the agent’s mean absolute estimation error $$\alpha $$ . The profitability condition $$s > 2\varphi \alpha $$ depends only on the sign of per-fill profit and holds for any positive fill-rate function. Market-quality improvement is a strictly stronger requirement: the agent tightens both sides of the book only when, in addition to quoting a narrower spread, its estimate lands close enough to the existing midpoint ( $$|\hat{p} - m| < (s_m - s^*)/2$$ ). Under the model’s error distribution this location condition holds with a closed-form probability whose zero boundary recovers the spread-width threshold, yielding a conditional long-tail result: the probability of improvement increases with the existing market spread. Accounting for the location condition shrinks the estimated improvable share of a 98-market Polymarket cross-section roughly by half (e.g., from 17% to 7% at the best measured calibration). On the empirical side, a realized profit-and-loss backtest on resolved questions—settling inventory at the true outcome under worst-case adverse selection—shows that realized profitability separates well- from poorly-calibrated models that are statistically indistinguishable on price alignment; an inventory-aware extension with position limits and inventory-skewed quotes reduces tail risk (5% expected shortfall and portfolio drawdowns) by 20–30% at negligible cost in mean profit, while a correlated-error stress test shows that shared model errors on correlated events, not per-market variance, are the binding risk. We present all calibrated parameters as illustrative: the spread–volume cross-section does not identify the fill-rate sensitivity $$\kappa $$ , and we report full diagnostics and wide sensitivity ranges. A complementary finding cautions against a common shortcut: five frontier LLMs track market prices closely (MAD $$\le 0.021$$ , measured with the market price in the prompt and therefore an anchoring measurement) yet exhibit realized outcome error up to an order of magnitude larger, and on a 60-question out-of-sample set no model beats a constant base-rate forecast on Brier score. Market-price agreement is not a safe calibration target, and the premise that current models are calibrated enough to clear the profitability condition is not yet established.

Читать у издателя →

Похожие статьи