Key Takeaways: Liquidity Provision and Market Making
Core Concepts
1. Market Making Is a Service — and a Business
Market makers provide immediacy, price discovery, and depth to prediction markets. Without them, traders must wait for natural counterparties, prices stagnate, and the market's information-aggregation function degrades. In return, market makers earn the bid-ask spread — but this spread is compensation for real costs, not free money.
2. The Bid-Ask Spread Decomposes Into Three Components
$$ s = s_{\text{adverse selection}} + s_{\text{inventory}} + s_{\text{operational}} $$ In prediction markets, the adverse selection component typically dominates. Understanding this decomposition is essential for setting spreads that are wide enough to survive but narrow enough to attract order flow.
3. The Glosten-Milgrom Model Is Foundational
The model shows how a zero-profit market maker sets bid and ask prices as conditional expectations of the contract value given the type of order observed. Key results: - Spreads widen as the fraction of informed traders ($\alpha$) increases. - Each trade triggers a Bayesian update of the market maker's probability estimate. - Spreads are widest at $\mu = 0.5$ (maximum uncertainty) and narrow near $\mu = 0$ or $\mu = 1$.
4. Inventory Management Is Non-Negotiable
Prediction market positions typically cannot be hedged because events are idiosyncratic. The Avellaneda-Stoikov framework provides the key insight: shift both bid and ask in the direction that reduces inventory. The adjustment is: $$ \Delta = -\gamma \cdot q \cdot \sigma^2 \cdot (T - t) $$ The more inventory you hold, the more risk-averse you are, and the longer until resolution — the larger the shift.
5. Subsidized Market Making Is Often Necessary
Many prediction markets cannot support unsubsidized market making because adverse selection costs exceed what the natural spread can sustain. Common subsidy mechanisms include: - LMSR (bounded maximum loss: $b \cdot \ln n$) - Direct payments for maintaining quotes - Loss protection guarantees - Fee rebates and liquidity mining
6. Adverse Selection Is Detectable
Four key metrics allow market makers to measure adverse selection in real time: - Toxicity — Do prices move in the direction of recent trades? - VPIN — What fraction of volume is directionally imbalanced? - Kyle's Lambda — What is the price impact per unit of order flow? - Realized Spread — Is the market maker earning less than the quoted spread?
When these metrics spike, the market maker should widen spreads, reduce size, or withdraw.
7. Multi-Market Diversification Is the Best Risk Management
A market maker operating across many uncorrelated markets benefits from diversification: losses in one market are offset by gains in others, reducing the variance of total P&L. This is especially powerful in prediction markets, where many events are naturally independent.
Practical Guidelines
For Building a Market-Making Bot
- Fair value estimation is the most critical component — get this right before optimizing anything else.
- Start with wide spreads and narrow gradually as you gain confidence in your model.
- Implement position limits and circuit breakers before going live.
- Log everything — you will need detailed trade-level data for post-trade analysis and model improvement.
For Managing Inventory
- Use nonlinear skewing (convex in inventory level) to become increasingly aggressive about reducing large positions.
- Set hard position limits per market and per portfolio.
- Reduce limits as markets approach resolution to avoid binary payoff risk.
For Detecting and Responding to Adverse Selection
- Monitor VPIN with a rolling window of 50–100 trades.
- Maintain an event calendar and pre-widen spreads around scheduled information events.
- When VPIN exceeds 0.8, consider pulling quotes entirely until conditions normalize.
- Resume quoting gradually after pulling — do not snap back to normal immediately.
For Risk Management
- Set daily loss limits and enforce them automatically.
- Track portfolio-level exposure, not just per-market positions.
- Account for correlations between markets (especially political markets in the same election cycle).
- Binary payoff risk means worst-case losses are calculable — use this to size positions appropriately.
Common Pitfalls
| Pitfall | Why It's Dangerous | How to Avoid |
|---|---|---|
| Ignoring adverse selection | Informed traders will extract systematic losses | Monitor AS metrics; widen spreads responsively |
| No inventory limits | Unbounded positions lead to catastrophic losses | Hard limits per market and portfolio |
| Treating correlated markets independently | Portfolio risk is much higher than the sum of parts | Estimate and monitor correlations; use cluster limits |
| Not adjusting near resolution | Binary payoff risk increases dramatically | Reduce limits, widen spreads, unwind early |
| Over-relying on historical backtest | Past AS patterns may not predict future events | Use adaptive, real-time models |
| Ignoring platform fees | Fees erode a significant share of spread revenue | Factor fees into spread calculations |
Key Formulas
| Formula | Description |
|---|---|
| $a_t = \frac{\mu_t(1+\alpha)/2}{\alpha\mu_t + (1-\alpha)/2}$ | Glosten-Milgrom ask price |
| $b_t = \frac{(1-\alpha)\mu_t/2}{\alpha(1-\mu_t) + (1-\alpha)/2}$ | Glosten-Milgrom bid price |
| $\Delta = -\gamma q \sigma^2 (T-t)$ | Avellaneda-Stoikov inventory skew |
| $C(\mathbf{q}) = b \ln(\sum_i e^{q_i/b})$ | LMSR cost function |
| $\text{VPIN} = \frac{\sum_n |V_n^B - V_n^S|}{n \cdot V_{\text{bucket}}}$ | Volume-synced probability of informed trading |
| $\Delta m_t = \lambda \cdot \text{OFI}_t + \epsilon_t$ | Kyle's lambda regression |
| $s_{\text{realized}} = 2 \cdot \text{sign}(q)(p_{\text{trade}} - m_{t+\Delta t})$ | Realized spread |
What to Remember for Later Chapters
- Chapter 30 (Scoring Rules) builds on the concept of subsidized information revelation. The LMSR is both a scoring rule and a market maker — understanding its dual nature is critical.
- Chapter 31 (Market Manipulation) examines what happens when participants deliberately exploit market makers — the adverse selection framework from this chapter provides the defensive lens.
- Chapter 32 (Platform Design) uses the market-making economics from this chapter to design optimal fee structures, subsidy programs, and market formats.