Case Study 2: Adverse Selection in Political Markets

Overview

Political prediction markets are among the most actively traded and closely watched prediction markets in the world. They are also among the most challenging for market makers due to the unique structure of political information: highly asymmetric, temporally clustered around events (debates, primaries, endorsements), and subject to rapid consensus shifts.

This case study examines the dynamics of adverse selection in political prediction markets, drawing on data patterns observed across major prediction market platforms during election cycles. We analyze how information flows, who the informed traders are, how adverse selection varies over time, and what strategies market makers can deploy to survive—and potentially profit—in this environment.

The Information Landscape of Political Markets

Types of Informed Traders

Political prediction markets attract several distinct types of informed participants:

1. Poll Insiders and Early Accessors Some traders gain early access to poll results through relationships with polling firms, media organizations, or campaign staff. Even minutes of advance notice on a major poll release can be worth significant edge in a prediction market.

Estimated fraction of flow: 5–10% Signal quality ($\theta$): 0.75–0.90 Trading pattern: Concentrated bursts before poll releases

2. Political Operatives and Campaign Staff Individuals with inside knowledge of campaign strategy, endorsement timing, and candidate health. Their information advantage is event-specific and often binary.

Estimated fraction of flow: 2–5% Signal quality ($\theta$): 0.85–0.95 Trading pattern: Sporadic, concentrated, large size

3. Data-Savvy Modelers Quantitative analysts who build sophisticated election models (similar to FiveThirtyEight-style analyses). Their edge comes from better information processing rather than private information.

Estimated fraction of flow: 10–20% Signal quality ($\theta$): 0.60–0.75 Trading pattern: Persistent, moderate size, correlated across markets

4. News-Reactive Traders Fast traders who react to breaking news within seconds. Their edge is speed rather than information asymmetry per se.

Estimated fraction of flow: 15–25% Signal quality ($\theta$): 0.55–0.70 Trading pattern: Concentrated immediately after news events

5. Retail/Uninformed Traders Individuals trading based on opinions, wishes, or entertainment value. They provide the "spread revenue" that market makers need to offset informed losses.

Estimated fraction of flow: 40–60% Signal quality: Random (no edge) Trading pattern: Distributed throughout the day

The Adverse Selection Lifecycle

Adverse selection in political markets follows a characteristic pattern over the lifecycle of an election market:

Adverse Selection Intensity Over Time

High |          *   *       *              * *   *
     |         * * * *    * *          *  * * * * *
     |     *  *   *  *   *  *     *  *  *   *   * *
Med  |    * **        * *    * * * **             * *
     |  **             *      * *                   **
Low  | *                                              *
     +--+----+----+----+----+----+----+----+----+----+-->
     Open    Primary Primary Conventions General  Election
             Season  Results             Campaign Day

Phase 1: Market Opening (Low AS) When a market first opens (e.g., 18 months before an election), information asymmetry is low. Few people have meaningfully better forecasts than the crowd. Spreads can be tight.

Phase 2: Primary Season (Rising AS) As primaries approach, campaign insiders and pollsters gain increasingly valuable information. Adverse selection rises, especially around primary election dates and debate performances.

Phase 3: Convention / Nomination (High AS) Major events like party conventions create sharp information asymmetries. Endorsements, vice-presidential picks, and platform changes are known to insiders before the public.

Phase 4: General Campaign (Variable AS) The general election campaign features periodic AS spikes (debates, October surprises, major endorsements) interspersed with quieter periods.

Phase 5: Election Week (Extreme AS) In the final days before the election, early voting data, internal polling, and turnout models create extreme information asymmetry. Exit polls on election day represent the ultimate informed trading signal.

Quantitative Analysis

Dataset Description

We simulate a dataset representative of political prediction markets, based on publicly available trading data patterns. The simulated market is "Will Candidate A win the presidential election?" tracked over 180 days.

Adverse Selection Metrics Over Time

We computed the four key adverse selection metrics at weekly granularity:

Pattern Analysis

Several patterns emerge from this data:

1. VPIN as a Leading Indicator VPIN consistently rose 2–3 days before major adverse selection events, as informed traders began positioning. This makes VPIN a valuable early-warning system for market makers.

2. Toxicity Spikes are Sharp but Short-Lived The toxicity metric (mid-price movement in the direction of trades) spiked sharply around events but reverted within 24–48 hours as information was digested by the broader market.

3. Kyle's Lambda Ratchets Upward Unlike toxicity, Kyle's $\lambda$ showed a secular upward trend, reflecting the increasing information content of order flow as the election approached. This is consistent with the theory that informed traders accumulate as the resolution event nears.

4. Realized Spread Inversion During the six most significant events, the realized spread turned negative—meaning the market maker was losing money on every trade. Between events, the realized spread was positive but thin.

Market Maker P&L Simulation

We simulated three market-making strategies over the 180-day period:

Strategy A: Fixed Spread (4 cents) No adverse selection adjustment. Constant spread regardless of conditions.

Strategy B: VPIN-Adaptive Spread widens when VPIN > 0.5 and withdraws when VPIN > 0.8.

Strategy C: Full Regime-Switching Uses the complete adverse selection detector from Chapter 29, with Bayesian regime detection and pre-emptive spread widening around scheduled events.

Key Finding: Strategy A (fixed spread) was deeply unprofitable. The adverse selection losses from event-driven informed trading far exceeded the spread revenue from uninformed traders. Strategy C, which combined multiple AS detection methods with pre-emptive widening, was the only strategy that would sustain a professional market-making operation.

The Anatomy of an Adverse Selection Event

Case: The Vice Presidential Announcement

Let us examine the VP announcement event in detail (Week 20 in our data). This event is representative of political information shocks.

Timeline:

• T-48 hours: Rumors begin circulating among political insiders. A few large buy orders appear in the "VP pick" conditional market. VPIN rises from 0.45 to 0.55.

• T-24 hours: A prominent political reporter tweets a cryptic hint. Trading volume doubles. The market maker's toxicity metric turns positive (price moving in the direction of trades). VPIN rises to 0.65.

• T-4 hours: An insider source leaks the pick to a news organization. A small number of traders who follow this source begin positioning aggressively. VPIN hits 0.75.

• T-0 (Announcement): The pick is officially announced. Within 30 seconds, the prediction market price jumps 8 cents. Market makers who were still quoting at pre-announcement prices are hit with a wave of toxic flow.

• T+5 minutes: The market stabilizes at the new price level. Market makers who pulled quotes at VPIN 0.75 avoided the worst of the adverse selection. Those who stayed suffered significant losses.

• T+2 hours: Volume returns to normal levels. Adverse selection subsides as the information is fully incorporated.

Losses by Strategy

Who Are the Informed Traders?

Flow Analysis

We can attempt to classify traders by examining trading patterns around information events:

Method: Post-Event Return Analysis

For each trader (identified by pseudonymous account), we compute:

$$ \text{Edge}_j = \frac{1}{N_j} \sum_{i \in \text{trades}_j} \text{sign}(q_{j,i}) \cdot (p_{\text{resolution}} - p_{\text{trade},i}) $$

This measures the average profit per contract for each trader, based on the final resolution price.

Distribution of Trader Edge:

Observation: The top 1% of traders (by edge) account for only 3% of volume but generate outsized adverse selection costs for market makers. A market maker who could identify and avoid these traders would dramatically improve profitability. However, most prediction market platforms enforce anonymity, making this impossible.

Timing as an Adverse Selection Signal

We find that the timing of trades is itself informative:

$$ \Pr(\text{informed} \mid \text{trade within 1 hour of event}) = 0.45 $$ $$ \Pr(\text{informed} \mid \text{trade at random time}) = 0.15 $$

This threefold increase in informed fraction around events is the primary driver of adverse selection spikes.

Optimal Market-Making Strategy for Political Markets

Based on our analysis, the optimal strategy for political prediction market making combines:

1. Calendar-Based Pre-Widening

Maintain a calendar of known political events: - Primary/election dates - Scheduled debates - Federal data releases - Supreme Court decision dates

Pre-widen spreads by 50–100% in the 4 hours before each event and keep them wide for 2 hours after.

2. Real-Time VPIN Monitoring

Continuously monitor VPIN with a 100-trade rolling window. Spread adjustment schedule:

3. Asymmetric Post-Event Resumption

After pulling quotes, resume gradually: - Start with wide spreads (2x normal) and small size (25% normal) - Narrow spreads and increase size over the next 30 minutes - Return to normal only when VPIN falls below 0.45 for at least 15 minutes

4. Cross-Market Hedging

When possible, use correlated markets to reduce directional exposure. For example, if accumulating long "Candidate A wins state X" inventory, sell "Candidate A wins nationally" to partially hedge.

5. Time-of-Day Adjustment

Political markets show strong time-of-day effects: - Night (00:00–06:00 EST): Low volume, low AS. Good for passive market making. - Morning (06:00–10:00 EST): Rising volume as news breaks. Moderate AS. - Midday (10:00–14:00 EST): Peak volume. AS depends on news cycle. - Afternoon (14:00–18:00 EST): Moderate volume. AS often lower. - Evening (18:00–00:00 EST): Events (debates, primaries) drive spike risk.

Adjust base spreads by time of day:

def time_of_day_multiplier(hour_utc: int) -> float:
    """
    Return spread multiplier based on time of day.
    Calibrated for US political prediction markets.
    """
    schedule = {
        range(5, 11): 1.0,    # Night (EST 00-06): low AS, passive
        range(11, 15): 1.2,   # Morning (EST 06-10): rising
        range(15, 19): 1.4,   # Midday (EST 10-14): peak
        range(19, 23): 1.1,   # Afternoon (EST 14-18): moderate
        range(23, 24): 1.5,   # Evening (EST 18-19): event risk
        range(0, 5): 1.5,     # Evening (EST 19-00): event risk
    }
    for time_range, multiplier in schedule.items():
        if hour_utc in time_range:
            return multiplier
    return 1.0

Implications for Platform Design

This analysis has implications for how prediction market platforms should design their liquidity programs:

1. Subsidies Should Scale with Adverse Selection Risk

Political markets have much higher adverse selection than sports or entertainment markets. Subsidy programs should compensate market makers for this risk, either through: - Higher per-trade subsidies for political markets - Loss protection that covers event-driven losses - LMSR-backed liquidity that absorbs the worst of the informed flow

2. Information Event Handling

Platforms could improve market quality by: - Implementing trading halts around scheduled events (similar to circuit breakers in equity markets) - Providing market makers with advance notice of known events - Offering enhanced fee rebates during event periods

3. Anonymity vs. Market Making Incentives

While anonymity protects trader privacy, it prevents market makers from identifying and avoiding toxic flow. A middle ground might be: - Providing market makers with aggregate flow statistics (% of flow from "top performers") - Tagging orders by account age or trading history without revealing identity - Offering market makers the option to trade only with "retail" accounts (at the cost of narrower spreads)

Conclusion

Adverse selection in political prediction markets is real, quantifiable, and manageable—but only with the right tools. A naive market-making strategy that ignores adverse selection will be systematically exploited by informed traders, particularly around political information events. The key to survival is a combination of:

1. Pre-emptive spread widening based on event calendars
2. Real-time monitoring of adverse selection metrics (especially VPIN)
3. Disciplined quote withdrawal during extreme conditions
4. Portfolio diversification across uncorrelated markets
5. Gradual post-event resumption of quoting

Political prediction markets are challenging, but their high spreads (when conditions are favorable) and the platform subsidies available make them potentially profitable for sophisticated market makers who respect the informational dynamics at play.

Discussion Questions

1. Should prediction market platforms reveal any information about the "informativeness" of past order flow to market makers? What are the privacy trade-offs?

2. If a market maker detects that 90% of their losses come from 5% of counterparties, and these are likely insiders, is it ethical to attempt to identify and avoid them?

3. How would the adverse selection dynamics differ in a corporate prediction market (where all traders are employees of the same company)?

4. The analysis shows that pulling quotes during high-AS periods reduces losses but also reduces uptime. From the platform's perspective, is it better to have intermittent tight spreads or continuous wide spreads?

5. Design a subsidy program that would make market making in political prediction markets profitable for a market maker with $10,000 in capital. What would it cost the platform?