Big-Dawg: Prediction Market Intelligence

Inspiration

Prediction markets move fast, but most users cannot tell if price action reflects real signal, crowd hype, or manipulation. We wanted to build an intelligence layer that answers one simple question: Can I trust this market?

What It Does

Big-Dawg analyzes live prediction market data and transforms raw trades into:

Integrity scoring (manipulation risk vs. healthy activity)
Information classification (Informed vs. Retail Momentum vs. Whale Dominance vs. Noise)
Confidence metrics (belief dispersion, volatility-based confidence, conviction)
Wallet intelligence (star wallets and capital-weighted signals)
Price prediction signals (cross-sectional alpha model)

Instead of watching price noise, users see where proven capital stands.

How We Built It

Data science and ML pipeline. We built a reproducible analytics stack from raw trades to signals.
Integrity Engine. Combines five risk sub-scores into a weighted risk index; the market health score is:

$$ H = 1 - \left(0.4\,R_{\text{whale}} + 0.2\,R_{\text{impact}} + 0.15\,R_{\text{liq}} + 0.15\,R_{\text{flip}} + 0.1\,R_{\text{cluster}}\right) $$

Information Engine. Classifies behavior using early-entry scores, momentum, whale dominance, and reaction latency with threshold-based decision logic.
Confidence Layer. Computes belief dispersion (std of per-wallet average entry price), volatility-based confidence (1/(1+\sigma)), and an optional conviction score (size (\times) hold time normalized by volatility).
Price Prediction Engine. Trains a cross-sectional linear model on the top 12 markets. Features are extracted over a 30-minute lead window (whale activity, order imbalance, volatility) with target next 30-minute return. We use LinearRegression with an 80/20 train-test split and report (R^2) and coefficients for interpretability.
SQLite Persistence Layer. A specialized data loader caches market scores and classifications with INSERT OR REPLACE keyed on market slugs for fast lookup during scouting and event-level aggregation.

Data Science and Machine Learning in the Project

We used a mix of supervised learning, statistical feature engineering, and rule-based synthesis.

Supervised learning: Our only trained model is a cross-sectional linear regression predicting the next 30-minute return (R_{t+1}) from behavioral features. We extract features per market (whale activity, order imbalance, volatility), stack them across the top 12 markets, and fit one global model with scikit-learn. When (n > 10) we use an 80/20 train-test split and report (R^2) and coefficients so the dashboard can show which factors currently drive predictability.
Feature engineering: All engines rely on hand-crafted features from trade-level and wallet-level data: volume concentration (top 5% wallets), buy-sell imbalance, price standard deviation, per-wallet average entry price (for belief dispersion), and size (\times) hold time (for conviction). The Information Engine uses early-entry analysis (whether trades precede price moves in the same direction) to score informed activity.
Rule-based layers: Integrity (weighted risk to health), Information (thresholded scores to Informed/Momentum/Whale/Noise), and Confidence (dispersion and volatility to confidence score) are deterministic. The Recommendation Engine is a decision tree over these outputs (for example, if Manipulation then INACTION; if Informed and (H > 0.7) and (\text{prob} < 0.3) then STRONG BUY). These verdicts are passed to Gemini for natural-language explanations, bridging statistical outputs and user-facing copy.

Challenges We Ran Into

Data quality drift across market metadata and outcome labels
Building reliable normalization for YES/NO pricing across markets
Maintaining fast analysis latency with large trade histories

Accomplishments We Are Proud Of

We successfully built a full intelligence stack in a hackathon timeframe, implementing a robust Trust Scoring system defined by:

$$ S_{\text{final}} = 0.4 + \left(S_{\text{raw}} \times 0.6\right) $$

At the heart of our user experience is the Recommendation Engine, a synthesis layer that converts abstract engine scores into actionable trade signals. It first evaluates the Integrity Engine health score to filter out manipulation risks, then cross-references the Information Engine classification against the Confidence Layer outputs. When the engine detects a Healthy integrity environment ((H > 0.7)) combined with Informed Activity, it triggers a high-conviction recommendation (for example, STRONG BUY if the price is undervalued). These discrete signals are then injected as contextual seeds into Gemini, which translates raw mathematical verdicts into plain-language explanations that guide users.

What We Learned

Wallet-level history is a significantly more powerful predictor than price alone. We formalized this using a cross-sectional linear regression model to predict short-term returns:

$$ R_{t+1} = \beta_0 + \beta_1 A_{\text{whale}} + \beta_2 I_{\text{order}} + \beta_3 \sigma_p + \epsilon $$

We also learned that simple statistical models like Belief Dispersion can uncover hidden market sentiment:

$$ \sigma_{\text{belief}} = \text{std}(\bar{P}_{\text{wallet},k}) $$

Finally, we found that smart money intensity is captured well by aggregating verified star wallet behavior into a Conviction Score:

$$ S_w = \min\left(\frac{N_{\text{stars}}}{10} + \frac{\sum C_{\text{star}}}{5000},\,1.0\right) $$

On the ML side, we learned that training the alpha model cross-sectionally yields more stable coefficients than single-market fits, and that (R^2) works well as a real-time diagnostic: when (R^2) drops below about 0.10, the market is likely driven by noise or external news rather than the behavioral features we track.