Autonomous Risk Constitution
Risk management is the constitutional layer of an AI trading bot. It defines which actions remain permissible when confidence, volatility, and market structure change.
These twelve rules translate broad risk principles into agent-operable constraints. BotTrade exposes how the resulting policy affects position size, drawdown, return, turnover, liquidation state, and portfolio behavior.
Set a hard ceiling for each position as a percentage of equity and a lower limit when liquidity or volatility demands it. The language model may recommend a smaller size but cannot exceed the rule. BotTrade exposes equity, cash, positions, and scenario leverage constraints. Log proposed and approved quantities so later concentration can be attributed to the policy or the risk layer.
Group positions by economic driver, sector, asset class, or explicit thesis and limit the combined allocation. Several tickers can express one risk even when the agent describes them as separate ideas. Calculate post-trade group exposure before BotTrade submission. Review simultaneous losses in the trade record to identify correlations the original grouping failed to capture.
Adjust quantity so assets with larger expected movement do not dominate portfolio risk by accident. Use a documented volatility window and cap sudden size changes when the estimate falls. vectorbt can help research volatility-scaling rules. BotTrade then reveals how the selected rule changes simulated positions, return, and maximum drawdown when an autonomous agent proposes trades.
Define minimum cash, conditions for deploying reserves, and whether unallocated capital represents uncertainty or a lack of opportunity. Require the agent to explain cash after every proposed order. BotTrade returns available cash and rejects unsupported quantities. A written cash policy distinguishes intentional defense from an agent that failed to identify or execute a valid position.
Every position needs observable evidence that would cause a reduction or exit, plus a review horizon. Store that condition with the BotTrade decision rationale and check it after later bars arrive. Do not let the model replace it after losses develop. The trade record can then show whether the policy followed its rule or improvised around a failing thesis.
Define thresholds that reduce gross exposure, concentration, or new risk when portfolio equity falls from its peak. A response should consider whether several losses share one driver instead of closing positions mechanically. BotTrade reports equity and maximum drawdown and preserves the sequence of decisions. Inspect the opportunity cost and recovery behavior before retaining the threshold.
Require a proposed rebalance to explain why the new portfolio improves expected return, risk, or thesis quality enough to replace current positions. Add minimum holding or materiality rules where appropriate. BotTrade trade count and history expose repeated small changes. Turnover controls should reduce low-value intervention without preventing exits when an invalidation condition is reached.
Entry logic establishes why a position deserves capital; exit logic responds to invalidation, target completion, risk limits, or superior alternatives. Do not require confidence to reverse before reducing risk. Record both rule sets with the agent configuration. BotTrade's sequential loop shows whether the policy held losses because it kept re-evaluating entry evidence instead of checking exit conditions.
When a proposal exceeds a defined allocation threshold, invoke a critic that searches for omitted downside, correlated exposure, and weak source evidence. The critic cannot trade and receives one bounded pass. Store whether the review changed the BotTrade order. This makes the cost and protection of independent review measurable rather than assuming debate always improves decisions.
Validate symbols, sides, quantities, cash, existing holdings, shorting permission, and leverage against the latest BotTrade state before submission. Never rely on the model's memory of an earlier portfolio. Typed Python helpers and MCP schemas reduce malformed actions, while server constraints provide the final boundary. Treat repeated rejections as an agent defect that requires engineering work.
Evaluate the written risk policy in several BotTrade scenarios that stress volatility, correlation, momentum, reversals, and event shocks. Inspect individual drawdown, liquidation, turnover, and missed opportunity before aggregating. A rule that protects one scenario may create harmful inactivity elsewhere. Preserve the agent and policy version with every run so those trade-offs remain traceable.
Return, Sharpe, Sortino, and drawdown summarize a completed run, but the trade sequence reveals how risk was actually taken. Review quantities, timing, concentration, realized outcomes, and decisions made during losses. BotTrade get_trades and BacktestResult expose that evidence. Use it to convert a vague concern into a specific sizing, exit, or portfolio-rule revision.
An agent becomes trustworthy as an engineered system when its freedom operates inside explicit portfolio boundaries. BotTrade enables builders to compare those boundaries empirically and preserve the combinations that improve risk-adjusted performance.