Detecting Informed Trading in a Prediction Market

Worked Solution

How to Think About It: This is a market microstructure problem, not a calculation -- the interviewer wants to see how you think about information asymmetry and price discovery. The central challenge: you cannot directly observe who is informed. You only see prices and quantities. So the question becomes: what behavioral fingerprints do informed traders leave behind?

Start by thinking about what informed traders do differently from noise traders. Informed traders trade in the direction of their signal, tend to hit the ask (or lift offers) when they are bullish, and their trades have persistent price impact -- meaning the price does not revert after they trade. Uninformed traders, by contrast, are more likely to be liquidity-driven, their trades are transient, and they may even trade against recent momentum.

Key Insight: Informed trading leaves two traces: (1) order flow imbalance that predicts future price moves, and (2) price impact that does not reverse.

The Method:

Part 1: Detecting Informed Trading

1. Order Flow Imbalance (OFI): Classify each trade as buyer-initiated or seller-initiated (e.g., using the Lee-Ready algorithm -- compare trade price to the prevailing midpoint). Compute the net signed volume over a rolling window. A persistent, one-sided imbalance that precedes price moves is a strong informed-trading signal. In a prediction market on a binary event, this is especially clean: if someone is aggressively buying probability, they likely have information.

2. Price Impact Persistence (Kyle's Lambda): Regress price changes on signed order flow: $\Delta p_t = \lambda \cdot Q_t + \epsilon_t$ where $Q_t$ is signed volume at time $t$. A large, statistically significant $\lambda$ means trades move prices permanently -- the hallmark of informed order flow. Uninformed trades have transient impact: the price reverts. Informed trades do not revert.

1. Bid-Ask Spread Decomposition: In a market-making context, the spread has two components: inventory cost and adverse selection cost. If you are a market maker in this prediction market and you keep getting picked off (your quotes get hit right before a big move), the adverse selection component of your realized spread will be large. You can estimate this by comparing your mid-quote at trade time to the mid-quote 10-30 minutes later. If you sold at 0.60 and the price drifted to 0.65, you were adversely selected.

1. Volume Clustering Around Information Events: Informed traders tend to front-run news. Look for volume spikes or unusual order imbalances in the hours before scheduled announcements (e.g., FOMC press conference). If trading activity is elevated and directional well before the event, informed flow is likely.

1. Trader-Level Signals (if you have the data): Some prediction markets (like Polymarket or internal corporate prediction markets) expose pseudonymous wallet or account IDs. You can build a track record for each account -- do they consistently trade before price moves? A trader who is right 65% of the time on binary events over 200+ trades is almost certainly informed.

Part 2: Profiting from Mispricings

First, distinguish the two sources of mispricing:

• Uninformed flow pushes price away from fundamentals. Example: a large uninformed seller drives the probability from 0.55 to 0.45, but the true probability (based on your model) is 0.55. You buy.
• Informed traders are pushing price toward true probability, but slowly. In this case, you want to follow the informed flow, not fade it.

The playbook for fading uninformed-driven mispricings:

1. Build a base model for true probability. For a Fed hike prediction market, this might be a regression on Fed funds futures, inflation prints, and FOMC speaker tone. If the model says 0.60 and the market says 0.45, you have a potential edge.

1. Wait for order flow confirmation. Before betting purely on your model vs. the market, check: is the order flow consistent with informed sellers (permanent impact, no reversion) or noise sellers (transient impact, price recovering)? If prices are bouncing back toward your model estimate within 30 minutes, that is uninformed flow. Fade it.

1. Size and timing. Enter when the spread between your model and the market price is widest. Prediction markets on binary events are illiquid -- a large position will move the market. Scale in gradually.

1. Define your exit. Either you exit when the market price converges to your model, or you exit when new information arrives that invalidates your model. Never hold through the event if you are not confident in your model.

Practical Considerations:

• Adverse selection risk: The most dangerous case is when you think you are fading noise and you are actually fading someone with better information. The Kyle lambda test helps here -- if impact is persistent, do not fade.
• Liquidity risk: Binary event markets can gap at expiry. Make sure you can unwind before the event resolves, or size appropriately for a binary outcome.
• Model risk: Your base probability model is never perfect. If you are running 5 such positions, diversification helps. If you are running 1 concentrated position, be humble about your model confidence.
• Regulatory and operational: In real prediction markets (Kalshi, Polymarket), be aware of position limits and settlement mechanics for binary contracts.

Answer: Detect informed trading via persistent order flow imbalance, non-reverting price impact (high Kyle lambda), and adverse selection in realized spreads. Profit from mispricings by building a base probability model, waiting for uninformed-flow signatures (transient price impact, mean reversion), and fading the mispricing with appropriate sizing and a defined exit.