AI Market Making Playbook: Trading Prediction Markets

# AI Market Making Playbook: Trading Prediction Markets **Market making on prediction markets** means continuously quoting both buy and sell prices on binary outcome contracts, capturing the bid-ask spread as your profit — and AI agents have made this strategy accessible, scalable, and significantly more profitable for individual traders. This playbook walks you through the mechanics, the math, and the automation stack you need to run a real market-making operation across platforms like Polymarket and Kalshi in 2025 and beyond. --- ## What Is Market Making on Prediction Markets? Unlike traditional equity markets, prediction markets trade contracts that resolve to either $0 or $1 (or $0 and $100 cents on Polymarket). A "YES" share on "Will the Fed cut rates in September?" might trade at 62 cents. A market maker sits on both sides of that contract, posting a bid at 60 cents and an offer at 64 cents, collecting the 4-cent spread from traders who cross the book. This is fundamentally different from **directional trading**, where you take a position because you believe an outcome is more likely than priced. Market makers are (ideally) agnostic about the outcome — they profit from volume and spread, not from being right. The catch: prediction market contracts carry **inventory risk**. If you're holding 10,000 YES shares at 62 cents and the contract suddenly prices at 35 cents after a breaking news event, you've just taken a significant loss. Managing that inventory risk is where AI agents earn their keep. --- ## The Core Math: Spread, Inventory, and Expected Value Before automating anything, you need to understand the three numbers that govern every market-making decision: ### 1. The Effective Spread Your **gross edge** per trade is simple: (Ask price − Bid price) ÷ 2. On a binary contract quoting 60/64, that's 2 cents per side. But your *effective* edge needs to account for: - **Adverse selection**: sophisticated traders hitting your quote because they know something you don't - **Platform fees**: Polymarket charges 2% on winnings; Kalshi charges a trading fee of 1-3% depending on the market - **Slippage on hedges**: if you're cross-market hedging, execution costs eat your edge A realistic effective spread after these costs might be 1.0–1.5 cents on a liquid contract. That sounds small, but at 500 fills per day across 20 markets, you're talking $100–$150 in daily gross profit — or roughly $36,000–$54,000 annually from a single modestly capitalized strategy. ### 2. The Inventory Skew Formula When your inventory gets long (too many YES shares), you need to **skew your quotes** to encourage selling. A simple formula: ``` Adjusted_Bid = Fair_Value − (Half_Spread) − (Inventory_Coefficient × Net_Position) Adjusted_Ask = Fair_Value + (Half_Spread) − (Inventory_Coefficient × Net_Position) ``` The **Inventory_Coefficient** is a tuning parameter — typically 0.001 to 0.005 per share of position. When you're long 500 shares, this pushes both quotes down by 0.5–2.5 cents, making it cheaper to sell to you (discouraging more longs) and cheaper for you to sell (encouraging unwinding). ### 3. Fair Value Estimation Everything hinges on your estimate of the **true probability**. If the market is at 62 and you believe fair value is 65, you're effectively taking a directional tilt *while* market making. AI agents can pull real-time data — polling averages, news sentiment, order flow signals — to continuously update this estimate. The more accurate your fair value model, the tighter you can quote without blowing up on adverse selection. --- ## Building Your AI Agent Stack Running a market-making operation manually is impossible at any meaningful scale. Here's how to structure the automation layer: ### Step-by-Step Setup 1. **Connect to exchange APIs** — Polymarket uses a CLOB (Central Limit Order Book) with a REST and WebSocket API. Kalshi has a similar structure. Authenticate with API keys and set up WebSocket streams for real-time order book data. 2. **Build a fair value model** — Start with a simple model: weighted average of external prediction sites, news sentiment score from an LLM, and recent price momentum. More sophisticated agents use **Bayesian updating** as new information arrives. 3. **Implement the quoting engine** — Your agent should recalculate bid and ask prices every 5–30 seconds (or on every order book update) and submit/cancel/replace limit orders accordingly. 4. **Add inventory management** — Set hard position limits (e.g., max ±2,000 shares per contract). Apply the inventory skew formula above. Trigger a "panic unwind" routine if position exceeds 150% of limit. 5. **Build a risk dashboard** — Track P&L in real time, adverse selection rate (what % of fills are immediately followed by price movement against you), and per-market edge. 6. **Deploy hedging logic** — For correlated markets (e.g., "Fed cuts in July" and "Fed cuts in September"), calculate cross-market exposure and hedge automatically. 7. **Monitor and tune** — Review performance weekly. Adjust spread widths, inventory coefficients, and position limits based on realized P&L and Sharpe ratio. Platforms like [PredictEngine](/) provide a pre-built infrastructure layer for steps 1–4, significantly reducing the time to go live. If you want to explore what full automation looks like end-to-end, the guide on [automating AI agent trading on prediction markets with PredictEngine](/blog/automating-ai-agent-trading-on-prediction-markets-with-predictengine) covers the technical stack in detail. --- ## Market Selection: Which Contracts Should You Make? Not all prediction market contracts are worth quoting. Market makers should focus on contracts with the right combination of liquidity, volatility, and resolution clarity. | Contract Type | Liquidity | Adverse Selection Risk | Typical Spread | Recommended? | |---|---|---|---|---| | High-profile elections | High | High (political insiders) | 3–6 cents | Selective | | Fed rate decisions | Medium-High | Medium | 2–4 cents | Yes | | Sports game outcomes | Medium | Low–Medium | 2–5 cents | Yes | | Crypto price levels | Medium | Medium-High | 3–6 cents | With caution | | Niche political events | Low | High | 8–15 cents | Avoid | | Economic data releases | Medium | High pre-release | 5–10 cents | Avoid pre-release | | Weather/climate events | Low-Medium | Low | 4–8 cents | Emerging opportunity | The sweet spot for new market makers is **Fed decisions, major sports outcomes, and broad political events** where public information is widely available (reducing adverse selection) and volume is sufficient to generate meaningful fill rates. For election-specific strategy with a concrete capital framework, the [election outcome trading playbook with a $10K portfolio guide](/blog/election-outcome-trading-playbook-10k-portfolio-guide) is an excellent companion read to this article. --- ## Risk Management: The Non-Negotiable Rules Market making can blow up fast without strict guardrails. These are the rules that separate sustainable operations from blown accounts: ### The Five Hard Rules - **Rule 1 — Never hold overnight without a stop**: Binary contracts can gap 30+ points overnight on news. Set maximum overnight position limits at 25% of intraday limits. - **Rule 2 — Kill the bot before resolution**: Pull all orders at least 2 hours before contract resolution. Spreads collapse, adverse selection spikes, and there's no edge to capture. - **Rule 3 — Correlation is your hidden enemy**: If you're long Fed-cut YES on Kalshi and long rate-sensitive crypto contracts elsewhere, you have concentrated macro exposure. Track net delta across all correlated contracts. - **Rule 4 — Size to 1% daily VaR**: Your total position exposure should not risk more than 1% of your capital per day under a stress scenario (e.g., all contracts move 20 points against you simultaneously). - **Rule 5 — Log every fill and review weekly**: Adverse selection analysis requires data. If >40% of your fills are immediately followed by a 3+ cent move against you, you're being picked off and need to widen your quotes. For a deeper look at cross-platform risk management, the [real-world prediction market arbitrage case study](/blog/real-world-prediction-market-arbitrage-june-case-study) illustrates how position concentration can create unexpected losses even in seemingly hedged trades. --- ## Cross-Platform Market Making and Arbitrage Synergies Advanced market makers don't just quote one platform — they run **coordinated strategies across Polymarket, Kalshi, and other venues**. When the same underlying event is tradeable on multiple platforms, you can: - **Stat arb the spread**: If Polymarket shows YES at 64 and Kalshi shows YES at 61, buy Kalshi and sell Polymarket for a near-riskless 3-cent capture (accounting for fees). - **Use one platform as a hedge**: If you accumulate excessive long inventory on Polymarket, offload it on Kalshi if prices are equivalent or better. - **Discover price dislocations**: AI agents scanning both order books simultaneously can identify arbitrage windows that last only seconds. The [Polymarket vs Kalshi arbitrage trader playbook](/blog/trader-playbook-polymarket-vs-kalshi-arbitrage-guide) covers the mechanics of cross-platform execution in detail, including how to handle the timing risk when transfers between platforms take minutes but prices can move in seconds. For a broader algorithmic comparison of both platforms, the [algorithmic approach to Polymarket vs Kalshi in 2026](/blog/algorithmic-approach-to-polymarket-vs-kalshi-in-2026) provides a useful framework for deciding where to allocate your quoting capital. --- ## AI Agents in Practice: What They Actually Do Let's be concrete about what an AI market-making agent does in a real trading session on a contract like "Will the Supreme Court issue a major ruling before August recess?": - **At market open**: Agent pulls current probability estimates from 4 sources (Metaculus, PredictIt, news sentiment model, internal momentum signal). Weights them into a fair value of 58%. - **Every 30 seconds**: Agent posts bid at 55.5, ask at 60.5 (5-cent spread, 2.5 cents per side). Adjusts based on inventory. - **Breaking news alert at 10:47am**: Agent's LLM-based news monitor detects a Reuters headline about a related case. Fair value model updates to 72%. Agent cancels all orders within 200ms, reprices quotes at 69.5/74.5 while spreading wider (8 cents) to compensate for elevated uncertainty. - **By end of day**: 340 fills, net position flat, gross P&L of $127. Adverse selection rate: 28% (acceptable). This is the kind of reactive, continuous operation that's physically impossible for a human trader but runs effortlessly for an automated agent. The [AI market making post-2026 midterms analysis](/blog/ai-market-making-on-prediction-markets-post-2026-midterms) explores how these systems perform under high-volatility political environments specifically. --- ## Frequently Asked Questions ## How much capital do I need to start market making on prediction markets? You can start with as little as **$500–$1,000**, but $5,000–$10,000 gives you enough capital to diversify across 10–15 contracts and absorb short-term inventory swings without blowing position limits. At $1,000, you're likely to be position-limited on any contract with more than $200 in volume per day. ## What's the biggest risk in AI-driven market making? **Adverse selection** is the primary risk — sophisticated traders with better information systematically hitting your quotes. The second-biggest risk is model failure: if your fair value estimate is significantly wrong and you're quoting a 4-cent spread around the wrong midpoint, you can lose 10–20 cents per fill before the bot detects the problem. Position limits and stop-loss routines are essential safeguards. ## Do I need coding experience to run an AI market-making agent? Not necessarily. Platforms like [PredictEngine](/) provide no-code and low-code tools that handle API connectivity, order management, and basic quoting logic out of the box. However, to customize your fair value model or implement advanced inventory skew logic, Python proficiency is a significant advantage. ## How do platform fees affect market-making profitability? Fees are the silent killer of narrow-spread strategies. Polymarket's 2% fee on winnings is effectively a per-contract cost that varies with win rate. Kalshi's volume-based fee structure ranges from 1–3%. On a 4-cent gross spread, a 1.5-cent effective fee load leaves only 2.5 cents of edge — which is still profitable at volume, but means you should never quote spreads narrower than 3 cents on most contracts. ## Can I market make on sports contracts profitably? Yes — sports contracts are often **lower in adverse selection** than political markets because outcomes are more random and insiders have less structural edge. The key is pulling your quotes before major injury news, starting lineup announcements, or live game events. Many traders find sports contracts an excellent starting point before moving to higher-volume political and macro markets. ## Is market making on prediction markets legal in the US? Kalshi is a **CFTC-regulated exchange**, so trading there is fully legal for US residents including algorithmic trading. Polymarket is technically geo-restricted for US users (though this is an evolving regulatory situation). Always consult current platform terms and applicable laws in your jurisdiction before deploying capital. --- ## Start Building Your Market-Making Edge Today Market making on prediction markets with AI agents is one of the most compelling systematic trading strategies available to individual traders right now — low competition compared to equity markets, growing liquidity, and a clear edge for anyone willing to build the right infrastructure. The playbook is here: understand your spread math, build a fair value model, automate your quoting engine, and manage inventory like a professional. [PredictEngine](/) is built specifically for traders who want to run this kind of operation without starting from scratch. From API connectivity to AI-powered quoting agents to real-time P&L dashboards, PredictEngine gives you the infrastructure layer so you can focus on strategy and optimization. [Explore PredictEngine's tools and pricing](/pricing) and get your first market-making agent live in days, not months.