Case Study 2: Edge Detection Across 500 Markets

Overview

This case study demonstrates a systematic, quantitative approach to finding edge across a large universe of prediction markets. Rather than analyzing individual markets one by one, we build an automated pipeline that scans 500 markets, identifies potential edge opportunities, filters by EV threshold, sizes positions with the Kelly criterion, and simulates portfolio outcomes. This approach is suited for traders who prefer a data-driven, systematic strategy over discretionary analysis.

Part 1: The Setup

The Universe

We simulate a universe of 500 prediction markets spanning multiple categories:

The Trader's Model

Our systematic trader (call him Alex) has built a quantitative model that produces probability estimates for each market. The model combines:

1. Base rates from historical data (weighted 30%)
2. Feature-based predictions from a logistic regression trained on past markets (weighted 40%)
3. Market price itself, used as a signal (weighted 20%)
4. Recency-adjusted trends capturing momentum in related data (weighted 10%)

Critically, the model is not equally good across all categories. Alex has domain expertise in economics and politics but limited knowledge of sports and entertainment. This means the model's edge varies by category.

Simulation Parameters

Part 2: The Scanning Pipeline

Step 1: Generate Probability Estimates

For each of the 500 markets, Alex's model produces: - A point estimate $\hat{q}$ - An uncertainty measure $\sigma_q$ (standard deviation of the estimate) - A confidence interval $[\hat{q} - 1.28\sigma_q, \hat{q} + 1.28\sigma_q]$ (80% CI)

Step 2: Compare to Market Prices

For each market, compute: - Raw edge: $\hat{q} - p$ (for YES) or $p - \hat{q}$ (for NO) - Best side: YES if $\hat{q} > p$, NO otherwise - Conservative edge: using the lower/upper bound of the CI as appropriate

Step 3: Filter by Edge Threshold

Only consider markets where the conservative edge (based on the 80% CI) exceeds the minimum threshold of 3 percentage points. This is a stringent filter that eliminates marginal opportunities where the model's uncertainty undermines the apparent edge.

Step 4: Compute Kelly Fractions

For each passing market, compute: - Full Kelly fraction - Quarter Kelly fraction (our chosen multiplier) - Capped at 5% of bankroll per position

Step 5: Portfolio Construction

Allocate capital to the highest-EV opportunities, subject to: - Maximum 40% total exposure - Maximum 5% per position - Diversification across categories (no more than 15% in any single category)

Part 3: Scanning Results

Raw Scan Output

Of 500 markets scanned:

41 markets pass all filters, representing 8.2% of the universe. This hit rate is realistic -- the vast majority of markets are either fairly priced or have edge too small to be confidently exploited.

Distribution of Edge

Among the 41 passing markets:

The distribution is right-skewed: most opportunities have modest edge (3-5 pp), with a few high-edge opportunities that disproportionately drive returns.

Edge by Category

Alex's edge is concentrated in politics and economics, consistent with his domain expertise. The model is weakest in entertainment and sports, where Alex has no informational advantage.

Part 4: Portfolio Construction

Capital Allocation

The 41 qualifying markets receive the following allocations:

Top 10 positions by allocated capital:

Portfolio summary:

Portfolio Diversification

The portfolio is concentrated in politics and economics, reflecting Alex's comparative advantage. While this reduces diversification, it concentrates capital where edge is highest.

Part 5: Simulation Results

We simulate the portfolio's performance over 6 months using Monte Carlo methods (1,000 simulation runs). In each run, markets resolve randomly according to the "true" probabilities (which we set to be close to Alex's estimates, with some noise to reflect model imperfection).

Model Accuracy Assumptions

To make the simulation realistic, we assume Alex's model is imperfect:

• For each market, the "true" probability is drawn from a distribution centered on Alex's estimate with standard deviation $\sigma_q$ (the model's own uncertainty estimate)
• On average, Alex's estimates are biased slightly toward the market price (the model is not perfectly independent of market information)
• The model is better calibrated in politics/economics and worse in sports/entertainment

Monte Carlo Results (1,000 Runs)

Interpretation

The portfolio has an 87.3% probability of being profitable over 6 months, with a mean return of 6.84%. This corresponds to an annualized return of approximately 14%, which is attractive for a strategy with limited drawdown.

The 5th percentile outcome (losing about 3%) is manageable. The worst-case drawdown of 11.8% (at 95th percentile) is well within tolerance.

The Sharpe ratio of 1.42 is strong, comparable to top-tier quantitative hedge fund strategies. This is partly because prediction markets have returns that are less correlated with traditional financial markets, which reduces portfolio-level risk.

What Drives Returns

Decomposing the mean return across categories:

Politics and economics together account for 76.3% of returns while using 63.3% of allocated capital. This confirms that Alex should concentrate on these categories.

The Impact of Filtering

To demonstrate the importance of edge filtering, we compare the full-portfolio strategy to alternatives:

The full pipeline dramatically outperforms trading all markets. Trading all 500 markets randomly (no model) produces negative returns due to fees. Trading only the top 10 markets by edge gives good returns but with more concentration risk.

The edge filter is critical. Without it, the many marginal trades (with edge near zero or negative) dilute the portfolio's returns and increase risk.

Part 6: Sensitivity Analysis

Sensitivity to Kelly Fraction

Full Kelly offers the highest returns but with a 22.4% average drawdown. Quarter Kelly provides a good balance of returns and risk management. The probability of profit is fairly stable across fractions, but drawdown risk increases significantly at higher fractions.

Sensitivity to Edge Threshold

There is an optimal trade-off: very low thresholds include too many marginal trades (diluting returns), while very high thresholds exclude too many trades (limiting the number of bets and increasing concentration risk). For this model and universe, a threshold around 3-5 pp maximizes the Sharpe ratio.

Sensitivity to Model Accuracy

We vary the model's accuracy by adjusting how much noise is added to Alex's estimates:

Model quality is the primary driver of profitability. A mediocre model with high noise produces negative returns even with Kelly sizing and edge filtering. This underscores the importance of investing in model development and honest assessment of model accuracy.

Part 7: Practical Considerations

Execution Challenges

The simulation assumes perfect execution (trading at the observed market price). In reality:

1. Slippage: Large orders move the price, especially in thin markets. Alex should break large orders into smaller pieces.
2. Latency: By the time Alex acts on his model's output, the market may have moved. Real-time integration with market APIs reduces this risk.
3. Market availability: Not all 500 markets may be available on a single platform. Cross-platform trading introduces additional complexity.

Correlation Risk

The simulation treats all markets as independent, but in reality:

• Political markets in the same election cycle are correlated
• Economic markets respond to common macro factors
• Multiple sports markets may depend on the same underlying team or player performance

Ignoring correlations leads to overexposure. Alex should group correlated markets and apply position limits at the group level, not just the individual market level.

Rebalancing

As markets resolve and new markets appear, the portfolio must be rebalanced: - Resolved markets free up capital for new opportunities - New markets should be scanned and added if they pass the edge filter - Existing positions should be reviewed if prices move significantly (edge may shrink or grow)

A weekly rebalancing cycle is practical for most prediction market traders. More active traders might rebalance daily.

Scaling Challenges

As bankroll grows, some markets become too small (insufficient liquidity) to absorb larger positions. Alex may need to: - Expand to additional platforms - Trade higher-liquidity markets (which tend to have smaller edges) - Accept lower returns per dollar as the strategy scales

Part 8: Key Takeaways

1. Systematic scanning dramatically outperforms discretionary trading. By evaluating 500 markets rather than relying on a handful of gut-feel trades, Alex identifies more opportunities and avoids selection bias.

2. Edge filtering is essential. Most markets do not offer meaningful edge. Trading everything dilutes returns. A 3-5 percentage point minimum edge threshold balances opportunity capture with quality.

3. Concentrate on your comparative advantage. Alex's edge is concentrated in politics and economics. Rather than trying to trade every category, he should focus where his model performs best.

4. Position sizing protects against model failure. Quarter Kelly with position caps limits drawdowns even when the model is wrong. This is especially important for a systematic strategy where errors can be correlated.

5. Model quality is the bottleneck. All the fancy portfolio construction in the world cannot compensate for poor probability estimates. Invest in model accuracy above all else.

6. The pipeline is more valuable than any single trade. The systematic process -- scanning, filtering, sizing, tracking -- generates consistent returns. No single trade matters much. What matters is the process repeated over hundreds of trades.

Code

The complete simulation code for this case study is available in code/case-study-code.py. It includes: - Market universe generation - Model simulation with configurable accuracy - Edge scanning and filtering pipeline - Kelly-based portfolio construction - Monte Carlo portfolio simulation - Sensitivity analysis across key parameters - Visualization of results