Risk & Limits

Turbine Studio helps users build and run automated prediction market strategies, but automation does not remove risk. It can make risk happen faster. Every strategy should have explicit limits before it is backtested or deployed.

Core risks

Prediction market strategies can lose money because of:

• bad signals,
• overfitting,
• poor fills,
• wide spreads,
• partial fills,
• fees,
• stale data,
• market close behavior,
• venue downtime,
• unexpected settlement,
• account or credential issues,
• strategy bugs,
• live behavior diverging from a backtest.

Studio is designed to make these risks easier to inspect, not to eliminate them.

Required limits

Every serious strategy should define:

Backtest risk

Backtests can mislead when:

• the sample is small,
• the best result comes from one market,
• the rules were repeatedly tuned to the same window,
• fill assumptions are too generous,
• future data leaks into the simulation,
• fees are ignored,
• live liquidity is thinner than history suggests.

Use backtests to reject weak strategies. Be slower to use them as proof.

Live deployment risk

Live strategies can behave differently because:

• markets move faster than the strategy loop,
• fills are not guaranteed,
• posted orders can become stale,
• data can arrive late,
• venue rules or access can change,
• the user's credentials or account permissions can fail,
• the runner can lose network connectivity.

This is why first deployments should be small.

AI agent rules

AI agents using Studio should follow these rules:

• Do not call a strategy risk-free.
• Do not promise profit.
• Do not remove risk limits without approval.
• Do not deploy before backtesting unless the user explicitly chooses to skip testing.
• Do not hide weak backtest assumptions.
• Do not treat missing data as favorable.
• Do not increase position size as a default optimization.
• Do not ask users to paste credentials into normal chat.

Conservative defaults

When a user has not specified limits, a conservative draft should:

• use small position sizes,
• avoid wide spreads,
• avoid late entries,
• require fresh data,
• include clear exits,
• pause on repeated errors,
• prefer fewer markets over broad exposure.

The AI should state that these are defaults and ask the user to confirm before deployment.

Human approval points

Studio should keep humans involved at important points:

• approving the strategy,
• accepting assumptions,
• moving from backtest to live,
• increasing exposure,
• adding a new market family,
• changing credential scope,
• restarting after a pause.

The point of Studio is not to make risk invisible. The point is to make strategy automation understandable enough that risk can be bounded.